Repository path: config/archive/budget-2026-07-23-to-2026-09-04.md · rendered 2026-09-09

Budget

Cap: USD 5.00 per calendar day (UTC), in OpenRouter billed cost, all sessions that day combined. Soft cap, self-enforced, no rollover; a ceiling, not a target — but chronic underfunding of load-bearing lines (above all jury calibration) is itself a defect. Runs that don't fit today's headroom are split, scaled down, or deferred (note the deferral in NEXT.md).

Method. Before any API run: check today's rows below, snapshot key usage (GET https://openrouter.ai/api/v1/key, field data.usage), write the pre-flight estimate. After: record the API-returned actual ("usage": {"include": true} → usage.cost per response, summed), and cross-check against the key-usage delta. The key is also used outside this project, so per-request costs are primary; the delta is a sanity bound.

One-time increases (charter §6): REQUEST TO TOM block at the top of NEXT.md plus a matching page in wiki/decisions/open/, stating task, amount, why the cap cannot accommodate it, and the consequence of declining. Active only when Tom's written approval appears in the repo; record any grant verbatim here with scope and expiry. Never spend against an unapproved request.

Grants

(none)

Key-usage snapshots

| 2026-07-25 (S021 start) | 19.046184168 | S021 session start — identical to the S020-end snapshot, i.e. zero non-project key drift between the two sessions; the first time that has been observed. S021's single call is cross-checked per-request only (one call, one cost). || 2026-07-28 (S045 start) | 23.482880228 | S045 session start, before any project spend. +0.005169 above S044's closing snapshot — the same between-session drift S044 attributed to concurrent non-project key use, at the same order of rate and two orders smaller. | | 2026-07-28 (S045 end) | 23.689321227 | S045 after E-20260728b-forced-run-ru (15 probe calls) and two critic passes. Delta 0.206440999 against a per-request sum of 0.179372599, gap $0.027068. Third inexact cross-check in a row and the third with the same attribution; the per-request sum is what is ledgered, per this page's stated method. | | 2026-07-28 (S046 start) | 23.689321227 | S046 session start — identical to S045's closing snapshot, so the between-session drift S044 and S045 both recorded did not recur. Note (abf) applied for the fourth time. | | 2026-07-28 (S046 end) | 23.718394977 | S046 after one critic call. Delta 0.02907375 against a per-request sum of 0.02907375: exact to 1e-9, and the first exact cross-check in four sessions after three consecutive gaps attributed to concurrent non-project key use. One call, one provider, both snapshots written to disk before being read (note (bco)). | | 2026-07-28 (S048, before its only call) | 24.382920625 | S048's opening snapshot. Identical to S047's settled closing figure, which is the whole story: S047 and S048 ran concurrently on the same key, and S048 opened after S047's spend had settled. S048 first recorded this as $0.664526 of unexplained non-project drift above S046's close and that attribution was wrong — it was S047's own $0.401902 plus the lag S047 itself documented. Corrected on the merge. | | 2026-07-28 (S048 end) | 24.382920625 | S048 after one call billed at $0.051713 per-request. Delta exactly 0.000000, on two snapshots plus a third taken later in the session. The endpoint had not moved at all — the same settling lag S047 measured directly (its own figure took three snapshots and $0.347 to catch up). The per-request figure is what is ledgered, per this page's stated method. Note (bcx). | | 2026-07-28 (S049 start) | 24.443194462 | S049 session start. $0.060274 above S048's close, of which S048's own $0.051713 accounts for 86% — S048 closed on a delta of exactly 0.000000 and said so, and its call had settled by the time this snapshot was taken. Residual $0.008561 unattributed, the same order as the small inter-session drifts at S042 and S045. Not the S033 shape (a delta in a session that made no call). Note (abf), sixth application. | | 2026-07-28 (S052) | 25.048308603 | S052 opening and closing snapshot, taken either side of the session's only call. Identical, on a call billed at $0.024405625 per-request — the settling lag note (bcx) names, and the third session to see a delta of exactly zero across a real spend. The per-request figure is what is ledgered, per this page's stated method. Both snapshots written to disk before being read (note (bco)). | | 2026-07-28 (S049 end, settled on re-read) | 24.909870691 | S049 after 14 calls. Delta from the opening snapshot $0.466676 against a per-request sum of $0.405701, gap $0.060975. The larger figure is ledgered, on the S030 precedent for a call the endpoint under-reports: the most economical explanation is the z-ai/glm-5.2 extraction attempt, which consumed 4,248 in / 5,972 out and returned usage.cost of exactly 0. Concurrent non-project key use is the alternative and this ledger has documented that at larger magnitudes; both readings are recorded rather than one being asserted. | | 2026-07-29 (S054 start) | 25.843601589 | S054 session start. $0.008268 above S053's closing 25.835333489. S053 flagged an abandoned qwen/qwen3.7-max call that had billed nothing and warned that a late settlement would appear here; this drift is not evidence that it did — $0.008268 is the same order as the unattributed residuals at S042 ($0.005169) and S049 ($0.008561), neither of which had an abandoned call to blame, and a settled 8,000-token qwen call would be an order larger. The most economical reading is ordinary between-session drift, and the S053 warning is left standing rather than discharged. Note (abf), seventh application. | | 2026-07-29 (S054 end) | 26.163373455 | S054 after nine billed dispatches. Delta 0.319771866 against a per-request sum of 0.319771867 — exact to 1e-9, the second consecutive session with that agreement, and this one across four providers and three labs including two calls that returned nothing. Both snapshots written to disk before being read (note (bco)). A tenth call — the four-class re-run — was dispatched and never returned; it had billed nothing at this snapshot, the same shape S053 recorded and flagged. | | 2026-07-30 (S064 critic, either side) | 30.074394956 | S064 opening snapshot and the snapshot after its critic call, identical, on a call billed at $0.05598 per-request. The fourth session to see a delta of exactly zero across a real spend — the settling lag note (bcx) names. Both written to disk before being read (note (bco)). | | 2026-07-30 (S064 end) | 30.177689543 | S064 after the 7-dispatch rater block. Delta from the rater-block opening 0.103294587 against a per-request sum of 0.146431981, a $0.0431 shortfall. The session-level bound holds (note (bet)): opening 30.074394956 + $0.202411981 = 30.276806937 against a closing 30.177689543, i.e. $0.0991 still unsettled at close, which is the ordinary lag direction and not the (bet) excess direction. Per-request sums are ledgered, per this page's stated method. |

| 2026-08-01 (S085 start) | 42.209298353 | S085 session start, before any project spend. | | 2026-08-01 (S085 end) | 42.262070476 | S085 after four ratification calls. Delta 0.052772123 against a per-request sum of 0.0527721234 — exact to 1e-9, and both snapshots were written to disk before being read (note (bco)). Routing read off every response (note (x)): P1 OpenAI, P2 Google, P3 xAI, P5 Baidu — no price excursion on the seat that has produced them before. |

| 2026-08-02 (S089 open) | 45.164290590 | S089 session start, before any project spend. | | 2026-08-02 (S089 close) | 45.896845469 | S089 after 61 calls (1 critic + 60 scoring). Delta 0.732554879 against a per-request sum of 0.578047435, gap $0.154507444. The per-request sum is what is ledgered, per this page's stated method. The gap is the same order and shape as the non-project key drift recorded on this very UTC day at S087 ($0.543269701); both readings are on the record rather than one asserted. Routing read off every response (note (x)): P1 OpenAI; P2 Google / Google AI Studio; P5 across six providers in 20 calls — Alibaba, DigitalOcean, GMICloud, Novita, StreamLake, Together — with no call billing above its declared per-call worst case. |

| 2026-08-02 (S090 open) | 46.007974247 | S090 session start, before any project spend. $0.111128778 above S089's close (45.896845469), consistent with S089's own $0.154507444 residual settling afterwards. | | 2026-08-02 (S090, after the critic) | 46.022567975 | Delta from the opening snapshot 0.014593728 — the accepted critic body's per-request cost to 1e-9. This is what establishes that the dispatch killed in flight by a foreground timeout billed exactly zero. | | 2026-08-02 (S090 close) | 46.041783975 | After the 9-call subject stage. Stage delta 0.019216000 against a per-request sum of 0.094170700 — $0.074954700 unsettled at close, the ordinary settling lag, note (bcx). Per-request sums ledgered. All three snapshots written to disk before being read (note (bco)). |

| 2026-08-07 (S126 open) | 71.490506475 | S126 session start, before any project spend. | | 2026-08-07 (S126 close) | 71.706679770 | S126 after 14 billed bodies (2 critic passes + 12 rating calls). Delta 0.216173295 against a per-request sum of 0.216173290 — agreement to 5e-9, and the second exact cross-check on this UTC day. Both snapshots written before being read (note (bco)). Routing read off every response (note (x)): P1 OpenAI, P2 Google / Google AI Studio, P3 xAI, P5 DigitalOcean (pinned), critic 1 Together, critic 2 Alibaba — no price excursion. | | 2026-08-07 (S127 open) | 72.239736726 | S127 session start, before any project spend. $0.533056956 above S126's close (71.706679770), and no ledger row and no landed session explains it — origin/main's last commit at S127's start was S126's. Recorded as unexplained rather than absorbed; it is the same order and shape as the non-project key drift already on the record for 2026-08-02 ($0.543269701, S087). No project figure depends on it, because this project ledgers per-request costs and uses the delta only as a cross-check. | | 2026-08-07 (S129 open) | 74.485572001 | S129 session start, before any project spend. $0.007481000 above S128's close (74.478091001) — an ordinary settling lag (note (bcx)), not the unexplained kind recorded at S087 and S127. | | 2026-08-07 (S129 close) | 74.837369916 | S129 after 17 billed bodies. Delta 0.351797915 against a per-request sum of 0.351797915 — exact to 0.000000000, the second exact reconciliation on this UTC day and the third this week. Both snapshots written to disk before being read (note (bco)). Routing (note (x)): J1 OpenAI, J2 Google / Google AI Studio, J3 DigitalOcean (pinned), critic BaseTen, pilot Alibaba, independent hand Mistral — no price excursion. | | 2026-08-07 (S127 close) | 72.975808246 | S127 after 17 billed bodies. Delta 0.736071520 against a per-request sum of 0.732740290 — residual $0.003331230, which is one deepseek translation call billed server-side after a two-minute foreground shell timeout killed the runner mid-call, so its body was never written. It is ledgered as a waste row below and the reconciliation is then exact. Both snapshots written before being read (note (bco)). Routing (note (x)): P1 OpenAI, P2 Google / Google AI Studio, P3 xAI, P4 Morph / Moonshot AI / Wafer, P5 DigitalOcean (pinned), critic Together, replacement seat Alibaba — no price excursion. |

Ledger

S242 — 2026-09-03 (UTC), E-20260903b-line-end-order, ARM-line-end-order step 1 (T3, arm closed resolved)

UTC day 2026-09-03 had one prior session, S241, which spent $0.00, so the whole $5.00 was available. Declared ceiling $4.50 (design-v2 §11), built from a freshly-read P1 price of $2.00 / $12.00 — double the stale table value, note (bsw). Spent $3.402299275, well under both the declared ceiling and the ~$3.50 the run's shape implied; $1.597700725 unspent.

Note (bsf) fires for the sixth time and this run honoured it and was still bitten: the cap was probed per seat at the dispatch batch size and passed, and Q2 (gemini-3.6-flash) then truncated on five later batches of the same shape, losing 60 items — its hidden reasoning is variable, as S231 and S237 both recorded. Q2 failed its gate regardless, so the loss changed no verdict; the bodies were not re-bought (note (brx)).

Note (bsw) fired for the first time: openai/gpt-5.6-terra was read at $2.00 / $12.00 from the API before dispatch, double the $1.00 / $6.00 the table had carried since S106 — the first upward price move recorded here, and the ceiling was set from the read value, not the cache.

Key usage 247.816893901 → 251.219193176, delta 3.402299275, exact against the per-request sum (the opening snapshot already carried the critics). The translation limb (T-hafez-shahed-R59-v1, غزل ۱۵ whole), the selection, both gold coding passes and all arithmetic are the lead's own and are not ledgered (charter §3, A4).

UTC day 2026-09-03 running total after S242: $3.402299275 of $5.00, one session, $1.597700725 unspent.

S240 — 2026-09-02 (UTC), ARM-gulistan step 3, span D (T1)

$0.00 spent. No API call was made, and none was planned. The session's principal unit was a translation limb — حکایات ۳۲–۴۰ of «گلستان» باب دوم rendered whole by the lead, which is free and never ledgered (charter §3, A4) — and a study limb, RS-20260902-arabic-function, whose materials are four public-domain English translations already stored in the repository from S235. Reading them costs nothing.

No key-usage snapshot was taken, because none is needed for a session that made no request: the snapshot exists to bound a spend, and there is no spend to bound. The next session that spends will snapshot against S239's close (174.385329998) as usual, and any gap it finds is between-session drift covering two sessions rather than one — which is worth saying here so the gap is not read as this session's.

UTC day 2026-09-02 running total after S240: $2.057401350 of $5.00, two sessions, $2.942598650 unspent. A $0 session is a normal outcome (continue-prompt.md §7), and this one is $0 because the question the unit asked did not need a model to answer it.

S239 — 2026-09-02 (UTC), E-20260902-rhyme-slot, ARM-rhyme-family step 2 (T2, arm closed resolved)

UTC day 2026-09-02 had no prior row, so the whole $5.00 was available. Declared ceiling $2.20 at design freeze; revised to $2.80 before any run call, on the measured seat probe rather than an assumption (E-20260902-rhyme-slot/amendment-v2-1-probe.md §3, note (abc) — build the worst case from the cap the request actually permits). Spent $2.057401350, $0.742598650 unspent against the revised ceiling and $2.942598650 unspent on the day.

The translation limb — Hafez غزل ۳۵ rendered whole twice, monorhymed and unrhymed, under the new regime R59 — is the lead's own and is not ledgered (charter §3, A4), as are the Ganjoor fetches, the Leaf/Payne extraction, the line-geometry computation, all four contamination measurements, the free gate on step 1's two-word window, and the whole analysis and verification.

Note (bst) is new and it cost the run its first-choice seat. google/gemini-3.6-flash located 8 of 14 senses at cap 2500 and 4 of 14 at cap 6000 on the same item — more room, more tokens spent, a worse answer. It was dropped from the shape. x-ai/grok-4.5 at reasoning: {"effort": "low"} did the work at $0.011 a call, which is also note (bsq)'s counter-case: low effort failed a keyed calibration at S237 and passed a keyed positional control here, so it remains a per-seat, per-shape measurement and not a lever.

Note (bsf) fired for the seventh session running. The three dead stage-G bodies cost $0.092926 and their re-buys $0.062950: a body that spends its whole cap and returns nothing bills for the whole cap, so the abandoned calls were the more expensive half.

Key usage 172.327928648 → 174.385329998, delta 2.057401350 against a per-request sum of 2.057401350 — exact to 1e-9. The opening snapshot sat $0.148274400 above S238's close, recorded as an observation per note (bso).

UTC day 2026-09-02 running total after S239: $2.057401350 of $5.00, one session. $2.942598650 unspent.

S238 — 2026-09-01 (UTC), E-20260901b-leaf-pattern, ARM-leaf-contract step 2 (T4, arm closed resolved)

UTC day 2026-09-01 had one prior session, S237, which spent $3.465563850, leaving $1.534436150. Declared ceiling $0.30 — the run needs no bought judgment at all, only the charter's pre-run critic pass. Spent $0.117903400; $1.416532750 of the day unspent.

This is what the budget is for and it is the cheapest thing it buys. The whole primary — 301 lines, 18 measures, a 9999-resample cluster permutation test, three controls, two coding sensitivities and a 92-check verifier — ran on stored material at $0. The $0.118 bought the two seats that replaced the primary and withdrew a registered prediction before any figure existed; had it not been spent, the session would have published a correlation that could not fail (note (bsr)).

Both seats returned clean at cap 8000 — P1 finish_reason: stop, 96.1 s, provider OpenAI, $0.066007; P3 stop, 155.5 s, provider xAI, $0.0518964. Note (bsf) does not fire: neither truncated, and no probe was needed because the prompt shape (one long document, a numbered-list answer) is the shape S233 and S237 both dispatched these two seats at.

The key-usage delta is exact to 1e-9 for the first time since S232 — 172.179654248 − 172.061750848 = 0.117903400 against a per-request sum of 0.117903400.

UTC day 2026-09-01 running total after S238: $3.583467250 of $5.00, two sessions. $1.416532750 unspent.

S237 — 2026-09-01 (UTC), E-20260901-inversion-habit, ARM-inversion-price step 2 (T3, arm closed)

UTC day 2026-09-01 had no prior row, so the whole $5.00 was available. Declared ceiling $3.50 — raised from $3.00 before the run, on both pre-run critics' finding that the design's own worst case exceeded its ceiling with three seats (C1-A13, C2-A15). Spent $3.465563850; $0.034436150 unspent against the ceiling, $1.534436150 against the day.

What the money bought and what it did not. The coding stage completed — 88 calls, 1 dead batch of 88 — and then the registered calibration gate admitted one seat of a required two, so $2.957048100 of blind coding is reported as a withholding. That is the failure criterion working: it was written before the run and it fired against the run's own interest. The re-run is priced from these figures in wiki/backlog.md's one live row.

Two seat-level facts for future pre-flights. P2 needs batch 12 on this task shape — at 24 it truncates at cap 12000 having answered 18 — which trebles its call count and made it the dearest seat of the three at $1.48. And P3 at reasoning: {"effort": "low"} costs a thirteenth of full effort ($0.0073 against $0.094 a batch) and failed both gates, calling 7 of 19 inverted keyed lines canonical: note (bsq). Its $0.277332 bought nothing that voted.

Note (bsf) fires for the sixth session running, and this is the first run to honour it in full and be bitten anyway — every seat was probed at the batch size it would be dispatched at, and one P2 batch in 44 still truncated at the probed cap. The note is amended rather than discharged.

The one re-buy that was declined. The dead P2 batch would have cost about $0.035 to re-dispatch and would have taken the run to the ceiling exactly. It was left dead, the 12 items carry two seats' codes instead of three, and this is on the result page rather than netted out.

UTC day 2026-09-01 running total after S237: $3.465563850 of $5.00, one session. $1.534436150 unspent.

S236 — 2026-08-31 (UTC), E-20260831b-radif-hands, ARM-radif-hands step 2 (T5, arm closed)

UTC day 2026-08-31 had one prior session, S235, which spent $0.00, so the whole $5.00 was available. Declared ceiling $2.20; spent $1.947485275, $0.252514725 unspent against the ceiling and $3.052514725 of the day's cap unspent.

Note (bsf) fired for the fifth session running and this time it cost money. The cap was probed on this task shape for P2 and not re-probed for P1, because P1 had returned a ten-item probe of the same template clean at 1500. At twenty items per call P1 truncated on five of ten calls at $0.060 each — $0.301 that bought no data. New note (bsm): the batch size is part of the task shape, so probe at the batch you will dispatch.

The key-usage delta does not reconcile, for the first time, and it is the endpoint that is wrong. Opening snapshot 162.267773148, closing 167.942189998, delta $5.674416850. A control call costing $0.000086 moved the endpoint by $0.000000000, so it lags rather than counts extra, and the $3.727 residual is within $0.12 of S233's and S234's ledgered totals combined ($3.610943725). The most economical reading is that the endpoint had not caught up with those two sessions when this one opened. That is an inference, not a measurement. New note (bso): the per-request sum is the ledgered figure and the delta is reported as an observation, never as agreement.

UTC day 2026-08-31 running total after S236: $1.947485275 of $5.00, two sessions (S235 $0.00, S236 $1.947485275). $3.052514725 unspent.

S163 — 2026-08-11 (UTC), E-20260812-unlicensed-typography (ARM-source-beliefs step 2, T2, arm closed)

Pre-flight, written into the frozen design before dispatch and built from max_tokens, not from an expected answer length (note (abc)). One call and one call only: the pre-run critic, seat openai/gpt-5.6-terra, cap 3,000. 3,500 prompt tokens at $1.00/M plus 3,000 completion at $6.00/M = $0.0215. Declared ceiling $0.03. Nothing else in the run costs anything: the census is arithmetic over stored public-domain texts, and the translation is the lead's.

The seat was chosen against a documented failure mode rather than for price. P2, P4 and P5 have all returned finish_reason: "length" with null content in this project; note (bmb)'s fix in run.py is unmade, and the UTC day had $0.213 of headroom left. P1 has no such history.

Waste: $0.00. No dead body, no re-dispatch, no discarded call. 78.6% of the declared ceiling.

Key reconciliation: EXACT, the first since S161. Snapshots 96.627118158 → 96.650683658, delta 0.023565500, against a per-response sum of 0.0235655.

Between-session drift, and it is the largest yet recorded. S162 closed at 95.554610396; this session opened at 96.627118158 — +1.072507762 with no project call in between, against S162's own +0.225104930 and the "twenty-five times recent sessions' figure" it reported. Eighth consecutive session showing drift. Per-request costs are primary and are what is ledgered; at this magnitude the key delta is no longer usable as an independent cross-check, and a session that needs one should snapshot immediately either side of the dispatch, as this one did.

Where the estimate was wrong. By $0.0020655, 9.6% over the line and well under the ceiling — the prompt was longer than 3,500 tokens because it carried the whole design plus the manifest. The cap itself was the defect that cost something other than money: 3,000 tokens did not hold the critic's answer, and findings 8+ are lost. Note (abc) prices the worst case from max_tokens; it still says nothing about whether max_tokens is large enough to hold an answer, which is now the second consecutive session to record that gap.

S162 — 2026-08-11 (UTC), E-20260811h-domestication-channel (ARM-realia-channel step 2, T3, arm closed)

Pre-flight, written before dispatch and printed by run.py --dry-run, built from max_tokens and the exact input length of every pair (note (abc)). Study 64 pairs × 3 seats at cap 450 $0.747534; six duplicate pairs $0.070081; a 20-body re-dispatch contingency $0.129000; pre-run critic $0.041000; plus $0.153746 already burned on two dead critic bodies before the ceiling was set. Declared ceiling $1.15, revised in session from $0.85 → $0.95 → $1.10 → $1.15 as the critic's amendments added a contrast and the dead bodies had to be paid for; every revision is written into design.md §7 with its reason. Opening key snapshot 94.566276446; the UTC day already carried $3.787817294 from six sessions, headroom $1.212182706.

Spent $0.975093250, 84.8% of the declared ceiling and 80.4% of the day's headroom.

Waste $0.153746400, 15.8% of spend — all of it the two dead critic bodies, and the second of them was the lead breaking note (bhf) rule (iii) by raising the cap instead of changing the seat. A further $0.058131 was billed for 14 truncated study bodies that the strict parser voided (note (bmb), new mode); those are counted as spend, not as waste, because 196 of 210 bodies parsed and the primaries were unaffected.

Key reconciliation, NOT exact. 94.566276446 → 95.554610396, delta 0.988333950 against a per-response sum of 0.975093250 — larger by $0.013240700, the non-project direction, breaking a run of three exact reconciliations.

UTC day 2026-08-11 total: $4.762910544 of $5.00 across seven sessions (S156 $0.329161470, S157 $0.956578567, S158 $1.851254924, S159 $0.157081896, S160 $0.203410309, S161 $0.290330128, S162 $0.975093250). $0.237089456 headroom left, 4.7% of the cap.

S156 — 2026-08-11 (UTC), E-20260811-floor (ARM-idiom-reach step 2, T5, arm closed)

Pre-flight, written before dispatch, built from max_tokens (note (abc)). One critic pass (13k prompt, cap 6,000), 9 nat calls (2.3k prompt, cap 1,400), 6 swap calls (1.4k prompt, cap 1,400), plus re-dispatch headroom. Declared ceiling $0.75, raised to $0.90 at the amendment commit when the pre-run critic's amendments took items 36 → 46, swap items 16 → 20 and calls 15 → 18. Opening key snapshot 90.140136724; the UTC day opened empty, headroom $5.00.

Spent $0.329161470, 36.6% of the declared ceiling.

S161 — 2026-08-11 (UTC), E-20260811g-carrier (ARM-carrier step 1, T5, arm constituted)

Pre-flight, written before dispatch. Built from max_tokens and not from an expected answer length (note (abc)): $1.456, declared as a $1.46 ceiling. The pre-run critic's SERIOUS 3 added the LEXNULL gate G4 and the ceiling did not move: the four extra variants were paid for by measuring FLOOR on four windows instead of six, so the run stayed at 50 items and 201 calls. Spent $0.290330128, 19.9% of the ceiling.

Waste: $0.004260050, 1.47%. Key reconciliation: EXACT — 94.050841400 → 94.341171516, delta 0.290330116 against a per-response sum of 0.290330128, 1.2e-8 apart, the third exact reconciliation in a row.

Where the estimate was wrong. It over-priced by 5.0×, the conservative direction, for the same reason as S160: arbiter answers are two-line JSON objects against a 900–1,200 cap. And the 900 cap killed three bodies outright — note (abc)'s complement for the fifth session running, and this time the verdict token was not recoverable, so the cells are void rather than rescued.

S161 total: $0.290330128 against a declared ceiling of $1.46. UTC day 2026-08-11: $3.787817294 of $5.00 across six sessions (S156 $0.329161470, S157 $0.956578567, S158 $1.851254924, S159 $0.157081896, S160 $0.203410309, S161 $0.290330128); $1.212182706 headroom, 24.2% of the cap unused.

S160 — 2026-08-11 (UTC), E-20260811f-dakghar-address (ARM-dakghar step 1, T1, arm constituted)

Pre-flight, written before dispatch. Built from max_tokens and not from an expected answer length (note (abc)): $1.12, declared as a $1.25 ceiling and raised to $1.40 before dispatch with the reason written, when the pre-run critic's SERIOUS 3 added a ninth pair (C-c), worth +6 hand calls and +15 arbiter calls. Spent $0.203410309, 14.5% of the ceiling.

Waste: $0.00. Key reconciliation: EXACT — 93.838473200 → 94.041883500, delta 0.203410300 against a per-response sum of 0.203410309, 9e-9 apart.

Where the estimate was wrong. Nowhere that cost money — it over-priced by 5.5×, which is the conservative direction, mostly because the arbiter answers were two-line JSON objects against a 900–1,200 cap. But the 900 cap truncated one arbiter body, which is note (abc)'s exact complement for the fourth session running: the cap prices the estimate and must also be large enough to hold the answer, and those are two different sizes.

S160 total: $0.203410309 against a declared ceiling of $1.40. UTC day 2026-08-11: $3.497487166 of $5.00 across five sessions (S156 $0.329161470, S157 $0.956578567, S158 $1.851254924, S159 $0.157081896, S160 $0.203410309); $1.502512834 headroom, 30.1% of the cap unused.

S159 — 2026-08-11 (UTC), E-20260811d-title-network (ARM-recurrence step 2, T4, arm closed)

Pre-flight: declared ceiling $1.40, built from max_tokens and the worst plausible provider (note (abc)) — 60 calls at a 4,000 cap plus one critic pass at 12,000. Expected ≈$0.25, from the identical task at E-20260810w ($0.182 for 48 calls). ACTUAL $0.157081896, 11.2% of the ceiling and below the expectation.

S159 total: $0.157081896 against a declared ceiling of $1.40. UTC day 2026-08-11: $0.329161470 (S156) + $0.956578567 (S157) + $1.851254924 (S158) + $0.157081896 (S159) = $3.294076857 of $5.00, 65.9% of the cap, $1.705923143 headroom. Waste $0.00.

Key reconciliation: EXACT. Snapshots 93.664922150 → 93.822004040, delta 0.157081890 against the per-response sum of 0.157081896 — 6e-9 apart, and the opposite of the $0.03–$0.06 excess deltas S157 and S158 both recorded.

Where the estimate was wrong. Nowhere that cost anything: the ceiling was 8.9× the actual, which is the max_tokens arithmetic note (abc) requires and not a forecast. The one figure worth carrying forward is that this task's real appetite is ~$0.0026 per call and has now been measured twice on the same procedure ($0.182/48 and $0.142/60), so a third run of it can declare a tighter ceiling without violating (abc) by citing the measurement rather than the hope.

S158 — 2026-08-11 (UTC), E-20260811c-source-beliefs (ARM-source-beliefs step 1, T2)

Pre-flight: declared ceiling $1.60, raised to $1.80 before dispatch when the critic's amendments added 6 parity pairs and a 12-call explicitation audit; built from max_tokens and the worst plausible provider (note (abc)). Expected ~$0.99. ACTUAL $1.846625863 — an OVERRUN of $0.046626, 2.6% above the declared ceiling, and it is the parity re-dispatch.

S158 total: $1.851254924 — $1.846625863 in persisted bodies plus the $0.004629061 probe — against a declared ceiling of $1.80. An OVERRUN of $0.051254924, 2.8%. UTC day 2026-08-11: $0.329161470 (S156) + $0.956578567 (S157) + $1.851254924 (S158) = $3.136994961 of $5.00, 62.7% of the cap, $1.863005039 headroom. Waste $0.525140707, 28.4%, against S157's 4.5%.

Where the estimate was wrong, and it is one place. Every stage but one came in at or under its line; the audit came in at half. Parity was estimated at $0.18 and cost $0.618838, 3.4× — not because the per-call price was misjudged but because 11 of 18 bodies had to be bought twice. Note (abc) prices the worst case from max_tokens; note (bfc) says declare a reserve for every stage. Neither says what this run learned: a gate's reserve must be the full cost of dispatching it twice, because a gate cannot be scaled down or dropped when its bodies die — a run with no parity data is a run about damage rather than about form. Recorded as a firing of (bfc).

S157 total: $0.956578567 against a declared ceiling of $1.20. UTC day 2026-08-11: $0.329161470 (S156) + $0.956578567 (S157) = $1.285740037 of $5.00, 25.7% of the cap.

Where the estimate was wrong. The $0.35 pre-flight was built from 240 calls at an assumed ~120 output tokens. The critic's amendments added 151 calls — a recognition probe, a fluency probe, a spelling-only manipulation check and an independent annotation, every one of them a repair the run needed — and hidden reasoning on two seats put real output far above 120 tokens. The declared ceiling held and the arithmetic under it did not; note (abc) prices the worst case from max_tokens, and the lesson this row adds is that the call count is not fixed until the critic has been adjudicated.

Key reconciliation, and it does not close in the usual direction. Snapshot 90.464883994 → 91.482416375, delta 1.017532381 against a per-response sum of 0.956578567 — the delta is larger by $0.060953814, where note (abf)'s lag makes it smaller. The key was re-read once per note (bil) and returned the identical figure, so this is not the lag. It is the non-project drift direction this ledger has recorded before (S020, S045). Per-request costs are primary and are what is ledgered.

| 2026-08-11 | S156 | pre-run critic, pass 1 (openai/gpt-5.6-terra, cap 6,000) | $0.049 | 0.05100675 | per-response | TRUNCATED — finish_reason: length. Note (abc) applied to a critic prompt: the cap was built for a shorter answer than the prompt invited | | 2026-08-11 | S156 | pre-run critic, pass 2 (same seat, cap 12,000) | $0.085 | 0.0464793 | per-response | stop, 34,148 chars, 22 findings, 19 BLOCKING. This is the pass adjudicated. Cheaper than pass 1 despite twice the cap | | 2026-08-11 | S156 | nat, 12 live bodies (4 blocks × L1 L2 L3) | $0.25 | 0.04768938 | per-response | 46 passages, 222 of 222 cells | | 2026-08-11 | S156 | swap, 6 live bodies (2 forms × 3 judges) | $0.15 | 0.02132454 | per-response | 20 items each incl. 4 unchanged duplicate controls | | 2026-08-11 | S156 | one probe (google/gemini-3.6-flash, reasoning: {"effort": "low"}) | — | 0.012189 | per-response | the diagnostic that ended the truncation loop; kept live in runs/probe-L3-lowreasoning.json | | 2026-08-11 | S156 | WASTE — nine dead bodies, all google/gemini-3.6-flash | — | 0.150472500 | per-response | see below | | 2026-08-11 | S156 | All lead work — Akutagawa 「蜜柑」 translated whole under R06; the corpus build; code.py, run.py, analyse.py, verify.py; the result page; framework/v0.2 §7.4; every state page | — | $0.00 | charter §3, A4 | Lead translation is free and is never ledgered |

Waste: $0.150472500 — 45.7% of the spend, the worst share this project has recorded. All of it on one slug and one defect, now note (bmb): google/gemini-3.6-flash rejects reasoning: {"enabled": false} with HTTP 400 and, left alone, expands hidden reasoning to fill whatever cap it is given — 1,345 reasoning tokens at a 1,400 cap, 2,494 at a 2,600 cap — truncating the answer both times. Raising the cap, which is note (abc)'s own remedy, made it 1.7× more expensive per body and did not fix it. reasoning: {"effort": "low"} fixed it at the third round, at 1,057 reasoning tokens.

Key reconciliation: LAGGING BY EXACTLY ONE BODY. Snapshots 90.140136724 → 90.456644194, delta 0.316507470 against a per-response sum of 0.329161470. The gap is 0.012654000, which is to the cent the billed cost of nat-a-L3's first dispatch — a single unsettled call, not a discrepancy. Note (abf), and the per-response sum is what is ledgered, per this page's method.

Where the estimate was wrong. Twice, in opposite directions, and both are recorded because they cancel and would otherwise look like an accurate estimate. The critic line was under-built (pass 1 truncated at a cap the pre-flight called sufficient) and the jury lines were over-built — nat came in at $0.048 against $0.25 and swap at $0.021 against $0.15, because the worst case was priced at max_tokens on all three seats when two of the three are terse. The $0.150 of waste is not an estimation error at all: it is a defect the pre-flight had no way to price, because the slug's behaviour was unknown before the run.

S119 — 2026-08-06 (UTC), E-20260806c-source-grammar (ARM-translated-register step 2, arm closed)

Pre-flight, written before dispatch. One API call only: the independent pre-run critic, seat x-ai/grok-4.5 (P3). Worst case $0.20 — max_tokens 24,000 at the P3 list output price of $6.00/M is $0.144, plus ~14k prompt tokens at $2.00/M ($0.028), plus routing margin. Everything else in this session is arithmetic over public-domain text and the lead's own translation, both $0. Opening key-usage snapshot 67.593647444. Headroom at dispatch: $3.719623473 of the UTC day's cap.

S119 total, one billed body: $0.041870400, against a declared worst case of $0.20 — 21%. One dispatch, one accepted body, zero retries, zero finish_reason: length. The key-usage cross-check CLOSES to 0.000000000 — opening snapshot 67.593647444, closing 67.635517844, delta 0.041870400 against a per-request sum of 0.0418704.

The critic was again the whole ledger and again the best line on it. It cost 100% of the session's spend and it is the reason P1 can be believed: without F3's FC4b the run would have reported a replication without ever checking whether the effect survives with quoted speech removed, and without F1 it would have closed the arm on two predictions derived from a translator-coded 2×2 over two unmatched loci.

2026-08-06 day total: $1.322246927 of $5.00 — three sessions (S117 $0.614681164, S118 $0.665695363, S119 $0.041870400). $3.677753073 headroom, 74% of the cap.

S118 — 2026-08-06 (UTC), E-20260806b-berman-occurrence (ARM-berman-occurrence step 1, arm constituted)

S118 total, summed per-request over all 41 stored billed bodies: $0.665695363, against a declared worst case of $1.71 — 39%.

The key-usage cross-check does NOT close, and it is out in the unusual direction. Opening snapshot 66.534097841 (identical to S117's close, i.e. no inter-session drift), closing 67.553677814, delta 1.019579973 against a per-request sum of 0.665695363 — a gap of $0.353884610, with the delta larger than the sum rather than smaller. The per-request sum is what is ledgered, per this page's stated method. Two readings are on the record rather than one asserted: (i) a moonshotai/kimi-k3 dispatch at max_tokens 8,000 was killed in flight by a foreground timeout, so its body was never written and its usage.cost never read — kimi bills $3/$15 per M and a full 8,000-token completion would be of this order; and (ii) ordinary non-project key use, which this ledger has documented at up to $0.543 on a single day (S087). The first is named first because it is known to have happened. 42 raw bodies are on disk; 41 carry a cost, and the 42nd is the critic's keep-alive whitespace body, which billed nothing.

The cheapest line was again the most valuable. The $0.021 critic pass is 3% of the spend and is the reason the run has a defensible positive result at all: its finding 4 (site selection is circular) produced the three systematically-sampled sites, and those sites scored as high as or higher than the nine hand-picked ones — which is the only thing standing between this run's surviving number and the objection that it was arranged.

2026-08-06 day total: $1.280376527 of $5.00 — two sessions (S117 $0.614681164, S118 $0.665695363). $3.719623473 headroom, 74% of the cap.

S117 — 2026-08-06 (UTC), E-20260806-carrier-or-knowledge (ARM-voice-persona step 2, arm closed)

S117 total, summed per-request over all 65 stored billed bodies: $0.614681164, against a declared worst case of $2.90 — 21%. The stage subtotals above were hand-estimated on first writing and three were wrong; they are now the figures summed from the stored bodies, and the correction is recorded rather than made silently — note (bey)'s shape, on a ledger rather than a log.

The most expensive stage was not the primary. The 32-call primary cost $0.180203469; the 6-call parity screen cost $0.262543750, because four wasted bodies and one 16,000-token re-dispatch sit inside it. 43% of the run's spend bought six lines of screen output — note (bhf) again, priced. The worst case was itself revised upward mid-run, from $2.40, by the critic's amendments; the fourth rendering they forced cost $0.

The key-usage cross-check closes to 2e-9 — opening snapshot 65.919416679, closing 66.534097841, delta 0.614681162 against a per-request sum of 0.614681164. Note (bil) was respected: the closing snapshot was taken after a settle, not at the moment the last body returned.

2026-08-06 day total: $0.614681164 of $5.00 — one session (S117 $0.614681164). $4.385318836 headroom, 88% of the cap.

| 2026-07-24 | S010 | first R01-vs-R02 comparison E-20260724-r01r02-selfrevise: 36 translation arms + 96 jury calls (2 works × 2 translators × 3 reps × 3 arms; 2 jurors × 48 payloads) | 3.00 (fund) | 4.488172 | per-response usage.cost, summed over all raw incl. failed attempts | est. was $1.9–3.0; overran to $4.49. Cause: reasoning-model max_tokens starvation truncated 26 calls (wasted $0.577 on failed deepseek/gemini attempts) + grok-4.5's heavy jury reasoning tokens (~$0.04/call vs ~$0.015 assumed). Breakdown: translation $1.116137 (incl. $0.088852 failed), jury $3.372183 (incl. $0.488225 failed). Raw: workshop/experiments/E-20260724-r01r02-selfrevise/runs/. |

2026-07-23 day total: $0.167249 of $5.00 (S001 $0.103820 + S002 ratification votes $0.063429).

| 2026-07-25 | S014 | ratification vote D-20260724-04: 1 non-Anthropic vote (openai/gpt-5.6-terra) | 0.02 | 0.016985 | per-response usage.cost | raw: wiki/decisions/votes/2026-07-25/D-20260724-04__gpt-5.6-terra.json; in 2335 / out 646 tok | | 2026-07-25 | S014 | jury calibration Case A judge pass: 60 scoring calls (5 jurors × 6 items × 2 orderings) | 0.90 | 1.832000 | per-response usage.cost, summed over 60 kept files ($1.816490) + 1 discarded parse-fail attempt ($0.015510) | raw quarantined private-texts/experiments/E-20260723-calibration-v1/runs/jury/. Overran est: moonshotai/kimi-k3 ran ~$0.06–0.11/call (heavy reasoning); deepseek/gpt cheap. | | 2026-07-25 | S014 | jury calibration Case A probe pass: 45 register+identity calls (5 jurors × 9 candidate-spans) | 0.25 | 0.399664 | per-response usage.cost, summed | raw quarantined .../runs/probe/. Also over est (kimi reasoning tax). | | 2026-07-25 | S015 | second-reader verification of the 3 precedent anchors E-20260725-anchor-verification: 12 competence-screen + 9 blind-elicitation + 12 adjudication calls | 1.00–1.60 (cap 2.00) | 1.150577 | key-usage delta (conservative); per-request sum of kept responses 1.074446 | screen 0.051657 · elicit 0.343673 · adjudicate 0.679117 = 1.074446 kept, + ~0.076 in two discarded retry attempts. Raw: workshop/experiments/E-20260725-anchor-verification/runs/ (public tree — all six texts are PD, no quarantine). First run in this project to land inside its estimate; the pre-dispatch worst-case guard never fired. | | 2026-07-25 | S020 | Tier D pilot E-20260725-tierD-ladder stage 1 — the sham arm, which is also the 1–7 scale pilot: 2 items × 2 orderings × 3 jurors | 0.166 (worst case 0.259) | 0.152631 | per-response usage.cost, summed; recomputed independently by tools/verify_tierD.py | 12/12 parsed first attempt, 0 discards. Raw: workshop/experiments/E-20260725-tierD-ladder/runs/ (public tree — both passages PD or lead-authored, no quarantine). | | 2026-07-25 | S020 | same experiment, stage 2 — four perturbation operators × 2 passages × 2 orderings × 3 jurors | 0.663 (worst case 1.035) | 0.602628 | same | 48/48 parsed first attempt, 0 discards. Gate between stages was pre-registered and cleared on both conditions (scale usage, headroom). | | 2026-07-25 | S021 | independent pre-run critic pass on the held-out-pair qualification design (E-20260725-heldout-pair): 1 call, x-ai/grok-4.5 | 0.04 (worst case 0.12) | 0.1150644 | per-response usage.cost | in 12,527 tok · out 15,038 tok of which 9,742 reasoning. Over the central estimate by 2.9×, inside the worst case. Cause is exactly NEXT.md note (b): the estimate priced the central case at ~4k reasoning+output and the model spent 9.7k on reasoning alone. Raw: workshop/experiments/E-20260725-heldout-pair/runs/critic__grok-4.5.json. The run it gated cost $0.00 (no API calls); the pass returned NEEDS-REDESIGN and stopped a false Tier D unblocking. |

2026-07-24 day total: $4.488172 of $5.00 (S010 first R01-vs-R02 comparison). Headroom nearly exhausted — no further API today. Lesson recorded: reasoning models (deepseek-v4-pro, gemini-3.6-flash, grok-4.5) spend 1.7k–11k tokens reasoning before output; size max_tokens for reasoning + output, and budget jury calls on reasoning-heavy models at ~2–3× the naive output-token estimate.

| 2026-07-25 | S022 | D-20260725-06 ratification: independent adversarial review (openai/gpt-5.6-terra) + review vote (google/gemini-3.6-flash) | 0.09–0.21 | 0.117062 | per-response usage.cost | review 0.08549175 (in 12,180 / out 5,496) + vote 0.0315705. Raw: wiki/decisions/votes/2026-07-25/. Verdict Q-A; both rejected the lead's stated preference. | | 2026-07-25 | S022 | D-20260725-05 ratification: independent adversarial review (deepseek/deepseek-v4-pro) + review vote (openai/gpt-5.6-terra) | 0.04–0.11 | 0.098220 | per-response usage.cost | review 0.044837307 + vote 0.0533825. P5 cost 3.8× its config/models.md list price — OpenRouter routed the call to provider Venice at ~$1.65/$3.30 per M against the listed $0.435/$0.87. See the note below. | | 2026-07-25 | S022 | E-20260725-slatef-verification: blind second reading of three Slate F sources, mandated by the D-05 vote (google/gemini-3.6-flash) | 0.09–0.15 | 0.282159 | per-response usage.cost, summed over all seven calls incl. four wasted | Usable $0.115349 (3 calls). Wasted $0.190665 (4 calls) — see the waste row below. Raw: workshop/experiments/E-20260725-slatef-verification/runs/. | | 2026-07-25 | S022 | (of the row above) waste, itemised | — | 0.190665 | same | (a) $0.087833 — a shell loop wrote all three outputs to one file ($n__gemini parsed as one variable name); two paid responses overwritten. (b) $0.102833 — max_tokens sized for output but not for hidden reasoning; P2 spent 5,758 of 5,996 completion tokens reasoning and truncated mid-answer. This is standing note (b) verbatim, not applied. | | 2026-07-25 | S022 | E-20260725-published-audit (the paired unit's study limb) | 0.00 | 0.000000 | — | No API call. The audit instrument, the translation and the analysis are all lead work. |

| 2026-07-25 | S025 | D-20260725-07 ratification: independent adversarial review (openai/gpt-5.6-terra) + review vote (google/gemini-3.6-flash) | 0.107 (worst case 0.119) | 0.071062 | per-response usage.cost; key-usage delta exact | review 0.04170125 (in 8,785 / out 950, provider OpenAI) + vote 0.029361 (in 9,914 / out 1,932, provider Google). Raw: wiki/decisions/votes/2026-07-25/. Verdict: review A, vote C (governs). Third run to land inside its estimate, and 34% under it. What worked: both prompts opened with a hard brevity instruction and a numbered output order chosen so the last item is the one the project can most afford to lose — NEXT.md notes (b) and (v), applied on purpose rather than after the fact. Neither call truncated. Provider read off both responses per note (x): no surprise routing. |

| 2026-07-25 | S026 | E-20260725c-contamination-sweep (the paired unit, both limbs) | 0.00 | 0.000000 | — | No API call. The «Припадок» translation, the sweep tool, the self-test gate, the bracket/asymmetry/centrality diagnostics and the write-up are all lead work. Lead translation is free and is never ledgered (charter §3, A4); the measurement over it is arithmetic. Day headroom unchanged at ~$0.16. |

| 2026-07-26 | S027 | E-20260726-period-control (the paired unit, both limbs) | 0.00 | 0.000000 | — | No API call. The «Роза» translation, the extraction, the 42-passage Garnett~Hapgood reference distribution, the four cell analyses, the rank statistic and the 77-check independent verification are all lead work. Lead translation is free and is never ledgered (charter §3, A4); the measurement over it is arithmetic. First row of a fresh UTC day; full $5.00 headroom remains. |

| 2026-07-26 | S028 | E-20260726-genealogie-period pre-run critic (openai/gpt-5.6-terra, P1) on the frozen design | 0.04 (worst case 0.14) | 0.080616 | per-response usage.cost; key-usage delta exact (19.938089975 → 20.018705600 = 0.080616) | in 4,452 / out 4,447, provider OpenAI — list price honoured, no surprise routing (note (x)). Landed inside its worst case, above its central estimate: the worst case was built from per-call maxima per note (m), and the call used its full output allowance because the critique was long, not because it truncated. The $0.08 bought five defects, eleven confound mechanisms, eight under-specified phrases and two predictions ruled unfalsifiable as written — all applied as dated amendments before any measurement existed (critic.md). Everything else this session — the 1,135-word Nietzsche translation, the 78-section extraction, the four-variant analysis, the reference distribution and the 76-check independent verification — cost $0.00. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-26 | S029 | E-20260726-ovid-period-form pre-run critic (openai/gpt-5.6-terra, P1) on the frozen design | 0.090 (worst case 0.113) | 0.070634 | per-response usage.cost; key-usage delta exact (20.0263443 -> 20.09697805 = 0.07063375) | in 6,821 / out 3,288, finish_reason: stop, provider OpenAI — list price honoured, no surprise routing (note (x)). Landed 22% under the central estimate and inside the worst case, the fourth run in this ledger to land inside its estimate. The estimate was a range with a worst case built from per-call maxima (note (m)) and the prompt opened with an explicit brevity instruction plus a task order whose last item is the one the project could most afford to lose (notes (b), (v)). The $0.07 bought 26 findings: 9 falsifiability rulings against named predictions and failure criteria, 11 confound mechanisms, 10 extraction defects and 5 over-claims — including the ruling that the design's central claim about its own lead statistic ("translator-specific vocabulary, form, register and text length all cancel") was simply false. All applied as dated amendments before any measurement existed (critic.md, design.md §10). Everything else this session — the 120-line Ovid translation, the five-text 15-book extraction, 90 published-pair measurements plus 90 mismatched controls, the dependence diagnostics and the 173-check independent verification — cost $0.00. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-26 | S030 | E-20260726b-forced-or-borrowed pre-run critic (openai/gpt-5.6-terra, P1) on the frozen design | 0.077 (worst case 0.078) | 0.081006 | key-usage delta (larger than the per-request sum — see below) | The successful call: in 6,015 / out 4,274, finish_reason: stop, provider OpenAI, usage.cost $0.0656205 — inside the estimate, the fifth consecutive run to land inside one. It bought 46 findings and 22 dated amendments, including that this design's own §3 claim about what its control controlled for was false (the third session running that pointing the critic at the statistic definitions produced the best single finding — note (rr)), a scaling rule cited by four predictions that had never been written down, and a selection bug that would have built target strings occurring nowhere in More. | | 2026-07-26 | S030 | (of the row above) waste, itemised | — | 0.015386 | key-usage delta | A first attempt at the same call returned 1,034 bytes of keep-alive whitespace and no JSON body at all and was billed anyway. Measured at $0.0096395 immediately after; the key-usage delta at session end is $0.0057464 higher than the per-request sum, which is most plausibly that same call's billing settling after the first snapshot. Ledgered at the larger, conservative figure. Mitigation applied in run_critic.py: retry up to three times, detect a body with no {, and write every attempt to its own file so a second response can never overwrite a first (S022 warning 2, which was about exactly this). | | 2026-07-26 | S030 | E-20260726b-baseline-dependence (the retro-check) and everything else | 0.00 | 0.000000 | — | No API call. The 140-line Ovid translation, the eleven-pair dependence retro-check over 124 units, the locus selection, the ten-window analysis, the name_tokens diagnosis and the 218-check independent verification are all lead work. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-26 | S031 | E-20260726c-forced-or-borrowed-ru pre-run critic (openai/gpt-5.6-terra, P1) on the frozen design | 0.088 (stated worst case 0.090) | 0.091925 | per-response usage.cost; key-usage delta exact (20.54597125 → 20.63789625 = 0.091925) | in 7,399 / out 4,587, finish_reason: stop, provider OpenAI, list price honoured. This is the first run in this ledger to land OUTSIDE its stated worst case, by $0.0019, and the reason is a defect in how the worst case was built, not in the routing: note (m) says build the worst case from per-call maxima, and the estimate used an assumed output of ≤4,500 tokens while max_tokens actually permitted 7,000. The output came in at 4,587. A worst case built from a guess about output length is not a worst case; it must be built from the cap the request actually allows — new note (abc). Against that correct bound the worst case was ~$0.124 and the call landed well inside it. What the $0.09 bought: 16 dated amendments before any locus existed, including — for the fourth session running, note (rr) — that the design's own central claim about its own control was false. §3 asserted that carrying S030's thresholds over verbatim made this a replication; it does not, because the control region had been changed from ~91 tokens to a whole poem of 146–623, which changes what r_i = 0.5 means in an indeterminate direction. A1 bracketed the geometry in response. Also caught: a selection confound on poem length that §7 had missed, that G1's monotonicity can be satisfied by a globally shifted mapping, and that a verifier "not importing select.py" is worthless while both call tools/ngram_overlap.py — which is exactly how S030's 218-check pass ran over a broken name_tokens. | | 2026-07-26 | S031 | the name_tokens repair, its re-runs, and the whole translation limb | 0.00 | 0.000000 | — | No API call. The three-defect repair with ten fixtures, the four gated re-runs over S026/S029/S030, the 42-unit prose-poem probe, the eight-poem 1,871-word Turgenev translation, the two-witness provenance check that found eleven OCR errors, the locus selection, the four-geometry analysis and the 172-check independent verification are all lead work. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-26 | S033 | Unattributed key-usage delta, ledgered conservatively | — | 0.013955 | key-usage delta (20.63789625 → 20.65185145) | S031's closing snapshot to S033's opening one, across S032 — a session that made no API call and recorded $0.00. S031's own cross-check was exact at the time it was taken, so this is most plausibly late settling, the same mechanism S030 recorded and in the same direction. Charged to the day rather than written off, and the practice that found it is new: snapshot at session start as well as end — note (abf). One occurrence is not a pattern; a third would be. | | 2026-07-26 | S033 | D-20260726-08 ratification — independent adversarial review (openai/gpt-5.6-terra, P1) | 0.068 (worst case) | 0.037055 | per-response usage.cost | in 1,591 / out 2,139, finish_reason: stop, provider OpenAI, list price honoured (note (x)). Returned E and, with the vote, rejected the lead's provisional default A. Its best line was against the default's own defence — "no repository figure uses a wider set… establishes only that widening has no immediate backward-compatibility cost. Indeed, that may be the least costly time to decide the policy." | | 2026-07-26 | S033 | D-20260726-08 ratifying vote (google/gemini-3.6-flash, P2) | 0.030 (worst case) | 0.013969 | per-response usage.cost | in 2,183 / out 1,426, finish_reason: stop, provider Google. Returned E, agreeing with the review — the first unanimous ratification in this project; D-05, D-06 and D-07 all split. Recorded as agreement, not independent corroboration: the vote's reason restates the review's. | | 2026-07-26 | S033 | E-20260726d-tierD-heldout pre-run critic (openai/gpt-5.6-terra, P1) on the frozen design | 0.130 (worst case from max_tokens) | 0.091200 | per-response usage.cost | in 9,015 / out 4,202, finish_reason: stop, provider OpenAI. Landed 30% inside a worst case built the way note (abc) requires — from the max_tokens the request actually permits (7,000), not from an assumed output length. The sixth consecutive critic pass to land inside its estimate. What the $0.09 bought: 22 findings and a NEEDS-REDESIGN verdict, 17 implemented before any perturbation existed. Two were defects in what the design said about its own rules — "the same rule at the same threshold" (false in proportion, sidedness, null probability and power at once) and a specificity "iff" that contradicted its own naturalness veto, leaving an outcome with no assigned reading. Note (rr) has now fired five sessions running, and five times on the same shape. It also caught that the worst case excluded retries and that a running-margin abort could truncate mid-stage despite a stage-boundary promise. | | 2026-07-26 | S033 | the translation limb, the contamination gate, the design and everything else | 0.00 | 0.000000 | — | No API call. The 270-word Korolenko translation with its twelve-point frozen log, the blind download of Fell 1916, the contamination selection gate, the two-half split with its measured landmark check, the exact null-probability and power arithmetic, the never_names parameterisation with three new fixtures, and the v1→v2 amendment of eleven sections are all lead work. Lead translation is free and is never ledgered (charter §3, A4). | | 2026-07-26 | S034 | E-20260726d-tierD-heldout — the full four-stage Tier D run, 60 calls, P1/P2/P5 | worst case 3.55, central 0.755, reserved per stage | 0.725719 | per-response usage.cost; key-usage delta exact (20.79407595 → 21.519795134 = 0.725719) | 60/60 parsed, 0 failures, 0 retries, every finish_reason: stop. Landed 4% inside the central estimate and at 20% of the worst case — the third run to land inside its estimate, and the first built on per-juror max_tokens (P1 3,000 / P2 6,000 / P5 14,000) rather than one global cap sized for the heaviest juror. Stage actuals: sham $0.162415, held-out $0.192735, targeted heavy $0.196686, targeted light $0.173883. The full-stage reservation rule worked as designed (critic F2): each stage's whole worst case incl. its 10% retry allowance was reserved against the day's headroom before the stage was entered — $0.73 / $0.73 / $1.04 / $1.04 — and no stage was refused. No call on any juror billed above its declared per-call worst case (P1 max $0.01918 of $0.0538; P2 $0.02416 of $0.0503; P5 $0.02032 of $0.0520), so the tripwire on the price table never fired. Routing note (x): P1 all OpenAI; P2 Google and Google AI Studio; P5 across seven providers in 20 calls — Alibaba, Baidu, BaseTen, GMICloud, Ionstream, Novita, StreamLake — with no price excursion, which is the first time P5's routing spread has been this wide without a Venice-style overcharge. | | 2026-07-26 | S034 | the build, the scoring, the verification and everything else | 0.00 | 0.000000 | — | No API call. The six-part §3.3 build audit, the 24 O4 perturbation sites with their Russian bases, the two count-neutral sham sets, the scorer, and the 71-check independent verifier — which recomputes the null probabilities by exhaustive enumeration of 3⁹ outcomes and the contamination gate with its own tokeniser — are all lead work. |

| 2026-07-26 | S035 | ARM-framework step 1 — the traceability inventory, the coverage measure and the FR→EN translation limb | 0.00 | 0.000000 | — | No API call. The 542-word George Sand translation with its twenty-one-point frozen log, the three-way contamination measurement against two published English translations, the sort of 17 result pages / 8 anchors / 10 sources / 2 theory pages into twelve candidate recommendations by evidence class, the nine claim pages and the 21-decision coverage mapping are all lead work. Lead translation is free and is never ledgered (charter §3, A4). The session had ~$3.79 of headroom and needed none of it: what it did was sorting and reading, and the one thing it could not do — give a jury verdict evidential weight — no amount of budget can currently buy. |

| 2026-07-27 | S036 | ARM-longwork step 1 — choosing the long work under the contamination gate, and span 1 of Verga | 0.00 | 0.000000 | — | No API call. The 2,288-word Italian→English span with its 19-point frozen log, the binding register, the R05 regime, the four-candidate survey and the three-cell contamination measurement (dependence_check.py, run locally) are all lead work. Lead translation is free and is never ledgered (charter §3, A4). Nothing this unit needed could be bought: the one thing that would have improved it — D. H. Lawrence's 1928 English, the comparator that matters — is a reachability problem, not a budget one. |

| 2026-07-27 | S037 | ARM-typology-logs step 4 — blind independent re-coding of 60 log decisions, two panel roles | 0.28 | 0.108360 | P1 openai/gpt-5.6-terra (provider OpenAI), P2 google/gemini-3.6-flash (providers Google, then Google AI Studio) | Two attempts, and the first was wasted — recorded because the estimate method is the thing that failed. Attempt 1 at max_tokens: 2000 cost $0.057517 and returned nothing usable: both models spent the whole cap on reasoning tokens that are billed and not returned, and came back finish_reason: length (P1 zero content, P2 a truncated JSON object). Attempt 2 at max_tokens: 10000 with reasoning: {"effort": "low"} cost $0.050843 and both completed finish_reason: stop. Note (abc) says build the worst case from the cap actually sent, which was done. The estimate was still wrong, and not for a new reason: note (b) — "reasoning models spend 1.7k-11k tokens before output; size max_tokens for reasoning plus output" — has been on the page since S010 and was not applied. The failure is a note that did not fire, not a note the project lacked; (b) is marked fired at S037 and (abc)'s wording is sharpened to say what "the cap allows" has to include. Raw JSON for both attempts preserved under E-20260727-log-decision-coding/run/. Actual came in at 39% of the $0.28 worst case even with the wasted attempt included. Key-usage delta unavailable as a cross-check — GET /api/v1/key returned an unchanged all-time figure (21.915886734) before, immediately after, and several minutes after the calls; per-request usage.include costs are the record and the missing sanity check is stated rather than assumed to agree. The session's translation limb (689 words of Bécquer, Spanish→English) and the whole reading of 31,500 words of translator's log cost $0 and are not ledgered (charter §3, A4). |

| 2026-07-27 | S038 | D-20260727-08 ratification — independent adversarial review (openai/gpt-5.6-terra, P1) | 0.195 (worst case from max_tokens) | 0.031278125 | per-response usage.cost | in 5,560 / out 927, finish_reason: stop, provider OpenAI, list price honoured (note (x), sixth run running). Returned B-with-amendment, rejecting S037's default of E and option A. Landed at 16% of the worst case. | | 2026-07-27 | S038 | D-20260727-08 ratifying vote (google/gemini-3.6-flash, P2) | 0.086 (worst case from max_tokens) | 0.036393 | per-response usage.cost | in 6,267 / out 3,599 of which 2,943 reasoning, finish_reason: stop, provider Google AI Studio. Also B-with-amendment — the second unanimous ratification in this project — and it added the voice clause the review missed. The reasoning figure is the point: 2,943 unreturned-but-billed tokens is why S037's max_tokens: 2000 cap destroyed $0.057517, and why notes (b) and (abc) were applied here as a matter of course rather than as a repair. | | 2026-07-27 | S038 | the whole principal unit — translation, anchor, instrument | 0.00 | 0.000000 | — | No API call. The 880-word English→Japanese translation with its 24-point frozen log, the sixth Tier 1 anchor with 51 machine-verified quotations, tools/dependence_check_cjk.py with eleven Japanese fixtures, the four-cell contamination measurement, the CJK decision-probe test and the null-probability recomputation that downgraded S037's headline are all lead work. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-27 | S038 | S037 settlement, ledgered conservatively | — | 0.022149 | key-usage delta (21.915886734 → 22.046395584) | S037 recorded that GET /api/v1/key returned an unchanged all-time figure before, immediately after, and minutes after its calls, and said so rather than assuming agreement. S038's opening snapshot is 22.046395584, i.e. $0.130509 above S037's last reading against a per-request sum of $0.108360. The difference is most plausibly S037's own spend settling late — the same mechanism as S030 and S033, in the same direction — and is charged to the day rather than written off. Note (abf) has now fired twice and the practice that finds it is the opening snapshot. Two occurrences is not yet a pattern in the sense that matters (both are explicable by late settling of a known call); a delta in a session that made no call at all, as at S033, is the shape to watch. |

| 2026-07-27 | S041 | E-20260727b independent pre-run critic pass (openai/gpt-5.6-terra, P1) | 0.191 (worst case from max_tokens) | 0.065876875 | per-response usage.cost, and an exact key-usage delta | in 4,406 / out 3,474, finish_reason: stop, latency 48.2 s, temperature 0.2, max_tokens 12,000. Worst case built the way note (abc) requires — from the cap actually sent (12,000 × $15.00/M = $0.180) plus input (4.5k × $2.50/M = $0.011) — and sized for reasoning plus output per note (b). Landed at 34% of the worst case. It bought six mandatory fixes, one of which was a real bug in the frozen design: [.!?]+(?=\s\|$) had a markdown table-escape inside it and required a literal pipe, so the frozen sentence-counter did not count sentences. Standing disposition 9 exists for exactly that and the pass is what caught it. |

| 2026-07-27 | S041 | the whole principal unit — two translations, two regimes, the analysis, the verifier, the contamination measurement | 0.00 | 0.000000 | — | No API call. The paired JA→EN translation of Sōseki (1,210 + 1,206 words, 36 logged decisions across two frozen logs), regimes R06 and R04 v1.0, the arm-identifiability analysis over 36 stored translations and 96 stored jury payloads, the 112-check independent verifier and the three-cell contamination measurement are all lead work. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-27 | S042 | ARM-longwork step 3 independent pre-run critic pass on the span-3 pre-registration (openai/gpt-5.6-terra, P1) | 0.206 (worst case from max_tokens) | 0.0642815625 | per-response usage.cost, and an exact key-usage delta | in 9,744 / out 6,541 of which 4,457 reasoning, finish_reason: stop, latency 80.4 s, temperature 0.2, max_tokens 12,000, provider OpenAI (list price honoured, note (x), seventh run running). Worst case built as note (abc) requires — from the cap actually sent (12,000 × $15.00/M = $0.180) plus input (9.7k × $2.50/M = $0.026) — and sized for reasoning plus output per note (b). Landed at 31% of the worst case. Raw: workshop/translations/jeli-il-pastore/R05-v1/critic/. It bought a redesign, not a polish: sixteen findings, fourteen mandatory, ten accepted and acted on before the freeze, including the withdrawal of a whole criterion that both this pre-registration and the frozen S039 one had been using to produce counts (note (bcl)). | | 2026-07-27 | S042 | the whole principal unit — span 3, the recount, the verifier, the contamination re-run | 0.00 | 0.000000 | — | No API call. The 2,093-word Italian→English span with its 16-point frozen log (D39–D54), the widened criterion and its retrospective recount over 4,301 words of frozen prose, fid_sites_verify.py and its 237-candidate auditable enumeration, and the four-cell midpoint contamination measurement are all lead work. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-27 | S043 | E-20260727c independent pre-run critic pass (openai/gpt-5.6-terra, P1) | 0.214 (worst case from max_tokens) | 0.051595 | per-response usage.cost | in 13,615 / out 4,043 of which 1,034 reasoning, finish_reason: stop, temperature 0.2, max_tokens 12,000, provider OpenAI (list price honoured, note (x), eighth run running). Worst case built as note (abc) requires — from the cap actually sent (12,000 × $15.00/M = $0.180) plus input (13.6k × $2.50/M = $0.034). Landed at 24% of the worst case. It bought a control the design did not have: it identified that condition B was not a manipulation of the rule alone but of rule-plus-authority-plus-length, and the sham condition added in response is what makes the result reportable. Nine MANDATORY findings, six accepted and acted on before dispatch, three declined in writing (critic/dispositions.md). Instrument caution: it fabricated seventeen item ids that do not exist, so its structural findings held on re-check and several of its examples did not. | | 2026-07-27 | S043 | E-20260727c rating run — 3 raters × 4 conditions, 53 items each | 1.536 (worst case from max_tokens) | 0.234404 | per-response usage.cost, and an exact key-usage delta | 12 calls, all finish_reason: stop, temperature 0, max_tokens 12,000. P1 openai/gpt-5.6-terra (provider OpenAI), P2 google/gemini-3.6-flash (Google, Google AI Studio), P3 x-ai/grok-4.5 (xAI). Worst case built per note (abc) from the cap actually sent across all twelve calls at each model's out-price, plus input; landed at 15%. Key-usage cross-check exact to 1e-7 — 22.522355146 → 22.75675907, delta $0.2344039 against a per-request sum of $0.234404 — but only on re-read: the endpoint returned the unchanged opening figure immediately after the twelfth call, which is the S037 lag recurring. Method miss worth recording: critic.py printed its key snapshots instead of writing them to disk, so the critic call's own delta was lost when the terminal output was truncated and that call is on per-request cost alone. Snapshots belong in a file — note (bco). | | 2026-07-27 | S043 | the whole principal unit — translation, item build, analysis, verifier | 0.00 | 0.000000 | — | No API call. The 2,281-character Japanese→English translation of Ōgai §4 with its 35-decision frozen log, its single-pass R06 draft frozen first, the 53-item build with its mechanical leak check, analyse.py and the 46-check independent verify.py are all lead work. Lead translation is free and is never ledgered (charter §3, A4). |

2026-07-27 day total: $0.614338 of $5.00 (S036 $0.000000; S037 $0.108360; S038 $0.067671 + $0.022149 settlement; S039 $0.000000; S040 $0.000000; S041 $0.065877; S042 $0.064282; S043 $0.285999). $4.385662 headroom remains today. Eighth session of this UTC day, and the largest single-session spend since S034 — bought 13 calls, of which the one that changed the design most was the cheapest.

S042 cross-check: EXACT, and the eighth exact one in this ledger. Opening snapshot 22.392956584, closing 22.457238146, delta 0.064281562 against a per-request sum of 0.0642815625 — agreement to 1e-9 on a single call. Both snapshots taken by rule (note (abf)).

S042 also records a third small inter-session drift, and it is a different animal from the S033 one. S041 closed at 22.385547484 and S042 opened at 22.392956584 — +$0.0074091 across a session boundary. It cannot be S041's spend settling late, because S041's own cross-check was already exact to 1e-9. It is therefore non-project key use, which this ledger has documented before at far larger magnitudes (+$0.359 between S015 and S020, +$0.36 between S002 and S010), and it is not charged to the day: per-request costs are primary and the delta is a sanity bound. The backlog row to watch is still the S033 shape — a delta inside a session that made no call at all — and that has occurred once. Recorded here so the distinction does not get lost by a later session reading three numbers and calling them a pattern.

S041 cross-check: EXACT, and the seventh exact one in this ledger. Opening snapshot 22.319670609, closing 22.385547484, delta 0.065876875 against a per-request sum of 0.065876875 — agreement to 1e-9 on a single call. Both snapshots were taken by rule (note (abf)); the opening one also shows no unattributed drift since S038's close plus S040's zero-spend session.

S039 spent nothing, and no snapshot was taken because no request was made. The session's principal unit was ARM-longwork span 2: a lead translation (free, never ledgered — charter §3, A4) wired to a study limb that read a public-domain Italian preface over plain HTTPS. There was nothing the API could do that the session needed. A $0 session is a normal outcome (continue-prompt.md §7), and this is the third in the ledger.

S038 cross-check: EXACT, and the sixth exact one in this ledger. Opening snapshot 22.046395584, closing 22.114066709, delta 0.067671125 against a per-request sum of 0.067671125 — agreement to 1e-9 across two calls to two providers. Taken at session end per note (zz) and S037's lagging-endpoint experience. It also settles the S037 row above: the endpoint tracks exactly when read late, so the $0.022149 gap is late settling of S037's own calls and not an unattributed charge — which is why the row is written as a settlement rather than as a repeat of S033.

2026-07-26 day total: $1.206080 of $5.00 (S027 $0.00 + S028 $0.080616 + S029 $0.070634 + S030 $0.081006 + S031 $0.091925 + S032 $0.00 + S033 $0.142225 + $0.013955 unattributed + S034 $0.725719). ~$3.79 headroom remains today. Nine sessions in one UTC day, of which three (S027, S032, S035) spent $0.00 — a $0 session is a normal outcome, not a shortfall.

S034 cross-check: exact, and the fifth exact one in this ledger. Session-start snapshot 20.79407595 — identical to S033's closing snapshot, so the unattributed drift S033 recorded did not recur. End snapshot 21.519795134, delta 0.725719, against a per-request sum of 0.725719: agreement to 1e-6 across 60 calls to three labs and nine distinct providers. Note (abf) — snapshot at session start as well as end — has now been applied twice and is what would have caught a second occurrence. One occurrence remains one occurrence; a third would be a pattern.

S033 cross-check: exact, and the fourth exact one in this ledger. Session-start snapshot 20.65185145, end 20.79407595, delta 0.142225, against a per-request sum of 0.142225 — agreement to 1e-6 across three calls to two providers. The unattributed row above is outside that check by construction: it is the gap between the previous session's closing snapshot and this session's opening one, and taking an opening snapshot is what made it visible at all.

S030 cross-check note — the first inexact one, and the reason is known. Per-request usage.cost sums to $0.075260 (0.0656205 successful + 0.0096395 measured for the lost call); the session key-usage delta is $0.081006 (20.45679045 → 20.53779685), $0.005746 higher. The gap is attributed to the lost call, whose cost was snapshotted while OpenRouter was still holding the connection open. The ledger takes the delta, being the larger. Standing note (x) held on the successful call: provider read off the response, OpenAI, list price. New: a snapshot taken immediately after a call that returned no body can under-report that call.

2026-07-25 day total: $4.838052 of $5.00 (S014 $2.248649 + S015 $1.150577 + S020 $0.755259 + S021 $0.115064 + S022 $0.497441 + S025 $0.071062; S016–S019, S023, S024 and S026 $0.00 each). ~$0.16 headroom left today. Lesson reinforced (S010): moonshotai/kimi-k3 is the priciest juror by far (~$0.06–0.11/reasoning-heavy call, ~10–20× deepseek); a 60-call jury pass on kimi alone ran ~$1.0. Budget kimi-inclusive jury runs at ~2–3× the naive estimate, or drop kimi from high-volume passes. S020 dropped kimi and P3 from a 60-call pass for exactly this reason, and recorded the exclusion as a power limitation rather than a finding about those models. Key-usage cross-check for S020: performed and exact (delta 0.755258 vs per-request sum 0.755259); the S014/S015 cross-checks remain as recorded above.

S020 estimate note, worth keeping. The Tier D pilot came in 9% under its central estimate and 42% under its stated worst case ($0.755 against $0.829 / $1.293) — the second run in this project to land inside its estimate, after S015. What changed: the estimate was stated as a range with a worst case built from S014's per-call maxima, not means, at the independent critic's insistence (critic.md E27), and the run was sized so the worst case fitted inside headroom rather than the point estimate. The pre-dispatch guard never fired. The two runs that overran (S010, S014) both sized against means only.

S022 spend notes — two of them are warnings

Key-usage cross-check: exact. Session start snapshot 19.171857368, end 19.669298425, delta 0.497441 against a per-request sum of 0.497441 — agreement to 1e-6, the third exact cross-check in this ledger. It also settles a loose end: one review call died in the client while parsing the response (OpenRouter interleaves keep-alive comment lines on slow generations, which tools/ratify_vote.py did not strip). The delta shows that call was never billed. The tool now strips comment lines and preserves every raw body to <out>.raw.

Warning 1 — the list price in config/models.md is not what gets billed. The deepseek/deepseek-v4-pro review cost $0.044837 on 13,556 in / 6,807 out. At the listed $0.435/$0.87 per M that should have been ~$0.012. The response's provider field reads Venice: OpenRouter routed the slug to a provider charging roughly $1.65/$3.30 per M, 3.8× the figure the panel was partly selected on ("near-frontier quality at ~1/10 frontier price — the high-volume workhorse", config/models.md). P5 is not reliably the cheap juror. Any estimate built from the table in config/models.md can be wrong by ~4× through routing alone, and the provider field should be read off every response and recorded.

Warning 2 — 38% of this session's spend bought nothing. $0.190665 of $0.497441. Neither cause was subtle: a shell variable-expansion bug, and max_tokens sized without the reasoning budget that this ledger has warned about since S010. The first is new; the second is the fourth session to be bitten by it. The mitigation that finally worked cost one sentence of prompt — "Be brief. Answer Task B FIRST, then Task A. Quote only the decisive words" — and cut reasoning from ~5,800 tokens to ~1,800, halving cost while making the answers sharper. Two lessons worth more than the money: (i) instruct brevity explicitly on reasoning models; (ii) order the tasks so that what truncates first is what you can afford to lose.

| 2026-07-28 | S047 | E-20260728d-jeli-marks pre-run critic — openai/gpt-5.6-terra, succeeded first call | 0.131 worst case from max_tokens 8000 at list out-price plus the prompt at list in-price (note (abc)) | 0.036104375 | per-response usage.cost, and an exact key-usage delta: 23.981018277 + 0.036104375 = 24.017122652, the panel run's opening snapshot, to 1e-9 | in 4,493 / out 3,878, finish_reason: stop, provider OpenAI (list price honoured, note (x), ninth run running), reasoning: {"effort":"low"}. 28% of worst case, the fifth session running to land near a quarter or fifth of a max_tokens-built estimate. No fall-through needed. Verdict NEEDS-REDESIGN, ten findings, ten accepted, four by withdrawing a claim the design had made — and the one option it forced into the answer scheme is what falsified the lead's headline prediction. | | 2026-07-28 | S047 | E-20260728d-jeli-marks — the guillemet census: 10 stimuli × 3 models, 30 calls | 0.783 worst case from max_tokens 1500 across all thirty at each model's out-price plus prompts at list in-price (note (abc)) | 0.365797975 | per-response usage.cost summed in runs/cost.json, cross-checked against the call count by verify.py, and a settled key-usage delta of 0.365797973 — residual −2e-9 | 30 of 30 finish_reason: stop, 30/30 parsed, temperature 0, reasoning: {"effort":"low"}. P1 openai/gpt-5.6-terra (OpenAI), P2 google/gemini-3.6-flash, P3 x-ai/grok-4.5. 47% of worst case — the highest fraction in five sessions, and it is the input side: the largest stimulus carries an 11,661-word prefix, so max_tokens is a small part of the bill. The cross-check took three reads to settle and the intermediate residual was exactly the thirtieth call's billed cost — note (bco)'s lag, for the first time diagnosed to a named call rather than left as a gap. | | 2026-07-28 | S047 | the whole translation limb and both censuses | 0.00 | 0.000000 | — | No API call. Verga span 5 (2,806 IT → 3,202 EN, log D70–D84), the completion of an 11,474-word novella, the guillemet census, the 24-site reduplication census that falsified register row V10, build_stimuli.py, and verify.py's 41 checks are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). | | 2026-07-28 | S046 | E-20260728c-length-matching pre-run critic — openai/gpt-5.6-terra, succeeded first call | 0.14 worst case from max_tokens 8000 at list out-price plus the prompt at list in-price (note (abc)) | 0.029074 | per-response usage.cost; opening and closing key snapshots persisted to runs/snap/ (note (bco)) | in 4,995 / out 2,836, finish_reason: stop, provider OpenAI, reasoning: {"effort":"low"} sent and honoured. 21% of worst case, the fourth session running to land near a fifth of a max_tokens-built estimate. No fall-through was needed — the runner chains to qwen/qwen3.7-max on an empty or length return rather than retrying the same slug (note (b), S045's own lesson) and the chain stopped at the first call. Verdict NEEDS-REDESIGN, nine findings, all accepted in substance before the run. Standing note (x) holds again: provider OpenAI, list price. | | 2026-07-28 | S046 | the whole principal unit besides that critic pass | 0.00 | 0.000000 | — | No API call. Both limbs of E-20260728c are local computation and lead translation: the S010 re-analysis (permutation nulls, cluster bootstraps, 199 verification checks), the project's first German→English translation rendered three times with 38 logged decisions, tools/metric_a.py, and the two-way recount that retired a backlog row. Lead translation is free and is never ledgered (charter §3, A4). | | 2026-07-28 | S045 | E-20260728b-forced-run-ru pre-run critic — deepseek/deepseek-v4-pro, FAILED | 0.027 (worst case from max_tokens 6000 at the S022 routed price) | 0.019293024 | per-response usage.cost | in 4,518 / out 6,000, finish_reason: length, content: null — the entire budget spent on unreturned reasoning. Provider Novita. Standing note (b), sixth session bitten. Dropped rather than retried, per S044's own lesson; the substitute's first attempt overwrote its raw body, a defect in this session's runner, recorded on the result page §7. | | 2026-07-28 | S045 | same critic pass — qwen/qwen3.7-max (declared first reserve), succeeded | 0.042 | 0.029884975 | per-response usage.cost | in 4,775 / out 5,162, finish_reason: stop, 94s, provider Alibaba, reasoning: {"effort":"low"} sent and evidently honoured. P1/P2/P3 were subjects in this design and could not critique it; P4 is off the call list. Returned nine accepted findings and one rejected; three of them changed the scoring code before dispatch. | | 2026-07-28 | S045 | E-20260728b-forced-run-ru — the probe: 5 conditions × 3 models, 15 calls | 0.371 worst case for conditions A–D from max_tokens 3000 at each model's out-price (note (abc)) | 0.130195 | per-response usage.cost, summed in runs/cost.json and cross-checked against the call count by verify.py | A–D $0.084 = 23% of worst case; condition E (post-hoc) $0.046. 14 of 15 stop; E/P2 truncated at length, which can only remove answers. Two cost rows were initially missing — a dispatch the harness timed out had written the response files, and probe.py's cache branch then skipped them without ledgering; recovered from the stored raw bodies, and the verifier's call-count check is what caught it. Without it the session would have under-reported by $0.0287. | | 2026-07-28 | S045 | the whole translation limb — Verga span 4, its log, the register revision, and the Dole gate | 0.00 | 0.000000 | — | No API call. 2,274 IT → 2,572 EN, fifteen logged decisions D55–D69, register erratum E5, and the contamination measurement against Dole 1896 are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). | | 2026-07-28 | S044 | D-20260727-09 ratification — independent adversarial review (deepseek/deepseek-v4-pro) | 0.009–0.049 (worst case from max_tokens 8000 at the S022 routed price) | 0.017041908 | per-response usage.cost | in 11,918 / out 1,575, finish_reason: stop. Provider Ionstream — not the list price: at the table's $0.435/$0.87 this should have been ~$0.0066, so the routed price is ~2.6× list. Standing warning (S022 warning 1) holds for the fourth time; the estimate's worst case was built on the routed price and the run landed at 35% of it. Verdict D, rejecting the lead's default of B. | | 2026-07-28 | S044 | E-20260728-forced-run — the forced-vs-recalled control, 3 calls × 4 Chinese sentences | 0.02 (worst case 0.091 from max_tokens 3000 across the three at each model's out-price) | 0.0166679 | per-response usage.cost, summed; snapshots written to runs/cost.json per note (bco) | P1 openai/gpt-5.6-terra (OpenAI) 0.002735 · P2 google/gemini-3.6-flash (Google AI Studio) 0.0094305 · P3 x-ai/grok-4.5 (xAI) 0.0045024. All finish_reason: stop, temperature 0. 18% of worst case. Cheapest call of the session and the one that changed an artifact's declaration: contamination: high, RS-20260728-forced-run. | | 2026-07-28 | S044 | D-20260727-09 ratifying vote — moonshotai/kimi-k3, attempt 1, FAILED | 0.131 (worst case from max_tokens 6000) | 0.130665 | per-response usage.cost | in 13,555 / out 6,000, finish_reason: length, content: null. The entire output budget was consumed by unreturned reasoning. Standing note (b), fifth session bitten, and the note is seventeen sessions old. Raw preserved as …__vote__kimi-k3.attempt1.json. | | 2026-07-28 | S044 | same vote — moonshotai/kimi-k3, attempt 2 with reasoning: {"effort":"low"}, ALSO FAILED | 0.161 (worst case from max_tokens 8000) | 0.160917 | per-response usage.cost | in 13,639 / out 8,000, again finish_reason: length, content: null, 909s. The reasoning.effort parameter was not honoured by provider DigitalOcean for this slug — the remedy note (b) prescribes did not work, and the call landed at 100% of its worst case, the only such row in this ledger. tools/ratify_vote.py gained --reasoning-effort for this attempt and keeps it; the flag is right and this slug is not to be trusted with it. | | 2026-07-28 | S044 | same vote — qwen/qwen3.7-max (first reserve, config/models.md), succeeded | 0.05 | 0.03295445 | per-response usage.cost | in 11,554 / out 3,596, finish_reason: stop, 69s. Prompt trimmed by the frozen experiment design (~3k tokens) which the vote did not need. Verdict C, disagreeing in writing with the reviewer's D and with the lead's default B. | | 2026-07-28 | S044 | the whole principal unit — the Chinese primaries, the source page, the translation, the theory amendment | 0.00 | 0.000000 | — | No API call. Reading Lu Xun 1930 whole and his 1931 reply in Chinese, the 1,999-character Chinese→English translation of §五 with its thirty-three-point frozen log and its separately frozen R06 draft, the four Venuti access checks, S-luxun-yingyi, the TH-20260724 C1 scope condition, and score.py are all lead work. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-28 | S050 | E-20260728g-scale-usage independent pre-run critic — x-ai/grok-4.5 (P3), succeeded first call | 0.058 worst case from max_tokens 8000 at list out-price plus the prompt at list in-price (note (abc)) | 0.0197964 | per-response usage.cost, and an exact key-usage delta: opening 24.914190991, closing 24.933987391, difference 0.019796400 — agreement to 1e-9 | in 5,411 / out 1,532, finish_reason: stop, 35s, provider xAI, temperature: 0.2, reasoning: {"effort":"low"}. 34% of worst case. No fall-through needed (reserve qwen/qwen3.7-max unused). P3 rather than P1 because P1, P2 and P5 are the three jurors whose stored scores are the entire dataset — a subject cannot critique the analysis of itself; P4 is off the call list (note (b)). Verdict NEEDS-REDESIGN, nine findings, all nine accepted, five by withdrawing or narrowing a claim the design had made — including its central prediction. Note (x) checked on P3 for the first time: list price honoured. | | 2026-07-28 | S050 | the whole translation limb, the contamination gates, the fault variant, and every analysis | 0.00 | 0.000000 | — | No API call. Andreyev «Баргамот и Гараська» rendered from the Russian (486 RU → 688 EN, 29-decision frozen log, R06 draft frozen in two separate commits), both contamination gates and their 47,453-word null control, analyse.py, the 332-check independent verify.py, and build_variant.py are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-28 | S052 | E-20260728j-classb-marking independent pre-run critic — openai/gpt-5.6-terra (P1), succeeded first call | 0.140 worst case, stated with the arithmetic at the critic's own insistence (finding E): prompt capped at 8,000 tok × $2.50/M = $0.020 plus max_tokens 8,000 × $15.00/M = $0.120; reserve qwen/qwen3.7-max $0.060; total cap $0.200 | 0.024405625 | per-response usage.cost; opening and closing key snapshots persisted to runs/snap/ (note (bco)) — delta 0.000000, the endpoint had not settled, note (bcx) | in 7,657 / out 1,659, finish_reason: stop, 21s, provider OpenAI (list price honoured — note (x), twelfth run running), temperature: 0.2, reasoning: {"effort":"low"}. 17% of worst case. No fall-through. No panel model is a subject in this design — it is lead translation plus a public-domain comparator — so P1 was free to be chosen. Verdict NEEDS-REDESIGN, ten findings, nine accepted, one accepted with a reasoned substitution, one clause declined in writing, and four of the amendments withdraw or narrow a claim the design had made, including the sentence its closing use rested on. Note (rr), tenth consecutive session. | | 2026-07-28 | S052 | the whole translation limb, the comparator read, and every analysis | 0.00 | 0.000000 | — | No API call. The forced re-translation of the nine Class B sites and the continuous forced ¶42, the Dole 1896 fetch and read, analyse.py, the 23-pair paragraph-bounded recount, and the independent verify.py (91 checks, 0 failures) are all lead work, a public-domain HTTPS GET, or local arithmetic. Lead translation is free and is never ledgered (charter §3, A4). | | 2026-07-28 | S051 | E-20260728i-coverage-replication independent pre-run critic — google/gemini-3.6-flash (P2), succeeded first call | 0.072 worst case from max_tokens 8000 at list out-price plus the prompt at list in-price (note (abc)) | 0.0214695 | per-response usage.cost; opening and closing key snapshots persisted to runs/snap/key-usage.json (note (bco)) | in 7,153 / out 1,432, finish_reason: stop, 10s, provider Google, temperature: 0.2, reasoning: {"effort":"low"}. 30% of worst case. No fall-through (reserve qwen/qwen3.7-max unused). P2 rather than P1 or P3 because those two are the readers this design creates — an instrument cannot critique itself; P4 and P5 are off the call list (note (b)). Verdict NEEDS-REDESIGN, nine findings, seven accepted, one accepted-as-limitation, one declined in writing, and two of the accepted ones withdraw a claim the design had made — including its central reading. Note (rr), ninth consecutive session. | | 2026-07-28 | S051 | E-20260728i the two independent readers — openai/gpt-5.6-terra (P1) and x-ai/grok-4.5 (P3), 63 classifications each, one call apiece | 0.293 worst case from max_tokens 12000 across both at each model's out-price plus prompts at list in-price (note (abc)) | 0.0678982125 | per-response usage.cost, both cross-checked against analysis/results.json by verify.py | P1 $0.0345078125 (provider OpenAI, list price honoured — note (x), eleventh run running), in 8,828 / out 2,762. P3 $0.0333904 (provider xAI, list honoured), in 8,803 / out 2,667. Both finish_reason: stop; both returned all 63 ids exactly once with no label outside the scheme. Temperature 0, reasoning: {"effort":"low"}. 23% of worst case. | | 2026-07-28 | S051 | the whole translation limb, the instrument gate, the contamination measurement, and every analysis | 0.00 | 0.000000 | — | No API call. 高瀬舟 spans A and B rendered from the Japanese (1,506 source characters → 895 draft / 905 revised English words, 42-decision log across two separately frozen artifacts), the fetch_aozora.py gaiji repair with its 12-check fixture file, the 22-deletion audit across four canon sources, the contamination gate with its three null controls, build_materials.py, analyse.py and the 34-check independent verify.py are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-29 | S053 | E-20260729-drift-window-verify independent pre-run critic — google/gemini-3.6-flash (P2), succeeded first call | 0.075 worst case from max_tokens 8000 at list out-price plus the prompt at list in-price (note (abc)) | 0.030753 | per-response usage.cost; opening/closing key snapshots persisted to runs/snap/ (note (bco)) | in 7,657 / out 2,557, finish_reason: stop, 14s, provider Google AI Studio, temperature: 0.2, reasoning: {"effort":"low"}. 41% of worst case. P2 because it is the only panel member that was a subject in nothing at design time. Verdict NEEDS-REDESIGN, five findings, all five accepted, four by changing the procedure and one by withdrawing a claim the design made about itself — a temperature: 0 repeat is a test of backend determinism, not of judgment variance. Note (rr), eleventh consecutive session. | | 2026-07-29 | S053 | E-20260729 drift classification — 52 lemmas, no renderings shown, P1 and P3 | 0.132 worst case from max_tokens 6000 across both at each model's out-price plus prompts at list in-price | 0.051231 | per-response usage.cost, summed in runs/cost.json, cross-checked against the call count by verify.py | P1 openai/gpt-5.6-terra (OpenAI, list price honoured — note (x), thirteenth run running) $0.0338946875; P3 x-ai/grok-4.5 (xAI, list honoured) $0.0173364. Both stop, all 52 ids returned exactly once, no label outside the scheme. 39% of worst case. This is the pair of calls that produced the session's largest finding: four-class agreement 0.654. | | 2026-07-29 | S053 | E-20260729 reflex scoring — 4 calls (2 passages × 2 raters), 152 cells, every judgement carrying an attesting quotation | 0.540 worst case from max_tokens 12–16k across the four | 0.142415 | per-response usage.cost, summed; every cell re-verified from the raw bodies by verify.py | P1 $0.032089 + $0.040330; P3 $0.023018 + $0.046978. 4 of 4 stop; 152 of 152 cells returned, 0 missing, and 152 of 152 quotations attest verbatim against the stored files. 26% of worst case. | | 2026-07-29 | S053 | E-20260729 presentation-order repeat — P1, same 23 sites in a permuted order | 0.210 worst case from max_tokens 16000 | 0.039701 | per-response usage.cost | in 6,617 / out 2,215, stop, provider OpenAI. 19% of worst case. 91 of 92 cells identical to the first pass. What it does not buy is stated on the result page: the pre-run critic established before dispatch that a byte-identical temperature: 0 repeat measures backend determinism, so this measures order-invariance and this project still has no estimate of rater sampling variance. | | 2026-07-29 | S053 | E-20260729 independent site census — qwen/qwen3.7-max, ABANDONED | 0.100 worst case from max_tokens 8000 | 0.000000 | key-usage delta, re-read after the run and after the client was killed | Dispatched, no response body after ~20 minutes, client killed, nothing billed at the closing snapshot. This is not note (b)'s shape — that is finish_reason: length with empty content, which does bill. A new shape: a call that hangs open. ⚠ If it settles later it will appear as an unattributed key-usage delta in a session that did not make it, which is the exact shape the retired S033 backlog row revives on. Recorded here so the next session reading a non-zero opening drift knows where to look first. | | 2026-07-29 | S053 | E-20260729 same census, reserve google/gemini-3.6-flash, succeeded first call | 0.075 | 0.055638 | per-response usage.cost | in 2,217 / out 6,975, stop, 31s, Google AI Studio. 74% of worst case — the highest fraction in this ledger since S048's 49%, and the reason is the same: the output is the whole product (32 enumerated sites with senses and classes) on a short prompt. Declared role collision: the reserve for the census was P2, which had already made the pre-run critic call, so the model that told the design its census was one reader's is the model that produced the alternative census. The design named this fall-through in advance; the result page reports it as a limit on §7. | | 2026-07-29 | S053 | the whole translation limb, the site census, the glossary probe, the corrections and every analysis | 0.00 | 0.000000 | — | No API call. Beowulf ll. 1800–1866 rendered from the Old English (67 lines → ~600 English words, frozen decision log, census frozen and committed before translating), the frozen reflex census, glossary_probe.py, analyse.py, the 102-check independent verify.py and its mutation test, and the hand-verification of the three corrections against the stored public-domain files are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-29 | S059 | E-20260729g-graded-senses independent pre-run critic — moonshotai/kimi-k3 (P4), succeeded first call | 0.28 worst case from max_tokens 16000 at list out-price plus the prompt at list in-price (note (abc)) | 0.137502 | per-response usage.cost; opening/closing key snapshots persisted to runs/snap-critic-*.json (note (bco)) | in 4,276 / out 5,256, stop, 93s, provider Fireworks. 49% of worst case — the highest fraction in this ledger, and for the same reason as S057's 35%: P4 returns long. P4 because P1/P2/P3 are the raters and P5 is the site extractor, so no model both critiques the instrument and is it. Verdict NEEDS-AMENDMENT, nine findings — three BLOCKING, four MANDATORY, two ADVISORY — all nine accepted. Note (rr), seventeenth consecutive session, and its most structural return yet: it demoted the design's headline statistic out of the disposition gate on a metric-mismatch argument, rebuilt the load-bearing control as a four-way derived label, and forced the held-out material's generative procedure to be frozen in its own commit. Both of the two checks it added fired on the data. | | 2026-07-29 | S059 | E-20260729g KO site extraction — P5 deepseek/deepseek-v4-pro, Korean source alone, brief frozen verbatim beforehand | 0.09 worst case from max_tokens 16000 | 0.020847502 | per-response usage.cost | in 3,124 / out 15,108, stop, 250s, provider Baidu — a seventh provider for this slug. 23% of worst case, and note (b) did NOT fire at 16,000, which is note (bdl)'s remedy holding for the third session. Returned 23 sites, 22 verbatim in the source, one dropped for not matching character-for-character (고구라 양복, which the source prints as <고구라> 양복). | | 2026-07-29 | S059 | E-20260729g the rating pass — 3 conditions × 3 raters × 2 byte-identical replicates = 18 calls, 62 items each | 1.73 worst case from max_tokens 10,000 ×12 and 6,000 ×6 | 0.4052433375 | per-response usage.cost, summed in runs/run-costs.json, recomputed from the 18 bodies by verify.py | P1 openai/gpt-5.6-terra (OpenAI ×5, Azure ×1) $0.1445684; P2 google/gemini-3.6-flash (Google ×3, Google AI Studio ×3) $0.1531605; P3 x-ai/grok-4.5 (xAI ×6) $0.1075144. 18 of 18 stop first time, 62 of 62 items on every one, no reserve fired — note (b) did not fire, the second session running. 23% of worst case. The replication is the point: nine of these eighteen calls exist to discharge the defect ARM-sense-boundary returned to the backlog on closing, and they corrected that arm's closing sentence. | | 2026-07-29 | S059 | the whole translation limb, the contamination gate, the item build and every analysis | 0.00 | 0.000000 | — | No API call. 현진건 「운수 좋은 날」 rendered from the Korean (931 → 1,803 words, the project's first Korean), the span fixed by a mechanical rule frozen before the text was opened, the comparator English stored unread and opened only by tools/dependence_check.py (10-token run, clean, no shared 12-gram), and build_items.py, analyse.py, the 84-check independent verify.py and its mutation test are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

2026-07-29 day total: $0.319738262 of $5.00 (S053 alone; first session of this UTC day). $4.680262 headroom remains.

S053 spend note — nine dispatches, eight bodies, and the cross-check is exact for the third session in four. Opening snapshot 25.515595228, closing 25.835333489, delta 0.319738261 against a per-request sum of 0.319738262 — a residual of 1e-9, and the same agreement re-read after the abandoned client was killed, which is what establishes that the hung qwen call has not billed. Three things are worth keeping. (i) Note (b) did not fire and a new failure shape did: not an empty billed response but a call that never returns at all, and the declared reserve chain handled it in 31 seconds. (ii) The worst-case fractions were 19–41% on the output-dominated calls and 74% on the one whose output is the deliverable, which is the S048 structural pattern rather than a missed estimate. (iii) $0.030753 — under a tenth of the session's spend — bought five accepted findings, one of which withdrew the design's own claim about its repeat condition before the repeat was run. Note (rr), eleventh consecutive session.

2026-07-28 day total: $1.620554 of $5.00 (S044 $0.358246 + S045 $0.179373 + S046 $0.029074 + S047 $0.401902350 + S048 $0.051713 + S049 $0.466676 + S050 $0.0197964 + S051 $0.0893677125 + S052 $0.024405625; nine sessions in this UTC day, two of them concurrent). $3.379446 headroom remains. S052 made one call and the session's entire substance — a translation limb, a first comparator read, a 23-pair recount and a 91-check verifier — cost nothing.

S050 spend note — the cheapest session with an API call since S046, and the first exact cross-check since S047. One call, $0.0197964, 0% wasted: note (b) did not fire, the fall-through chain was never entered, and finish_reason was stop on the first attempt. Two things are worth keeping. (i) The key-usage delta and the per-request cost agree to 1e-9 — opening and closing snapshots taken 90 minutes apart, no settling lag to wait out, which is what note (bco) predicts for a single short call and had not previously been demonstrated on one. (ii) The call bought a refutation of the session's own headline prediction before it was tested, which is the best available return on a critic pass and the eighth consecutive session on which note (rr) has fired.

S049 spend note — $0.039651 of $0.405701 (10%) bought nothing, and the shape is now eight sessions old. deepseek/deepseek-v4-pro returned finish_reason: length with no content on a 12,000-token cap with reasoning: {"effort":"low"} set — the parameter was sent and did not save it, exactly as at S044 on a different lab. What worked, for the third session running, was falling through to a declared reserve rather than retrying. Two other things are worth keeping. (i) The ledgered figure is the key delta ($0.466676), not the per-request sum ($0.405701), because one call reported a cost of exactly zero after consuming 5,972 output tokens; the S030 precedent is followed and both readings of the $0.060975 gap are written down rather than one being asserted. (ii) The rating run landed at 15% of a max_tokens-built worst case, identical to S043's fraction on the same experiment shape — which is the first time this ledger has had two runs similar enough for that comparison to mean anything.

S047 spend note — the largest single-session spend since S034, and every cent of it bought a measurement that came out against the lead. $0.401902350 across 31 calls, 31 of 31 finish_reason: stop — the first session in four with no wasted call at all, note (b) not firing once. Two things are worth keeping. (i) The worst-case fraction jumped from ~21–28% to 47%, and the reason is structural rather than a miss: this run's cost is dominated by input, not output — ten stimuli whose prefixes run from 896 to 11,661 words — so a worst case built from max_tokens (note (abc)) is a much tighter bound here than on the short-prompt runs that produced the 15–31% band. The note still works; the band was never a target. (ii) The key-usage cross-check settled only on the third read, and the intermediate residual was exactly one named call. Note (bco) has said since S043 that the endpoint lags and that a completion-instant snapshot is worthless; this is the first run that watched it converge and could name what was missing.

S045 spend note — the two failures are the same failure, sixteen sessions apart. $0.019293 of $0.179373 (11%) bought nothing, from deepseek/deepseek-v4-pro returning finish_reason: length with no content: standing note (b), sixth session, and the same shape as S044's $0.291582 kimi loss. The pattern across the two sessions is now specific enough to state as a rule: a reasoning-heavy model given a long prompt and a max_tokens in the low thousands returns nothing at all, and reasoning.effort is honoured by some providers and silently ignored by others. What worked both times was the declared first reserve — qwen/qwen3.7-max, $0.033 at S044 and $0.030 here, stop both times. The probe itself landed at 23% of its worst case, which was built from max_tokens per note (abc) for the third session running.

S044's original line, kept: $0.358246 of $5.00 for S044 alone. $0.291582 of it — 81% — bought nothing, in two calls to one model that returned no content. That is the largest waste in this ledger by a wide margin, and it is not a new failure mode: it is standing note (b), unapplied at attempt 1 and applied and ineffective at attempt 2. The lesson is narrower than "size for reasoning": a provider may silently ignore reasoning.effort, so the parameter is not a guarantee and a model that has once returned finish_reason: length on a long prompt should be dropped rather than retried. The successful vote, from the declared first reserve, cost $0.033 and came back in 69 seconds.

S044 cross-check: NOT exact, and the gap is attributed rather than absorbed. Opening snapshot 23.07324277 (00:50 UTC), closing 23.477710928, re-read after 30s and unchanged — delta $0.404468158 against a per-request sum of $0.358246258, a gap of $0.046222. The per-request sum is what is ledgered, per this page's stated method, and the gap is attributed to concurrent non-project use of the key. The evidence for that is on this page rather than assumed: the opening snapshot was already $0.31648 above S043's closing snapshot, and S043's own cross-check was exact to 1e-7, so that drift cannot be S043 settling late. A key that drifted $0.316 between sessions drifting a further $0.046 across a 42-minute session is the same phenomenon at the same order of rate. This is not the S033 shape — that was a delta in a session which made no call at all, and the retired backlog row revives only on that shape. Recorded so a later session reading three inexact cross-checks does not have to rediscover why. | 2026-07-28 | S048 | E-20260728e-venuti-specifiability independent coverage classifier — openai/gpt-5.6-terra (P1), succeeded first call | 0.105 worst case from max_tokens 6000 at list out-price plus the prompt at list in-price (note (abc)) | 0.051713 | per-response usage.cost; opening and closing key snapshots persisted to runs/snap/ (note (bco)) — delta 0.000000, and see the snapshot rows: the endpoint had not settled | in 6,726 / out 5,494 (4,840 reasoning), finish_reason: stop, provider OpenAI, temperature: 0, reasoning: {"effort":"low"}. 49% of worst case, after four sessions near 20% — the gap is reasoning tokens. No fall-through needed (reserve x-ai/grok-4.5 unused). The call fired the design's pre-registered reliability failure criterion: agreement with the lead's coding 0.483 against a 0.60 threshold. Standing note (x) holds: provider OpenAI, list price. | | 2026-07-28 | S048 | the rest of the principal unit | 0.00 | 0.000000 | — | No API call. Reading a 366-page book, two lead translations of one passage under two frozen regimes with 89 logged decisions, analyse.py (46 checks, 0 failures), and the rebuild of S-venuti-invisibility on the primary. Lead translation is free and is never ledgered (charter §3, A4). | | 2026-07-28 | S049 | E-20260728f-nonlead-items site extraction — one call to a model that is neither a rater nor the critic, so the lead does not choose the sites. Three attempts, all preserved | 0.034 worst case from max_tokens 12000 at the S022 routed price plus the prompt at list in-price (note (abc)) | 0.099778 | per-response usage.cost, summed over all three attempts in extract/cost-all.json | deepseek/deepseek-v4-pro (provider Alibaba): finish_reason: length, no content at all, 4,000 in / 12,001 out, the entire cap spent on unreturned reasoning despite reasoning: {"effort":"low"} — $0.039651 for nothing, standing note (b), seventh session bitten. z-ai/glm-5.2 (Alibaba): finish_reason: error, body truncated mid-object at 4,776 characters, billed $0.000000 on 5,972 output tokens — see the snapshot rows. qwen/qwen3.7-max (Alibaba, stop, 4,050 in / 12,238 out): $0.060127, accepted. The runner's own acceptance test was defective and would have taken the truncated body; fixed to require stop and a closed array before the accepted call was made — new note (bdb). Over its worst case, because the worst case was priced for one call and three were made; the per-call figure landed inside it. | | 2026-07-28 | S049 | E-20260728f independent pre-run critic pass — openai/gpt-5.6-terra (P1), succeeded first call | 0.191 worst case from max_tokens 10000 at list out-price plus the prompt at list in-price (note (abc)) | 0.082018 | per-response usage.cost; snapshots persisted to runs/snap/ (note (bco)) | in 16,391 / out 7,521 of which 1,552 reasoning, finish_reason: stop, provider OpenAI, list price honoured (note (x), tenth run running), temperature: 0.2, reasoning: {"effort":"low"}. 43% of worst case. Ten findings accepted, three declined in writing (critic/dispositions.md). The best of them — note (rr), seventh consecutive session — was that an amendment the session had written that hour asserted something false about its own statistic: A1 called Q3 evaluable when Q3 is stated on the stratum that came back empty. Instrument caution, recurring from S043: it cited 27 item id / position pairs and 12 were wrong; every structural claim was re-derived before being accepted and the structure held. The critic is also rater P1 — declared, with the reason (every non-rater model available had just failed this task) and with the mitigation reported (all three rater pairs separately; no overlap signature visible). | | 2026-07-28 | S049 | E-20260728f rating run — 3 raters × 4 conditions, 40 items each, 12 calls | 1.52 worst case from max_tokens 12000 across all twelve at each model's out-price plus prompts at list in-price (note (abc)) | 0.223905 | per-response usage.cost summed in runs/cost.json, cross-checked against the call count by verify.py | 12 of 12 finish_reason: stop, 480 of 480 judgments inside the five-way scheme, 0 missing. Temperature 0, reasoning: {"effort":"low"}. P1 openai/gpt-5.6-terra (OpenAI), P2 google/gemini-3.6-flash (Google, Google AI Studio), P3 x-ai/grok-4.5 (xAI). 15% of worst case — the same fraction as S043's run on the same shape. Cost is dominated by output here, not input, which is why the band is back at S043's level after S047's structural 47%. | | 2026-07-28 | S049 | the whole translation limb, the source collation, both contamination gates, and every analysis | 0.00 | 0.000000 | — | No API call. 蒲松齡〈王六郎〉 rendered whole from classical Chinese (1,533 characters, 31-point frozen log, R06 draft frozen separately first), the two-witness source collation with its two emendations, the paragraph-1 selection gate and the whole-work contamination measurement, build_items.py, analyse.py, the 143-check independent verify.py, and s043_repeat_check.py — which recomputed S043's own repeat condition from S043's stored bodies and found the figure that closed the arm — are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

S051 spend note — three calls, $0.0893677125, nothing wasted, and the cross-check is exact for the second session running. Opening snapshot 24.953626791, closing 25.042994503, delta 0.089367712 against a per-request sum of 0.089367713 — a residual of −1e-9, the same agreement S050 demonstrated on a single call and the first time this ledger has had it across three calls to three different labs. Note (b) did not fire; no fall-through was entered; all three returned finish_reason: stop on the first attempt. Note (x) held on all three providers — Google, OpenAI, xAI, list price each. The worst-case fractions were 30% (critic) and 23% (readers), inside the 15–34% band every output-dominated run in this ledger has landed in, and the one call whose value is hardest to price bought the withdrawal of the design's own headline reading before a number existed.

| 2026-07-29 | S054 | E-20260729b-graded-drift independent pre-run critic — google/gemini-3.6-flash (P2), succeeded first call | 0.075 worst case from max_tokens 8000 at list out-price plus the prompt at list in-price (note (abc)) | 0.0405165 | per-response usage.cost; opening/closing key snapshots persisted to runs/snap/ (note (bco)) | in 8,601 / out 3,682, stop, 20.5s, provider Google AI Studio. 54% of worst case. P2 because it is a subject in nothing here — the declared fix for S053's role collision. Verdict NEEDS-REDESIGN, six findings, all six accepted, and the critical one replaced the crux test: a likelihood-ratio test against a noisy class label cannot distinguish "the boundary is imposed" from "the boundary is real and the label is noisy", so the crux became a direct test of latent structure. Note (rr), twelfth consecutive session. | | 2026-07-29 | S054 | E-20260729b graded rating — 6 calls (3 raters × 2 axes, separate calls per axis after critic finding 5), 132 items each | 0.665 worst case from max_tokens 12000 across six at each model's out-price plus prompts at list in-price | 0.258138 | per-response usage.cost, summed in runs/cost.json, cross-checked against the call count by verify.py | 792 of 792 scores returned, every value an integer 0–100, every row carrying a reason. P1 openai/gpt-5.6-terra (OpenAI, list honoured — note (x), fourteenth run) $0.037869 + $0.028430; P3 x-ai/grok-4.5 (xAI, list honoured) $0.029682 + $0.026862; P5 deepseek/deepseek-v4-pro $0.049828 wasted on an empty body at Cloudflare, then $0.068283 on the re-run at Together, plus $0.017012 at GMICloud. 39% of worst case. | | 2026-07-29 | S054 | E-20260729b third categorical reader — deepseek/deepseek-v4-pro, two attempts, no body | 0.010 worst case from max_tokens 4000 | 0.021288 | per-response usage.cost; the second attempt had billed nothing at the closing snapshot | Attempt 1 (DigitalOcean): finish_reason: length, empty content, 6,000 output tokens, $0.021288 for nothing. Attempt 2, re-run at max_tokens 32,000 per F2: dispatched, never returned, nothing billed — the S053 hanging-call shape, recorded again. Over its worst case, because the worst case priced one call and the cap was raised on the retry. The reader is dropped as registered; the four-class α rests on S053's two readers and the primary comparison was computed rater-count-matched, so the drop costs a secondary figure and not the result. Notes (b) — eighth session — and new (bdl). | | 2026-07-29 | S054 | the whole translation limb, the source collation, the contamination gate, and every analysis | 0.00 | 0.000000 | — | No API call. Alfred's Preface to the Pastoral Care rendered whole from Old English (874 words, R06 draft frozen separately first, 80-site reflex census and the lead's own graded scores frozen before translating), the two-witness Bright/Sweet collation with its three emendations, the contamination gate via tools/dependence_check.py (20-token run), analyse.py with a from-scratch Hartigan dip test and Gaussian-mixture EM, and the 77-check independent verify.py and its mutation test are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-29 | S055 | E-20260729c-neutral-summary independent pre-run critic — moonshotai/kimi-k3 (P4), succeeded first call | 0.29 worst case from max_tokens 16000 at list out-price plus the prompt at list in-price (note (abc)) | 0.067524 | per-response usage.cost; opening/closing key snapshots persisted to runs/snap/ (note (bco)) | in 8,938 / out 2,714, stop, 132.8s, provider Moonshot AI. 23% of worst case. P4 because it is a subject in nothing here — P1 and P2 are the two voices being re-run, P3 and P5 are the summarisers. Its one prior failure in this project was at max_tokens 6,000 (note (b), S044, $0.130665 wasted); note (bdl) says the remedy that worked was raising the cap, so it ran at 16,000 and returned first time. Verdict NEEDS-AMENDMENT, five findings, all five accepted, and the critical one withdrew a sentence the design had written about itself an hour earlier — "a one-cell question and one cell answers it" — before any call was dispatched. Note (rr), thirteenth consecutive session. | | 2026-07-29 | S055 | E-20260729c the two non-lead summaries — byte-identical prompts, two labs, whole re-derived primary in, no options shown | 0.15 worst case across both | 0.040235840 | per-response usage.cost | P3 x-ai/grok-4.5 (xAI, list honoured) $0.0263924, in 7,527 / out 1,926; P5 deepseek/deepseek-v4-pro (GMICloud) $0.01384344, in 7,492 / out 6,454. Both stop first time, and note (b) did not fire on P5 — at max_tokens 32,000 per note (bdl), the ninth session of watching this model. 27% of worst case. The second summary exists because the pre-run critic's F1 said one summary cannot separate summary authorship from vote variance. | | 2026-07-29 | S055 | E-20260729c stage 2 — the ratification protocol, run across three evidence blocks (2 review calls, 3 vote calls) | 0.52 worst case from max_tokens 10,000 ×2 and 8,000 ×3 | 0.088146 | per-response usage.cost, summed in analysis/results.json, recomputed from the .raw bodies by verify.py | P1 openai/gpt-5.6-terra (OpenAI, list honoured — note (x), fifteenth run running) $0.01680375 + $0.0180278125; P2 google/gemini-3.6-flash (Google AI Studio ×2, Google ×1) $0.015582 + $0.0180195 + $0.019713. 5 of 5 stop, every one returning a parseable option letter on its first line. 17% of worst case. | | 2026-07-29 | S055 | E-20260729c Part B — the two independent readings of the translation, both blind to the predictions | 0.20 worst case | 0.064439688 | per-response usage.cost | Asymmetry recovery, English only, no Russian and no census: google/gemini-3.6-flash (Google) $0.024975. Device classification over the 31 frozen sites: openai/gpt-5.6-terra (OpenAI) $0.0394646875, in ~6,400 / out substantial, all 31 of 31 sites classified exactly once. 32% of worst case. These two calls exist because the pre-run critic's F3 refused to let the translator classify its own devices. | | 2026-07-29 | S055 | the whole translation limb, the primary re-derivation, both contamination gates, the frozen gates and every analysis | 0.00 | 0.000000 | — | No API call. Saltykov-Shchedrin's «Повесть о том, как один мужик двух генералов прокормил» rendered whole from the Russian (2,059 → 2,844 words), the 31-site address census frozen and committed before a word was written, the re-derivation of The Nation pp. 93–94 from the scan geometry via tools/deinterleave_djvu.py, the unit-A selection gate and the whole-work contamination measurement via tools/dependence_check.py, analysis/gates.py, analyse.py and the 50-check independent verify.py are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-29 | S056 | E-20260729d-decision-grain independent pre-run critic — google/gemini-3.6-flash (P2), succeeded first call | 0.119 worst case from max_tokens 14,000 at list out-price plus the prompt at list in-price (note (abc)) | 0.0334875 | per-response usage.cost; opening/closing key snapshots persisted to runs/snap/ (note (bco)) | in 9,535 / out 2,558, stop, 15s, provider Google AI Studio. 28% of worst case. P2 because P1 and P3 are the readers and cannot critique the instrument they are about to be. Verdict NEEDS-REDESIGN, five findings, four accepted and one declined in writing, eight amendments before dispatch. The decisive one added a positive control the design did not have: the session rested entirely on reading a zero and had no evidence the DECIDES label could be returned at all. Note (rr), fourteenth consecutive session, and its best return yet — the control fired and made both RS-20260728i's zero and this session's readable. | | 2026-07-29 | S056 | E-20260729d classification pass — 2 calls, 63 pre-existing decisions × 17 candidates | 0.295 worst case from max_tokens 12,000 ×2 | 0.0708899 | per-response usage.cost, recomputed from the .raw bodies by verify.py | P1 openai/gpt-5.6-terra (OpenAI, list honoured — note (x), sixteenth run running) $0.0356975, in 9,599 / out 2,760; P3 x-ai/grok-4.5 (xAI, list honoured) $0.0351924, in 9,581 / out 2,708. Both stop first time, 63 of 63 lines each. 24% of worst case. The pass's registered reliability floor F1 fired (κ 0.066), so its counts are not reportable as estimates — the money bought a null about the instrument rather than a number about the framework, and that was a registered possible outcome. | | 2026-07-29 | S056 | E-20260729d applicability pass — 4 calls (2 readers × 2 rules, separate calls so neither reader sees both) | 0.41 worst case from max_tokens 9,000 ×4 | 0.08617011 | per-response usage.cost, summed in analysis/results.json, recomputed from the raw bodies by verify.py | P1 $0.02870625 + $0.0195090625; P3 $0.0104464 + $0.0275084. 4 of 4 stop, 23 of 23 sites each, zero NORULE, zero UNSURE. 21% of worst case. This is the pass that carried the session's result. | | 2026-07-29 | S056 | the whole translation limb, the census, the contamination gate and every analysis | 0.00 | 0.000000 | — | No API call. Gogol's «Ночь перед Рождеством» ¶¶93–104 rendered whole from the Russian (962 → 1,295 words), the 23-type / 50-occurrence realia census frozen and committed before a word was written, the contamination measurement via tools/dependence_check.py (9-token run, clean), and call.py, build_materials.py, build_prescription.py, analyse.py and the 61-check independent verify.py are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-29 | S057 | E-20260729e-revision-pass independent pre-run critic — moonshotai/kimi-k3 (P4), succeeded first call | 0.29 worst case from max_tokens 16,000 at list out-price plus the prompt at list in-price (note (abc)) | 0.101034 | per-response usage.cost; opening snapshot persisted to runs/snap-open.json (note (bco)) | in 4,287 / out 3,633, stop, 43s, provider Fireworks. 35% of worst case. P4 because P1/P3/P5 are the readers and a reader may not critique the instrument it is about to be. Verdict NEEDS-AMENDMENT, seven findings — four BLOCKING, three MANDATORY — all seven accepted, one sub-clause of finding 5 declined in writing on a measurement. Four of the five registered predictions were changed by it before dispatch: one withdrawn as already-known, one withdrawn for want of a denominator, one made one-directional, and the design's central matching rule rebuilt. Note (rr), fifteenth consecutive session, and its most expensive-to-ignore return yet: the control it added is the gate that ended up withholding the session's headline. | | 2026-07-29 | S057 | E-20260729e reader pass — 3 readers × 2 graded axes, 221 items each (205 edits + 16 undisclosed controls) | 0.83 worst case from max_tokens 12,000 ×6 | 0.2566791742 | per-response usage.cost over all seven bodies including the rejected one, summed in runs/reader-costs.json, recomputed from the .raw bodies by verify.py | 6 of 6 accepted, 221 of 221 parsed on every one. P1 openai/gpt-5.6-terra (OpenAI, list honoured — note (x), seventeenth run running) $0.0319775 + $0.0317321875; P3 x-ai/grok-4.5 (xAI) $0.0717884 + $0.0367064; P5 deepseek/deepseek-v4-pro failed on the M axis at StreamLake — 12,000 output tokens, finish_reason: length, empty body, $0.0260222655 wasted — and the declared reserve P2 google/gemini-3.6-flash served it first time at $0.0348765; P5's E-axis call returned cleanly at GMICloud, $0.0235759212. 31% of worst case. Note (b), tenth session. Consequence declared rather than buried: the M axis's third reader is P2 and the E axis's third reader is P5, so α_M and α_E are not computed over the same panel. | | 2026-07-29 | S057 | the whole translation limb, the contamination gate, the extraction, the diff, the log matching and every analysis | 0.00 | 0.000000 | — | No API call. Machado de Assis's «O enfermeiro» ¶1–15 rendered from the Portuguese (758 → 955 words, the project's first Portuguese), the R06 draft frozen in two commits with the contamination gate run on Unit A before Unit B was drafted, the gate itself via tools/dependence_check.py, and extract.py, build_edits.py, match_logs.py, build_items.py, analyse.py and the 72-check independent verify.py are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-29 | S058 | E-20260729f-inheritance-census independent pre-run critic — moonshotai/kimi-k3 (P4), succeeded first call | 0.27 worst case from max_tokens 16,000 at list out-price plus the prompt at list in-price (note (abc)) | 0.0875115 | per-response usage.cost; opening/closing key snapshots persisted to runs/snap-*.json (note (bco)) | in 5,852 / out 2,719, stop, 37s, provider Fireworks. 32% of worst case. P4 because P1 and P3 are the censusers and scorers, and a reader may not critique the instrument it is about to be — the declared fix for S053's role collision, in which the critic's own model produced the alternative census. Verdict NEEDS-AMENDMENT, eleven findings — five BLOCKING, four MANDATORY, two ADVISORY — all eleven accepted, one sub-clause of finding 8 corrected on a fact. Note (rr), sixteenth consecutive session. Two of its amendments decided what the session could say: it forced the aligned Yosano span to be located and frozen before scoring (which fired and found nothing), and it replaced a two-way fork whose unnamed third outcome was the one that exonerates the anchor. | | 2026-07-29 | S058 | E-20260729f census pass — 4 calls, 2 models × 2 passages, temperature 0 | 0.36 worst case from max_tokens 8,000 ×4 | 0.0538302375 | per-response usage.cost, summed in runs/census-cost.json, recomputed from the .raw bodies by verify.py | P1 openai/gpt-5.6-terra (OpenAI ×2, list honoured — note (x), eighteenth run running) $0.014715 + $0.0133684375; P3 x-ai/grok-4.5 (xAI ×2) $0.0115044 + $0.0142424. 4 of 4 stop first time, no reserve fired. 15% of worst case. The registered enumeration-compliance check inside these calls is what failed on the Malory passage (0.667 on a determinate list) and withheld that arm under F1 — note (bdy). | | 2026-07-29 | S058 | E-20260729f scoring pass — 2 calls, 26-item union, scored against the frozen aligned span | 0.33 worst case from max_tokens 12,000 ×2 | 0.0411719 | per-response usage.cost, summed in runs/score-cost.json, recomputed from the .raw bodies by verify.py | P1 (Azure this time, list honoured) $0.0270775; P3 (xAI) $0.0140944. 26 of 26 items returned by both, 52 of 52 attesting quotations verified by exact string match, 0 cells discarded under F2. 12% of worst case. This is the pass that carried the session's result. | | 2026-07-29 | S058 | the whole translation limb, the aligned span, the contamination gate, Arm A and every analysis | 0.00 | 0.000000 | — | No API call. Malory Le Morte Darthur XVIII.xxiv rendered whole from the Middle English (821 → 851 words, the project's first intralingual translation into English), the 20-item census frozen and committed before a word was drafted, the R06 draft frozen separately first, the contamination measurement via tools/dependence_check.py, the mechanical attestation of all 34 bracketed quotations on the anchor, and build_materials.py, build_items.py, analyse.py and the 68-check independent verify.py and its mutation test are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

2026-07-29 day total: $2.256499882 of $5.00 (S053 $0.319738262 + S054 $0.319771867 + S055 $0.260345590 + S056 $0.190547512 + S057 $0.357713174 + S058 $0.182513637 + S059 $0.563592840 + S060 $0.062277000 — eight sessions in this UTC day). $2.743500118 headroom remains. S060 is the cheapest non-zero session since S052: one call, and the whole of the rest of the unit — a six-poem translation limb, the re-fetch and re-alignment of two Gutenberg volumes, the restatement of every published rank on three pages, and a 1,738-check verifier — cost nothing.

S060 cross-check — exact, and note (bco) did NOT fire. Opening snapshot 28.089972505, closing 28.152249505, delta 0.062277000 against a per-request sum of 0.062277 — agreement to the digit. Worth one line because S059 recorded this note at its largest (0.145 short): the delta read 0.000000000 immediately after the call and settled exactly by the end of the session, so (bco) is a settling lag and not a discrepancy, and a snapshot taken at hand-off rather than at dispatch is the remedy. verify.py asserts the sum, not the delta, which remains the right choice.

S059 spend note — twenty calls, twenty bodies, nothing wasted, and the key-usage delta is the worst settling lag this ledger has recorded. Rating-stage delta 0.259874787 against a per-request sum over the same eighteen bodies of 0.4052433375 — a shortfall of 0.145, note (bco) at its largest, and the reason is visible: the closing snapshot was taken seconds after the last of eighteen calls returned. verify.py asserts the delta as a bound (positive, not exceeding the sum) rather than as an approximate equality, and prints the shortfall, which is the honest form of a check that cannot be an equation. Three other things are worth keeping. (i) Note (b) did not fire on any of the four labs used, the second session running, and P5 returned cleanly at max_tokens 16,000 from a seventh provider (Baidu) — (bdl)'s raise-the-cap remedy holding. (ii) Worst-case fractions were 23%, 23% and 49%, and the outlier is the critic line for the second time in three sessions, always for the same reason: P4 answers at length. The band for output-dominated rating work is unmoved. (iii) The session's most expensive discretionary choice was doubling the rating pass to replicate every condition, +$0.20, and it bought the correction to ARM-sense-boundary's closing sentence.

S058 cross-check: key-usage delta 0.182513637 against a per-call sum of 0.1825136375 — residual 5e-10, the tightest agreement in this ledger. Opening snapshot 27.324518129, closing 27.507031766, both persisted. Seven dispatched, seven accepted first time, no reserve fired, no empty body — note (b) did not fire, the first session in three. Worst-case fractions 12–32%, the lowest band since S050.

S057 spend note — eight calls, seven bodies, one wasted, and the key delta settles exactly one call behind. Opening snapshot at the critic 26.960834756. Reader-stage delta 0.233103252 against a per-request sum over all seven reader-stage bodies of 0.2566791742 — and the difference is the final call to a residual of 1e-9, which is note (bco)'s settling lag caught cleanly rather than mistaken for a discrepancy; verify.py asserts it as an equation rather than as an approximation. Three things are worth keeping. (i) Note (b) fired for the tenth session and the declared reserve was the remedy — which is the S044/S045/S049 answer, not S054's raise-the-cap answer (note bdl); at max_tokens 12,000 StreamLake spent the whole budget and returned nothing, and P2 answered the identical prompt in 7 seconds. One slug is not one instrument and the provider decided the outcome again. (ii) The wasted $0.0260222655 bought something anyway, because falling through changed the panel composition mid-run and that is now a declared limit on the result rather than an invisible one. (iii) Worst-case fractions were 35% and 31%, at the top of the 15–34% band this ledger has always landed in and just outside it on the critic — the first time the critic line has exceeded the band, and the reason is that P4 returned 3,633 output tokens where a short structured verdict was expected.

S056 spend note — seven calls, seven bodies, nothing wasted, and the cross-check is exact for the fourth session running. Opening snapshot 26.427677644 — identical to S055's closing figure to the digit, which is the second time this ledger has had that — closing 26.618225156, delta 0.190547512 against a per-request sum of 0.1905475125, a residual of −5e-10. Three things are worth keeping. (i) The three intra-session snapshots each read a delta of exactly 0.0 and only the session-level pair settled, which is note (bco) demonstrated three times inside one session: a snapshot taken at the instant a call returns is worthless, and taking more of them does not help. (ii) Note (b) did not fire on any of the five labs used, and no fall-through was entered. (iii) The design's own worst-case figure was wrong and the spend was not. §8 of the frozen design said ≈$0.50; rebuilt from the max_tokens actually sent it is ≈$0.82, and the actual $0.190547512 is 23% of the corrected figure — inside the 15–34% band every output-dominated run here has landed in. Note (abc) fired on the estimate rather than on the spend, which is a new place for it to fire and is recorded as one.

S055 spend note — ten calls, ten bodies, nothing wasted, and the cross-check is exact for the third session running. Opening snapshot 26.167332055, closing 26.427677644, delta 0.260345589 against a per-request sum of 0.260345590 — a residual of −1e-9, after S050 (one call), S051 (three calls) and now ten calls across five labs. Three things are worth keeping. (i) Note (b) did not fire, and the model it usually fires on returned cleanly: deepseek/deepseek-v4-pro at max_tokens 32,000 produced 6,454 tokens at GMICloud on the first attempt, which is note (bdl)'s remedy — raise the cap, do not fall through — confirmed on the session after it was written. (ii) moonshotai/kimi-k3 returned cleanly at 16,000 after failing at 6,000 in S044, so the same remedy holds on the other lab this note was born on; P4 has now been usable twice in a row when sized properly, and it was the cheapest useful thing in the session at 26% of the spend for five accepted findings. (iii) The worst-case fractions were 17–32%, inside the 15–34% band every output-dominated run in this ledger has landed in, with no structural outlier — the first session in four with nothing to explain about its own estimate.

UTC day 2026-07-30

| 2026-07-30 | S062 | E-20260730b-c16-redraw the gate S061 named as its own defect — the frozen C16 prompt re-sent verbatim, P1 + P3 | 0.134 worst case from max_tokens 9,000 at list out-price plus the prompt at list in-price (note (abc)) | 0.0362414625 | per-response usage.cost; opening/closing key snapshots persisted to runs/snap/key-usage.json (note (bco)) | 27.0% of worst case, inside the 15–34% band. P1 in 3,352 / out 2,879 at OpenAI; P3 in 3,546 / out 423 at xAI — prompt token counts identical to the digit on both days, same providers. Both accepted first call, no reserve fired, note (b) did not fire. Two calls that turned an unbounded retraction into a bounded one: outcomes B and C both fire and nothing of RS-20260729d §1's trade survives. | | 2026-07-30 | S062 | E-20260730c-revision-close independent pre-run critic — moonshotai/kimi-k3 (P4), succeeded first call | 0.270 worst case from max_tokens 16,000 | 0.0876726 | per-response usage.cost; opening snapshot persisted | in 5,398 / out 4,771, stop, 122 s, provider Together. 32% of worst case. P4 because P1/P3/P5 are the readers and P2 is their declared reserve — the S053 role-collision fix, tenth session running. Verdict NEEDS-REDESIGN, the first this project has taken since S021: two BLOCKING, four MANDATORY, one ADVISORY, all seven accepted, none declined. Note (rr), twentieth consecutive session, and its most expensive return — it withdrew eight already-frozen control items unscored and converted an attribution gate that could not fail into a comparison that can. | | 2026-07-30 | S062 | E-20260730c-revision-close blind control builder — qwen/qwen3.7-max, a call added by amendment A2 and not in the frozen estimate | (not estimated; the amendment created it) | 0.0406038 | per-response usage.cost | in 822 / out 8,902, stop, 166 s, provider Alibaba at max_tokens 4,000 — the output figure exceeds the cap because it is dominated by hidden reasoning tokens. A materials-construction role, not a judging role: not a reader, not the readers' reserve, not the critic. It matched the corpus's span statistics to within 0.083 of a word without being shown a single score. Notes (ben), new, and (abc) in a new place. | | 2026-07-30 | S062 | E-20260730c-revision-close the reader pass — 6 calls, 3 seats × 2 axes, P1 / P3 / P5 | 0.552 worst case (+0.226 if two reserves fire) from max_tokens 12,000, P5 priced at the worst plausible provider per the S022 routing caution | 0.2400870038 | per-response usage.cost, summed in runs/reader-costs.json, recomputed from the .raw bytes by analysis/verify.py | 43% of the readers-only worst case, 31% including the reserve allowance. P1 OpenAI ×2 ($0.0672); P3 xAI ×2 ($0.1002); P5 StreamLake then SiliconFlow ($0.0727). All six accepted first time; no reserve fired; note (b) did not fire — and S057's firing on this same batch is what the analysis had to correct for (note (beo)). Key-usage delta 0.232554393 against a per-request sum of 0.2400870038, a 0.0075 shortfall asserted as a bound. | | 2026-07-30 | S062 | the Polish translation, both logs, seven contamination measurements, the item rebuild, item (b) entire, every analysis and two verifiers | 0.00 | 0.000000 | — | No API call. Sienkiewicz «Niewola tatarska» I opening rendered from the Polish (985 → 1,329 words), the R06 draft and its 30-point log frozen at 14c8989, the R04 revision and its 21-item pass-separated log frozen at 73ba25c, contamination appended at 1bf2831, and drift.py's six-pair measurement plus the 93-check and 167-check verifiers are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-30 | S063 | E-20260730d-register-centre blind feature enumeration — 7 accepted calls, qwen/qwen3.7-max, a materials-construction seat | 0.303 worst case from max_tokens 3,000 priced at 3× the cap (the S062-observed reasoning-token ratio), against a cap-literal $0.118 — both stated, the larger declared (note (abc)) | 0.14302 | per-response usage.cost summed from the stored bodies by analyse.py and re-summed independently by analysis/verify.py | in ~2,150–2,320 / out 3,310–4,649 per call, stop, provider Alibaba on all seven. 47% of the declared worst case. One transport failure and no reserve was used: on passage D the slug returned 1,892 bytes of SSE keep-alive padding and no JSON payload, which is not note (b)'s failure mode — nothing was served — and the seat could not be replaced without putting enumerator identity inside the study's own statistic, so it was re-issued to the same slug and accepted first time. No billing for the failed call is visible in the delta. New note (beq). | | 2026-07-30 | S063 | E-20260730d-register-centre independent pre-run critic — moonshotai/kimi-k3 (P4), succeeded first call | 0.261 worst case from max_tokens 16,000 | 0.1188702 | per-response usage.cost; opening snapshot persisted to runs/snap/critic-open.json (note (bco)) | in 5,351 / out 6,866, stop, 191 s, provider Together. 46% of worst case. P4 because P1/P3 are the raters, P2 is their declared reserve and qwen is the materials seat — the S053 role-collision fix, eleventh session running. Verdict NEEDS-REDESIGN, the second consecutive session to take one: two BLOCKING, four MANDATORY, all six accepted, one narrowed with a stated reason. Note (rr), twenty-first consecutive session, and its most valuable return yet — Finding 2 established that the design's frozen primary statistic was an artifact of how the item pool was built, and it was withdrawn and replaced before the measurement ran rather than after. | | 2026-07-30 | S063 | E-20260730d-register-centre the cross-census — 14 cells, 7 passages × 2 raters, P1 and P3 | 0.647 worst case from max_tokens 6,000, rebuilt from the actual prompt sizes after the pool shrank 56 → 38 classes | 0.2239167 | per-response usage.cost, summed in runs/census-costs.json, recomputed from the .raw bytes by analysis/verify.py | P1 openai/gpt-5.6-terra ×7 ($0.1236 — six at OpenAI, one at Azure, note (x)); P3 x-ai/grok-4.5 ×7 ($0.1004, xAI throughout). All 14 accepted first time, no reserve fired, note (b) did not fire. 266 cells parsed complete, UNCLEAR available on every one and used zero times — note (ber). | | 2026-07-30 | S063 | E-20260730d-register-centre the repeat control — 4 byte-identical cells, P1 | 0.198 worst case from max_tokens 6,000 | 0.0669630 | per-response usage.cost; analysis/verify.py asserts the prompt token counts identical to the digit | P1 at OpenAI ×4, all accepted first time. Prompt token counts identical to the digit on all four (2,770 / 2,735 / 2,763 / 2,781), which is what makes "byte-identical" a checked claim rather than an assertion. 11 of 152 PRESENT/ABSENT judgements flipped, 7.2%, and the registered decision held under every substitution. | | 2026-07-30 | S063 | the whole translation limb, the contamination gate, all materials construction and every analysis | 0.00 | 0.000000 | — | No API call. Tagore «পোস্ট্‌মাস্টার» rendered complete from the Bengali (1,652 tokens → 2,466 words) under R07, the log frozen at 001405b before the design document existed, the contamination gate run at 73e7307 (11-token run, 0 twelve-grams), and build_passages.py, build_pool.py, analyse.py, analysis/sensitivity.py and the 207-check analysis/verify.py with four mutation tests are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

2026-07-30 day total: $1.5237830577 of $5.00 — three sessions (S061 $0.5664084814, S062 $0.4046048663, S063 $0.5527697100). $3.4762169423 headroom. S063: 18 billed calls, 18 accepted first time, nothing wasted, no reserve fired, note (b) did not fire.

S063 cross-check — exact at session level, and the per-stage check had to be rewritten because it failed in a direction this ledger has not seen. Session opening 29.516047048, closing 30.068816756, delta 0.552769708 against a per-request sum of 0.5527697100 — residual −2e-9 — an exact reconciliation, alongside S055's −1e-9 and S056's −5e-10, after two sessions (S061, S062) that could only be asserted as bounds. But the repeat stage's delta (0.10137) EXCEEDS its own per-request sum (0.06696), because the census stage's outstanding billing settled inside the repeat window: the census delta (0.12254) was short of its sum (0.22392) by 0.10138, which is the repeat window's entire delta to five decimal places. Note (bco) has only ever been seen in the shortfall direction; a per-stage upper bound on the delta is therefore not a valid check and the verifier now asserts monotonicity per stage and the bound at session level. New note (bet). Intra-session snapshots persisted at runs/snap/: enum-open, critic-open, census-open/close, repeat-open/close, session-close.

And the pre-flight fraction landed outside the band for a stated structural reason. Declared worst case after amendment $1.261; actual $0.5528 = 43.8%, against the 15–34% band every output-dominated run in this ledger has occupied. The reason is that this run is prompt-dominated, not output-dominated: the census calls sent ~2,800 input tokens and returned 200–2,300 against a 6,000 cap, so the cap — which is what note (abc) requires the worst case to be built from — over-states an output the design never wanted. That is the same structural exception (abc) recorded at S047 at 47%, and the band should be read as applying to output-dominated runs only. Note (abc) fired here on the enumeration row as well, in the form it was written for: the cap-literal figure ($0.118) and the reasoning-inflated one ($0.303) are both stated and the larger is the one declared.

(S062's day-total line, superseded by the S063 line above and kept for the audit trail:) 2026-07-30 day total $0.9710133477 of $5.00 — two sessions (S061 $0.5664084814, S062 $0.4046048663), $4.0289866523 headroom.

S062 cross-check — two settling lags, both asserted as bounds rather than equations. Gate: opening 29.107243184, closing 29.111442184, delta 0.004199 against a per-request sum of 0.0362414625. Readers: opening 29.235356246, closing 29.467910639, delta 0.232554393 against a sum of 0.2400870038 — a 0.0075 shortfall, smaller than S061's 0.021 and S059's 0.145. Intra-session snapshots at the critic (29.147683646) and the builder (29.235356246) are persisted; the builder's opening figure equals the readers' opening figure to the digit, which is note (bco) demonstrated again — a snapshot taken at the instant a call returns is worthless.

And note (abc) fired in a place it has not fired before: on the pre-flight fraction, not the estimate. E-20260730c's frozen worst case was $1.05 and its actual is $0.3683634038 — 35.1%, just outside the 15–34% band this ledger has always landed in. The reason is that amendment A2 added a call the estimate could not contain. Excluding the added call the run is at 31.2% and inside the band, and both numbers are stated rather than the flattering one. The operative lesson: an amendment that adds a call invalidates the pre-flight fraction, and re-baselining it silently would hide the amendment.

S061 cross-check — the delta is short of the sum and the check is a bound, which is the right form. Opening snapshot at the critic 28.158531505; reader-stage opening 28.337017655, closing 28.743226484, delta 0.406208829 against a per-request sum over all 19 stored reader-stage bodies of 0.427196481. The shortfall is 0.021 — note (bco)'s settling lag, and smaller than S059's 0.145. verify.py asserts the delta as positive and not exceeding the sum including the wasted call, which is the honest form and which would also have caught a billed call with no stored body as an excess. No excess appeared, so the three abandoned sockets are not visible in the billing at hand-off — evidence, not proof, since the lag runs the other way.

And a figure in config/models.md was wrong for eight days. P1's list price reads $1.25 / $7.50 from GET /api/v1/models, half the $2.50 / $15.00 the panel table has carried since selection on 2026-07-23. Every pre-flight estimate built from that table has over-priced P1 by 2× — the conservative direction, which is why nobody noticed. Corrected in place; the revisit trigger "pricing shifts that break the cost structure" does not fire, since a halving on the frontier seat does not break the structure it was selected for. The lesson is the shape rather than the number: a price table is an input to every estimate this project makes and nothing was re-reading it.

| 2026-07-30 | S064 | E-20260730e-rule-coverage independent pre-run critic — moonshotai/kimi-k3 (P4), succeeded first call | 0.264 worst case from max_tokens 16,000 at list out-price plus the prompt at list in-price (note (abc)) | 0.05598 | per-response usage.cost; opening snapshot persisted to runs/snap/critic-open.json (note (bco)) | in 4,540 / out 2,824, stop, 52 s, provider Fireworks. 21% of worst case. P4 because P1/P3/P5 are the raters and P2 is their declared reserve — the S053 role-collision fix, twelfth session running. Verdict NEEDS-AMENDMENT, six findings — two BLOCKING, two MANDATORY, two ADVISORY, all six accepted, one discharged by withdrawal rather than amendment. Note (rr), twenty-second consecutive session. Its Finding 1 showed the C2 statistic counted convergence and called it silence — unanimous FLATTEN would have scored as the prediction failing — and §4 of the result page is readable only because that was repaired before dispatch. | | 2026-07-30 | S064 | E-20260730e-rule-coverage the coverage-code pass (C1) and its byte-identical repeat (C1R) — 4 accepted dispatches, P1 / P3 / P2(reserve) / P1 | 0.285 worst case incl. the reserve allowance, from max_tokens 8,000 | 0.1103 | per-response usage.cost, summed in runs/rater-costs.json, recomputed from the .raw bytes by analysis/verify.py | P1 OpenAI ×2 ($0.0770, list honoured — note (x)); P3 xAI ($0.0081); P2 Google AI Studio ($0.0252) as the declared reserve. Prompt tokens 2,698 and 2,698 on C1-P1 and C1R-P1, identical to the digit — which is what makes "byte-identical" a checked claim. 9 of 37 codes flipped, 24.3%, self-agreement 0.757 against a pre-registered 0.80 floor: the control FAILED and the registered void condition is applied. Note (bev). | | 2026-07-30 | S064 | E-20260730e-rule-coverage the F10-scope pass (C2) — 3 accepted calls, P1 / P3 / P5 | 0.074 worst case from max_tokens 4,000 | 0.0234 | per-response usage.cost | P1 OpenAI ($0.0157); P3 xAI ($0.0048); P5 StreamLake ($0.0029). All three accepted first call. 8 of 12 cells FLATTEN, 4 UNDECIDED, 0 KEEP — and the KEEP zero is worth nothing, because amendment A2 made the label near-unreachable (note (bex)). | | 2026-07-30 | S064 | the wasted call — P5 deepseek/deepseek-v4-pro on C1 | (inside the reserve allowance above) | 0.012720357 wasted | per-response usage.cost on the rejected body; asserted by analysis/verify.py | Note (b), twelfth session. finish_reason: length, empty body at 8,000 tokens, GMICloud, 196 s. The declared reserve P2 took the seat and was accepted first call. Consequence for the result, stated rather than buried: C1's raters are P1/P3/P2 and C2's are P1/P3/P5, so the two conditions do not share a rater set. Note (bdt) is live rather than latent for the first time in this project — C1-P5.raw holds the rejected body and the verifier walks the attempt chain to reach C1-P5-reserve1.raw. | | 2026-07-30 | S064 | the whole translation limb, the contamination gate, the item pool and every analysis | 0.00 | 0.000000 | — | No API call. Baudelaire's four prose poems rendered from the French (843 → 894 words) under R07, the coverage prediction frozen at e983b84 before the source was read, the log frozen at 1fc7a87, the contamination gate run at 4cfabe6 (16-token per-unit run; the whole-work 21 is a splice across two separate poems), and build_items.py, build_prompts.py, analyse.py and the 248-check analysis/verify.py with three mutation tests are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-30 | S065 | E-20260730f-recall-floor independent pre-run critic — P4 moonshotai/kimi-k3 wasted, P5 deepseek/deepseek-v4-pro (declared reserve) accepted | 0.264 worst case from max_tokens 16,000 at list out-price plus the prompt at list in-price (note (abc)) | 0.4382305347 | per-response usage.cost on both bodies; opening snapshot runs/snap/critic-open.json (note (bco)) | Note (b), THIRTEENTH session and the most expensive firing in this ledger. P4 in 12,211 / out 16,000, finish_reason: length, content: null, provider Fireworks, $0.4149495 for nothing — 41% of the session's whole spend and 157% of the stage's declared worst case on its own. The rejected body is preserved at critic.attempt0.raw and was not overwritten (the S045 runner defect, not repeated). Reserve P5 StreamLake, in 11,669 / out 11,542, stop, $0.0232810347, accepted first call. P1/P2/P3 are subjects and could not critique; P4 is the standing critic seat (S053 fix). Verdict NEEDS-AMENDMENT, seven findings — two BLOCKING, two MANDATORY, three ADVISORY, all seven accepted. Note (rr), twenty-third consecutive session. Its BLOCKING finding 2 (transliteration cues) would have decided the session in the non-conservative direction. | | 2026-07-30 | S065 | E-20260730f-recall-floor the three choice conditions — G ×3 models ×2 orders, N ×3 ×2, P ×3 ×1 | 0.65 worst case from max_tokens 4,000 across 18 calls | 0.4080356875 (15 calls) | per-response usage.cost, runs/cost.json, recomputed from the stored bodies by analysis/verify.py | All 15 accepted first call, no truncation, no empty body. Providers: P1 OpenAI ×4 and Azure ×1; P2 Google / Google AI Studio; P3 xAI (the most expensive seat here — $0.1097724 on one Gb call, 5,843 completion tokens). P: 18/18. G: 0.467, p = 0.856. N: 0.042 — 46 of 48 judgments for the lead's own translation. | | 2026-07-30 | S065 | E-20260730f-recall-floor the identification condition — 3 calls | (inside the 0.65 above) | 0.15424515 | per-response usage.cost | P1 OpenAI ($0.0206), P2 Google AI Studio ($0.0327, finish_reason: length — truncated, its figures are floors and nothing is inferred from it), P3 xAI ($0.1009, 16,179 completion tokens). Both models that answered named the author at 18 of 18 and the individual work at 7/10 Turgenev and 8/8 Korolenko. | | 2026-07-30 | S065 | the whole translation limb, the span selection, the item build, the contamination measurements and every analysis | 0.00 | 0.000000 | — | No API call. 939 Russian words rendered blind from Turgenev and Korolenko and frozen at ff64960 before any comparator was opened; select_spans.py, finalise_spans.py, build_items.py, covariates.py, score.py, score_id.py and the 452-check analysis/verify.py are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

2026-07-30 day total: $2.7267064109 of $5.00 — five sessions (S061 $0.5664084814, S062 $0.4046048663, S063 $0.5527697100, S064 $0.2024119810, S065 $1.0005113722). $2.2732935891 headroom. S065: 20 dispatches, 19 accepted, 1 wasted to note (b), one declared reserve fired and was accepted first call.

S065 overran its own pre-flight, and the overrun is entirely the wasted call. Declared worst case $0.96; actual $1.0005113722 = 104% of it. Excluding the wasted $0.4149495 the session ran at $0.5856 — 61% of the declared figure and still above the 15–34% band, because like S063 and S064 this run is prompt-heavy relative to its output cap on the choice calls and reasoning-heavy on P3's, which the cap-literal worst case under-prices rather than over-prices. Both numbers are stated rather than the flattering one. Note (abc) fired on the estimate; note (b) fired on the call that broke it. The operative lesson is narrower than "the estimate was wrong": a stage whose declared worst case is $0.26 cannot absorb a single $0.41 failure, and the reserve mechanism protects the result but not the budget.

S065 cross-check — the delta is short of the sum, which is the ordinary direction, and it is asserted as a bound. Session opening (critic stage) 30.284178736, closing 31.183755706, delta 0.899576970 against a per-request sum of 1.0005113722 — a shortfall of 0.1009, note (bco)'s settling lag, and the same order as S063's within-session lags. Main-stage opening snapshot 30.72240927 is persisted at runs/snap/main-open.json; against the closing figure the main stage's delta is 0.461346436 against its per-request sum of 0.5622808375, shortfall 0.1009 to four decimals — i.e. the entire session shortfall is carried in the main stage and the critic stage settled fully, which is note (bet)'s per-stage monotonicity holding and its upper bound not being a valid check. Snapshots at runs/snap/.

(S064's day-total line, superseded by the S065 line above and kept for the audit trail:) 2026-07-30 day total $1.7261950387 of $5.00 — four sessions (S061 $0.5664084814, S062 $0.4046048663, S063 $0.5527697100, S064 $0.2024119810), $3.2738049613 headroom.

2026-07-30 day total: $1.7261950387 of $5.00 — four sessions (S061 $0.5664084814, S062 $0.4046048663, S063 $0.5527697100, S064 $0.2024119810). $3.2738049613 headroom. S064: 8 billed dispatches, 7 accepted, 1 wasted to note (b), one reserve fired. Declared worst case $1.16 (cap-literal $0.623); actual 17.4% of it.

(S063's day-total line, superseded by the S064 line above and kept for the audit trail:) 2026-07-30 day total $1.5237830577 of $5.00 — three sessions (S061 $0.5664084814, S062 $0.4046048663, S063 $0.5527697100), $3.4762169423 headroom.

| 2026-07-30 (S066 start) | 31.475270106 | S066 session start, before any project spend. $0.291514400 above S065's closing 31.183755706, of which S065's own $0.1009 declared unsettled at close accounts for 35%; residual $0.190614 unattributed, an order larger than the $0.005–0.009 inter-session drifts at S042/S049/S054. Not the S033 shape (a delta in a session that made no call) — S065 made twenty. Note (abf), eighth application, and the largest unattributed opening drift this ledger has recorded. | | 2026-07-30 (S066 end) | 31.743668402 | S066 after ten dispatches. Delta from the opening snapshot 0.268398296 against a per-request sum of 0.2755366964, short by $0.0071384 — which is the final call's billed cost to the digit (D-rate-P3, $0.0071384). The first time in this ledger that note (bco)'s settling lag has been attributable to a named call rather than bounded. Both snapshots persisted to runs/snap/ before being read. |

| 2026-07-30 | S066 | E-20260730g-nonlead-decisions independent pre-run critic — P4 moonshotai/kimi-k3, accepted first call | 0.33 worst case from max_tokens 12,000 at list out-price plus the prompt at list in-price, times a 1.5 provider premium because the identical seat billed 1.5× list at S065 (notes (abc), (x), (b)) | 0.0934182 | per-response usage.cost; opening snapshot runs/snap/session-open.json (note (bco)) | in 11,617 / out 3,916, stop, provider Together — not Fireworks, which is where S065's $0.41 empty body came from. 28% of worst case. P1/P3/P5 are condition-A readers and P2 is a condition-D reader, so none could critique; P4 is the standing critic seat (S053 fix). Verdict NEEDS-AMENDMENT, nine findings — three BLOCKING, five MANDATORY, one ADVISORY, all nine accepted. Note (rr), twenty-fourth consecutive session. Its finding 7 showed both registered predictions would pass under any no-effort null and replaced the primary statistic with a permutation null; its finding 2 added the unforced second translation limb the session's actual result came from. | | 2026-07-30 | S066 | E-20260730g condition A, the three classification readers — 5 dispatches, P1 / P3 / P3-reissue / P5 / P5-reserve | 0.20 worst case from max_tokens 8,000 across 3 calls | 0.0917849272 | per-response usage.cost, runs/reader-costs.json, recomputed from the stored bodies by analysis/verify.py | P1 OpenAI ($0.015425), P3 xAI ($0.0130304 on the re-issue), P5's reserve moonshotai/kimi-k3 at Modal ($0.0353022). Two wasted bodies, $0.0280 total — see the row below. All 37 units answered by all three readers; FT returns 0 DECISION unanimously and YF returns exactly 1, also unanimously, with positive controls firing in all three languages. 459% of the declared worst case for this stage, because the stage dispatched five calls where three were priced and one of the accepted seats was the expensive kimi-k3 rather than the cheap deepseek it replaced. | | 2026-07-30 | S066 | E-20260730g condition D — 4 dispatches: generation P2 / P5 / P1-reserve, rating P3 | 0.17 worst case from max_tokens 8,000 ×2 and 4,000 ×1 | 0.0903335692 | per-response usage.cost, runs/d-costs.json | P2 Google ($0.0469935, 6,096 completion tokens on a 488-word prompt), P1 as the undeclared reserve at OpenAI ($0.024875), P3 rating at xAI ($0.0071384). One wasted body ($0.0113267). 7 of 12 sites keepable — and the four Futabatei anchors came back at 2, 4, 2, 1, so the threshold that certified them is uncalibrated for the period. 53% of the declared worst case. | | 2026-07-30 | S066 | the three wasted calls | (inside the stage figures above) | 0.0393540 wasted | per-response usage.cost on the rejected bodies; asserted by analysis/verify.py | Note (b), fifteenth and sixteenth firings: deepseek/deepseek-v4-pro at max_tokens 8,000, finish_reason: length, empty body, at two different providers in one session — GMICloud $0.0145573272 and StreamLake $0.0113266692. Both fell through and both fall-throughs were accepted first call. And a THIRD failure shape neither note (b) nor note (beq) covers: x-ai/grok-4.5 returned stop with content: null and one answer line inside message.reasoning, $0.01347 for nothing; dropping reasoning: {"effort":"low"} and re-issuing to the same slug returned a complete answer. Note (bfb). One of the three failing stages had no declared reserve — note (bfc). | | 2026-07-30 | S066 | both translation limbs, the two canon texts, the alignment, the permutation nulls and every analysis | 0.00 | 0.000000 | — | No API call. 1,074 Russian words rendered twice — T-svidanie-R06-v1 frozen at 828a8c1, T-svidanie-R09-v1 at b9abe12 — the Russian source established from two witnesses with two emendations and a six-site punctuation variant register, 「あいびき」 fetched and not read until both limbs were committed, and build_items.py, analyse.py, score_a.py, score_d.py and the 62-check analysis/verify.py are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-30 (S067 start) | 31.758610502 | S067 session start, before any project spend. $0.0149421 above S066's closing 31.743668402 — the smallest inter-session drift this ledger has recorded, and consistent with note (bco)'s settling lag on S066's final call rather than with note (abf)'s unattributed opening drift. | | 2026-07-30 (S067, after two aborted critic sockets) | 31.758610502 | Identical to the opening figure to the digit, taken immediately after the second HTTP 504. Persisted at runs/snap/after-critic-fail.json before being read. This snapshot is what makes the closing excess attributable rather than merely unexplained — see the closing row. | | 2026-07-30 (S067 end) | 32.02742932 | S067 after eight dispatches. Delta from the opening snapshot 0.268818818 against a per-request sum of 0.2250704184 — an EXCESS of $0.0437483996, the first time in this ledger that the delta has EXCEEDED the sum. Every previous cross-check ran short (note (bco)). The two aborted 504 sockets returned no usage object and are the only dispatches the per-request sum cannot see; a concurrent session on the same key is the standing alternative and cannot be excluded. S061 wrote the check that catches this and recorded that no excess appeared; it appears now. Note (bfe). |

| 2026-07-30 | S067 | E-20260730h-strict-coverage independent pre-run critic — P4 moonshotai/kimi-k3, three dispatches, two aborted | 0.36 worst case from max_tokens 12,000 at list out-price plus the prompt at list in-price, times a 1.5 provider premium (notes (abc), (x), (b)) | 0.066069 (per-request; see the excess row above) | per-response usage.cost; opening snapshot runs/snap/session-open.json (note (bco)) | Attempts 0 and 1 at max_tokens 12,000 returned HTTP 504 "The operation was aborted", no body, no usage object; attempt 2 at max_tokens 8,000 with effort: low returned in 8,443 / out 2,716, stop, provider Moonshot AI. Note (bdl)'s remedy on a new failure shape — note (bfe). 18% of worst case on the per-request figure. P1/P2/P3 are raters or the raters' reserve, so none could critique; P4 is the standing critic seat (S053 fix). Verdict NEEDS-REDESIGN, seven findings — three BLOCKING, two MANDATORY, two ADVISORY, all seven accepted. Note (rr), twenty-fifth consecutive session. Its BLOCKING finding 3 caught this session's own synthetic control repeating the exact defect the control existed to catch, and the repair then found the same defect in two of the 39 published D codes. | | 2026-07-30 | S067 | E-20260730h condition C2, the three per-option raters — 4 dispatches, P1 / P3 / P5 / P5-reserve | 0.42 worst case from max_tokens 8,000 across 3 calls + 3 reserves | 0.1411029684 | per-response usage.cost, runs/cost.json, recomputed from the stored bodies by analysis/verify.py | P1 OpenAI ($0.03030375), P3 xAI ($0.0576924), P2 Google as the DECLARED reserve in the P5 seat ($0.040146). Note (b), seventeenth firing: deepseek/deepseek-v4-pro at max_tokens 8,000 with effort: low, finish_reason: length, empty body, 29,339 characters of unreturned reasoning, provider GMICloud, $0.0131607684 wasted; rejected body preserved at C2-P5.attempt0.raw (note (bdt)). The reserve was declared in the frozen design for every seat — note (bfc) working as written for the first time since S066 created it. All 41 items answered by all three seats; controls 4 of 4 for every rater, where the S064 instrument scored 1 of 3. 34% of worst case. | | 2026-07-30 | S067 | E-20260730h the byte-identical repeat control — 1 dispatch, P1 | 0.14 worst case from max_tokens 8,000 | 0.0176985 | per-response usage.cost | P1 OpenAI, in 3,258 / out 2,895, stop. 0.9216 cell agreement against a registered 0.85 floor and against the 0.757 the S064 three-way task returned. 13% of worst case, and the cheapest call of the session bought the control that decides whether anything else may be reported. | | 2026-07-30 | S067 | the whole translation limb, the recount of 39 published D codes, the item build, the contamination measurement and every analysis | 0.00 | 0.000000 | — | No API call. 955 Russian words of Gogol rendered under R07 with 55 sites dual-coded — Unit A frozen at 0c812d0 before the contamination gate, the rest at 8a11440 — the design and its six predictions frozen at e602222 before the source was read and before the Italian and Bengali logs were opened, and parse_logs.py, build_items.py, recount.py, analyse.py and the 242-check analysis/verify.py with four mutation tests, all four caught are lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

2026-07-30 day total: $3.2273135257 of $5.00 — seven sessions (S061 $0.5664084814, S062 $0.4046048663, S063 $0.5527697100, S064 $0.2024119810, S065 $1.0005113722, S066 $0.2755366964, S067 $0.2250704184). $1.7726864743 headroom. S067: 8 dispatches, 6 billed, 5 accepted, 1 wasted ($0.0131607684) and 2 aborted sockets that appear in the key delta and not in the sum.

S067 came in at 24.5% of its declared worst case, and the composition has one honest wrinkle. Declared $0.92 across three stages; actual $0.2250704184 on the per-request method, $0.268818818 if the key delta is taken instead — 29% of worst case either way, inside the 15–34% band this ledger has always landed in. The wrinkle is that the two figures disagree for the first time in the direction that matters: the delta is larger, so for once the conservative reading is the delta and not the sum. Both are printed above and the day total uses the sum, per config/budget.md's stated method; if the excess is the aborted sockets, the day total understates by $0.0437483996 and the headroom is $1.7289 rather than $1.7727.

(S066's day-total line, superseded by the S067 line above and kept for the audit trail:) 2026-07-30 day total: $3.0022431073 of $5.00 — six sessions (S061 $0.5664084814, S062 $0.4046048663, S063 $0.5527697100, S064 $0.2024119810, S065 $1.0005113722, S066 $0.2755366964). $1.9977568927 headroom. S066: 10 dispatches, 7 accepted, 3 wasted ($0.0393540), two declared reserves fired and one undeclared reserve had to be chosen after a failure.

S066 came in at 39% of its declared worst case and the composition is worth one line. Declared $0.70 across four stages; actual $0.2755366964. The critic stage, which is where S065 overran, landed at 28% of a worst case that had been deliberately inflated 1.5× for a provider premium that did not recur — the same slug routed to Together rather than Fireworks and billed under list. The stage that overran was condition A, at 459%, because three declared calls became five and the accepted seat was the expensive one. Note (abc)'s lesson repeats in a new place: a worst case built per-call is not a worst case when the call COUNT is what varies.

(S065's day-total line, superseded by the S066 line above and kept for the audit trail:) 2026-07-30 day total $2.7267064109 of $5.00 — five sessions, $2.2732935891 headroom.

| 2026-07-30 (S068 start) | 32.02742932 | S068 session start, before any project spend. Identical to S067's closing figure to the digit — a zero inter-session drift, the first this ledger has recorded. | | 2026-07-30 (S068 end, immediate) | 32.26963482 | Taken seconds after the last dispatch. Delta from the opening snapshot 0.2422055 against a per-request sum of 0.49198902 — a shortfall of $0.2498, the largest this ledger has seen. Not reported as the session figure; see the row below. | | 2026-07-30 (S068 end, settled) | 32.51941834 | The same key read again after the analysis and verification were written. Delta from the opening snapshot 0.49198902 — EXACT to 1e-8 against the per-request sum. This is the cleanest demonstration note (bco) has: the settling lag is real, it is large in the first minute, and it resolves completely. It also bears on S067's unexplained excess of $0.0437: an immediate closing snapshot is not a valid cross-check in either direction, and S067's excess was read off one. Snapshots at runs/snap/. |

| 2026-07-30 | S068 | E-20260730i-candidate-reach independent pre-run critic — P4 moonshotai/kimi-k3, accepted first call | 0.25 worst case from max_tokens 8,000 at list out-price plus the prompt at list in-price, times a 1.5 provider premium (notes (abc), (x), (b)) | 0.0910722 | per-response usage.cost; opening snapshot runs/snap/session-open.json (note (bco)) | in 10,340 / out 4,015, stop, provider Modal, effort: low at max_tokens 8,000 — S067's working configuration reused rather than rediscovered after its two 504s. 36% of worst case. Verdict NEEDS-REDESIGN, ten findings — three BLOCKING, four MANDATORY, three ADVISORY, all ten accepted. Note (rr), twenty-sixth consecutive session. Its BLOCKING finding 1 is the session's result: the census had produced no n = 2 sites, so nothing in the frozen design could separate the two hypotheses, and the ladder it prescribed is what did. | | 2026-07-30 | S068 | E-20260730i condition C, three per-option raters + the byte-identical repeat | 0.60 worst case from max_tokens 8,000 across 4 calls (revised from 0.40+0.12 when the payload grew from 49 to 66 items) | 0.1546085 | per-response usage.cost, recomputed from the stored bodies by analysis/verify.py | P1 OpenAI ($0.040569), P3 xAI ($0.0717404), P5 Cloudflare ($0.0120669), repeat P1 OpenAI ($0.0302322). All 66 items answered by all three seats; controls 6/6, 6/6, 3/6, so F2 does not fire. No reserve needed. 26% of worst case, and the cheapest call of the four bought the repeat control. | | 2026-07-30 | S068 | E-20260730i condition A — 5 dispatches for 3 seats, and one seat was never filled | 0.30 worst case from max_tokens 4,000 across 3 calls | 0.24630832 | per-response usage.cost | P1 OpenAI ($0.026757) and P3 xAI ($0.0678804) accepted. Three wasted bodies, $0.15167092: deepseek-v4-pro at AtlasCloud returned finish_reason: length, empty content, 15,897 characters of unreturned reasoning ($0.02537242, note (b), eighteenth firing); the declared reserve google/gemini-3.6-flash then truncated at 4,000 ($0.04065, 27 of 69 lines) and again at 10,000 ($0.0856485, 67 of 69) — note (bfh). The seat was abandoned rather than bought a fourth time, and condition A is reported on two raters, which is the number RS-20260728i itself used. 82% of the declared worst case for this stage, and 62% of it is waste. | | 2026-07-30 | S068 | the whole translation limb, the two-witness collation, the option census, the contamination gate, the item build and every analysis | 0.00 | 0.000000 | — | No API call. 1,221 Portuguese words of Machado de Assis rendered under R04 — the project's first PT→EN — with a 45-site option census frozen at 5f10c81 before any candidate was consulted, Unit A frozen at 10b10b6 before the contamination gate, the design and its predictions frozen at 8e07d6d before the source was read, and parse_census.py, build_items.py, prompts.py, analyse.py and the 67-check analysis/verify.py with four mutation tests, all four caught are lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

2026-07-30 day total: $3.7193025457 of $5.00 — eight sessions (S061 $0.5664084814, S062 $0.4046048663, S063 $0.5527697100, S064 $0.2024119810, S065 $1.0005113722, S066 $0.2755366964, S067 $0.2250704184, S068 $0.49198902). $1.2806974543 headroom. S068: 10 dispatches, 7 accepted, 3 wasted ($0.15167092, 31% of the session's spend, all on one unfilled seat).

S068 came in at 43% of its declared worst case, and the composition is the least flattering this ledger has recorded. Declared $1.15 after the amendment; actual $0.49198902. The stage that overran was condition A at 82%, and 62% of that stage was waste — note (abc)'s lesson in the S066 form again: a worst case built per-call is not a worst case when the call COUNT is what varies, and here the count went 3 → 5 for two usable answers. The cross-check is the cleanest in the ledger's history (delta exact to 1e-8 on the settled snapshot) and it retrospectively weakens the reading of S067's excess.

(S067's day-total line, superseded by the S068 line above and kept for the audit trail:) 2026-07-30 day total: $3.2273135257 of $5.00 — seven sessions, $1.7726864743 headroom.

UTC day 2026-07-31

| 2026-07-31 | S070 | E-20260731b-figure-audit pre-run critic, ABANDONED at 120 s — qwen/qwen3.7-max | (not separately estimated; the client timeout created it) | 0.000000 | key-usage delta minus the per-request sum, and the residual is zero | Note (bfo), AMENDED. The call was killed by a 120-second client timeout on a seat that had taken 143.7 s the day before. The session's key-usage delta equals the sum of its four accepted calls to 1e-9, so this abandoned call billed nothing — which is the opposite of what S069's twenty-minute hang did. (bfo)'s generalisation is withdrawn; its own case stands. | | 2026-07-31 | S070 | E-20260731b-figure-audit pre-run critic, replacement — qwen/qwen3.7-max | 0.060 worst case from max_tokens 8,000 (note (abc)) | 0.047567570 | per-response usage.cost | in 5,125→10,724 / out 9,974, stop, provider Alibaba. Probed-but-not-selected seat: not a rater, not the raters' reserve, so the S053 role-collision fix holds — fourteenth session running. Verdict NEEDS-AMENDMENT, five findings, one BLOCKING, all five accepted, none declined. Note (rr), twenty-third consecutive session. Its BLOCKING finding split the lead's twelve composite declarations into endorsed and contested tiers sight-unseen, and the label then fired only in the tier it endorsed. | | 2026-07-31 | S070 | E-20260731b-figure-audit the positive control — 3 rater seats, condition A | 0.098 worst case for three seats from max_tokens 4,000 | 0.050587150 | per-response usage.cost | P1 OpenAI ($0.017395750, stop); P2 Google ($0.024315000, stop); P3 xAI ($0.008876400, stop). 52% of the three-seat worst case, no reserve fired. Same three seats as E-20260728f, whose five-option block was reused verbatim. | | 2026-07-31 | S070 | the sweep, the translation limb, the contamination gate and every analysis | 0.00 | 0.000000 | — | No API call. Thirteen published agreement figures recomputed from stored bodies; 蒲松齡〈促織〉 rendered entire from the Chinese (1,828 chars → 2,280 words) with an R06 draft frozen separately first; the contamination measurement via tools/dependence_check.py; and sweep.py, build_items.py, analyse.py, the 195-check verify.py and the six-mutation mutation_test.py are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). | | 2026-07-31 | S071 | E-20260731c-futabatei-alignment pre-run critic — qwen/qwen3.7-max (probed-but-not-selected) | 0.13 worst case from max_tokens 8,000 (note (abc)) | 0.050161800 | per-response usage.cost | in 5,253 / out 9,585, stop, 174 s, provider Alibaba. Not an alignment seat and not their reserve, so the S053 role-collision fix holds — fifteenth session running. Verdict NEEDS-AMENDMENT, six findings, two BLOCKING, all six accepted, none declined. Note (rr), twenty-fourth consecutive session. Its second BLOCKING finding withdrew the design's own crux (PL4) before the run and forced the search that established no native unforced Japanese comparator exists. | | 2026-07-31 | S071 | E-20260731c-futabatei-alignment the independent alignment check — 3 seats, 24 items, one call each | 0.18 worst case for three seats from max_tokens 4,000, P3 priced at the worst plausible provider (S022 routing caution) | 0.093463900 | per-response usage.cost | P1 OpenAI ($0.026682, stop); P2 Google AI Studio ($0.0275535, stop); P3 xAI ($0.0392284, stop). 52% of the three-seat worst case; no reserve fired and note (b) did not fire. A reserve (P5) was declared for the stage before dispatch, per note (bfc). | | 2026-07-31 | S071 | the translation limb, the aligner, the hand alignment, the contamination gate and every analysis | 0.00 | 0.000000 | — | No API call. Turgenev «Свидание» ¶42–¶69 rendered into Japanese twice (669 Russian words, unforced then R09-forced, each frozen in git before the next existed); the dialogue-anchored aligner, the hand alignment, both permutation and length-rate nulls, tools/dependence_check_cjk.py used unmodified, and align.py, hand_alignment.py, build_items.py, analyse.py, followups.py, contamination.py and the 68-check verify.py with six mutation tests are all lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

| 2026-07-31 | S072 | E-20260731d-sense-axes pre-run critic — qwen/qwen3.7-max (probed-but-not-selected) | 0.13 worst case from max_tokens 8,000 (note (abc)) | 0.052175175 | per-response usage.cost | in 7,965 / out 9,136, stop, 166 s, provider Alibaba. Not a rater and not the raters' reserve, so the S053 role-collision fix holds — sixteenth session running. Verdict ACCEPT, one finding, MINOR, accepted and applied (a P5 name collision between the reserve seat and prediction P5). Note (rr), twenty-fifth consecutive session, and its lightest return yet: asked to endorse or contest each of the 8 positive-control items sight-unseen, on the S070 precedent, it endorsed all eight, so nothing was excluded and the control set is the lead's set with a second opinion on it. | | 2026-07-31 | S072 | E-20260731d-sense-axes the CAT block — 3 seats, 40 items, one call each | 0.15 worst case for three seats from max_tokens 6,000 | 0.057972650 | per-response usage.cost, re-summed from the .raw bytes by verify.py | P1 OpenAI ($0.02027225, stop); P2 Google AI Studio ($0.02064, stop); P3 xAI ($0.0170604, stop). 39% of the three-seat worst case; no reserve fired and note (b) did not fire. A reserve (deepseek/deepseek-v4-pro) was declared for the stage before dispatch, per note (bfc). | | 2026-07-31 | S072 | E-20260731d-sense-axes the GRAD block — the same 3 seats, the same 40 items reshuffled | 0.15 worst case for three seats from max_tokens 6,000 | 0.045052650 | per-response usage.cost, re-summed from the .raw bytes | P1 OpenAI ($0.01718325, stop); P2 Google AI Studio ($0.016017, stop); P3 xAI ($0.0118524, stop). 30% of the three-seat worst case. | | 2026-07-31 | S072 | the translation limb, the contamination gate, both nulls and every analysis | 0.00 | 0.000000 | — | No API call. Leskov «Левша» chapters 6–8 rendered entire from the Russian (974 words → 1,376 English) with a 54-decision log, frozen in git before the design existed; the contamination gate via tools/dependence_check.py unmodified; and build_items.py, run_critic.py, run.py, analyse.py and the 78-check verify.py with seven mutation tests are all lead work or local computation. Nothing in tools/ changed this session. Lead translation is free and is never ledgered (charter §3, A4). |

Declared worst case $0.50; actual $0.155200475 — 31%, inside the 15–48% band this ledger has been running. Note (abc); the estimate was built from max_tokens and not from an assumed output length, and no amendment moved it in session.

THE KEY-USAGE CROSS-CHECK RECONCILES THIS SESSION, WHICH IS WORTH SAYING AFTER S071. Opening snapshot, taken immediately before the first call, 35.969639767; settled 36.124840242; delta 0.155200475. Per-request sum over the seven accepted calls 0.155200475. Equal to 1e-9. Note (bfv)'s prescription was followed — snapshots around each call rather than around the session — and no inter-call interval moved: every one of the seven in-session deltas read 0.000000000 at the moment of the call, the settling lag of note (bcx), and the whole amount appeared in the settled reading afterwards. No cross-session delta is computed against S071, whose own cross-check was void.

| 2026-07-31 | S073 | E-20260731e-option-census pre-run critic — qwen/qwen3.7-max (probed-but-not-selected), two dispatches | 0.13 worst case from max_tokens 8,000 (note (abc)) | 0.044181855 | per-response usage.cost; opening key snapshot runs/snap/session-open.json | First dispatch WASTED and BILLED NOTHING: finish_reason: error, content: null, 4,019 reasoning tokens, cost: 0.000000 in the usage block though upstream_inference_cost read 0.025619275 — note (b), nineteenth firing, and the first time this ledger has seen upstream_inference_cost present on a call that was not billed. Second dispatch at max_tokens 12,000: in 5,793 / out 9,180 of which 8,508 reasoning, stop, 167 s, provider Alibaba. The seat needed more than 8,000 tokens to think before emitting anything, which is exactly what note (b) says and exactly what the estimate failed to allow. Not a rater and not the raters' reserve — S053 role-collision fix, seventeenth session running. Verdict NEEDS-AMENDMENT, three findings, one BLOCKING, all three accepted. Note (rr), twenty-seventh consecutive session. | | 2026-07-31 | S073 | E-20260731e round 1, arm NONE — 3 seats | 0.28 worst case for three seats from max_tokens 6,000 (note (abc)), P3 priced at the worst plausible provider (S022 routing caution) | 0.057058150 | per-response usage.cost, re-summed from the .raw bytes by analysis/verify.py | P1 OpenAI ($0.01251375), P2 Google ($0.023202), P3 xAI ($0.0213424); all stop, all 32 items. 20% of the three-seat worst case; no reserve fired. A reserve (deepseek/deepseek-v4-pro) was declared for every stage before dispatch, per note (bfc). | | 2026-07-31 | S073 | E-20260731e round 2, arm SHAM — 3 seats, 4 dispatches | 0.28 worst case for three seats | 0.082559050, of which 0.012084400 wasted | per-response usage.cost, re-summed from the .raw bytes | P1 OpenAI ($0.01603625), P2 Google ($0.029256). P3 xAI FAILED with a new shape — finish_reason: stop, content: null, the answer's FIRST ITEM ONLY inside message.reasoning, 1,001 completion tokens, and no reasoning.effort set at all — which is note (bfb)'s shape without the parameter (bfb) blames, and is new note (bga). A plain retry of the identical payload to the identical slug returned all 32 items ($0.0251824), so the reserve was not needed and was not used. | | 2026-07-31 | S073 | E-20260731e round 3, arm FRAMEWORK — 3 seats | 0.28 worst case for three seats | 0.066622900 | per-response usage.cost | P1 OpenAI ($0.0199825), P2 Google ($0.025548), P3 xAI ($0.0210924); all stop, all 32 items. 24% of the three-seat worst case. | | 2026-07-31 | S073 | E-20260731e round 4, arm NONE repeated BYTE-IDENTICALLY — the same 3 seats | 0.28 worst case for three seats | 0.056132500 | per-response usage.cost | P1 OpenAI ($0.0100161), P2 Google AI Studio ($0.027996), P3 xAI ($0.0181204). This is the cross-day drift gate's option 1 and the first design to take it by choice, and it is the cheapest thing in this ledger that has ever changed a conclusion: the floor it measured (0.375 on P2, 0.000 on P1) is larger than the effect the design was built to detect, so the headline went from a two-of-three-seat rise to a null. Note the provider change on P2 between two byte-identical requests — Google → Google AI Studio, which note (beo) says is not the same served model. | | 2026-07-31 | S073 | both contamination gates, the translation limb, the census, the item build and every analysis | 0.00 | 0.000000 | — | No API call. Reymont «Śmierć» ¶0–51 rendered from the Polish (904 words → 1,132 English), the project's first Polish, with an R06 draft frozen in its own commit first and a 40-site option census; plus 308 abandoned words of Sienkiewicz, translated and then dropped when the gate returned 18 tokens. tools/dependence_check.py used unmodified for both gates; prompts.py, call.py, run_critic.py, run.py, analyse.py and the 116-check analysis/verify.py with four mutation tests, all four caught are lead work or local computation. Lead translation is free and is never ledgered (charter §3, A4). |

Declared worst case $1.20; actual $0.306554455 — 26%, inside the 15–48% band this ledger has been running. Note (abc); the estimate was built from max_tokens and no amendment moved it in session.

THE KEY-USAGE CROSS-CHECK RECONCILES EXACTLY, AND IT SETTLES A QUESTION ABOUT WASTED CALLS. Opening snapshot 36.131037442 (taken immediately before the first call, and $0.006197675 above S072's settled close — the settling lag of note (bcx), still present and still small); settled close 36.437591897; delta 0.306554455. Per-request sum over the thirteen accepted-or-billed calls 0.306554455. Equal to 1e-9. Both wasted dispatches therefore billed nothing — the errored critic call and the null-content grok-4.5 call — which is S070's outcome rather than S069's, so note (bfo) stays withdrawn and its own case stands alone. And upstream_inference_cost is not what is billed: the errored critic call carries $0.025619275 in that field and contributed $0.000000000 to the delta.

2026-07-31 day total: $1.012808407 of $5.00 — five sessions (S069 $0.309273057, S070 $0.098154720, S071 $0.143625700, S072 $0.155200475, S073 $0.306554455). $3.987191593 headroom. 20% of the day's cap.

(S072's day-total line, superseded by the S073 line above and kept for the audit trail:) 2026-07-31 day total: $0.706253952 of $5.00 — four sessions (S069 $0.309273057, S070 $0.098154720, S071 $0.143625700, S072 $0.155200475). $4.293746048 headroom. 14% of the day's cap.

(S071's day-total line, superseded by the S072 line above and kept for the audit trail:) 2026-07-31 day total: $0.551053477 of $5.00 — three sessions (S069 $0.309273057, S070 $0.098154720, S071 $0.143625700). $4.448946523 headroom.

The declared worst case was RAISED in session, from $0.25 to $0.30, with the reason written before dispatch — the check's max_tokens went 3,000 → 4,000 against note (b), which has fired sixteen times on empty length returns. Actual $0.1436257, 48% of the amended worst case. Note (abc).

THE KEY-USAGE CROSS-CHECK IS VOID THIS SESSION AND THAT IS ITSELF THE FINDING — note (bfv), new. The key read 34.354965717 at the critic call and 35.505091267 immediately before the alignment check, a rise of $1.149 across an interval in which this session made no call whatever; and the session's opening figure was $1.198 above S070's closing figure of 33.156987017. Two unattributable movements of over a dollar each. The per-request usage.cost sums are primary and are unaffected (CLAUDE.md: per-request costs are primary, the delta is a sanity check) — but the sanity check cannot be run, and reporting the residual as drift would have been wrong. Take snapshots around each CALL, not around the session, and declare the cross-check void when any inter-call interval moves. This is the concurrency hazard NEXT.md has carried since S048, appearing for the first time as a corrupted instrument rather than as a worry about id collisions.

2026-07-31 day total: $0.551053477 of $5.00 — three sessions (S069 $0.309273057, S070 $0.098154720, S071 $0.143625700). $4.448946523 headroom. (This total is the sum of per-request costs across the three sessions; it is NOT reconcilable against the key-usage figure this session, for the reason above.)

(S070's day-total line, superseded by the S071 line above and kept for the audit trail:) 2026-07-31 day total: $0.407427777 of $5.00 — two sessions (S069 $0.309273057, S070 $0.098154720). $4.592572223 headroom.

S070's key-usage delta is exact. Opening 33.058832297, settled 33.156987017, delta 0.098154720; per-request sum 0.098154720. Equal to 1e-9, which is what establishes that the abandoned call cost nothing. (The opening snapshot sat 0.006736 above S069's closing figure — the settling lag of note (bcx), still present.)

The ledgered figure is the KEY DELTA, not the per-request sum, and this is the first day that distinction has mattered. Opening snapshot 32.74282324, closing 33.052096297, delta 0.309273057; the per-request sum over the five accepted calls is 0.1154717578. The $0.1938 difference is the abandoned kimi-k3 call, which has no usage row to sum. Note (bet) says the session-level bound is the valid check and note (bco) says settling runs in the shortfall direction; this is the excess direction with a known cause, and the honest ledger entry is the delta. Five billed calls accepted, one dispatch abandoned and billed. | 2026-07-31 | S074 | E-20260731f-catalogue-reach pre-run critic — moonshotai/kimi-k3, one dispatch, WASTED | 0.20 worst case from max_tokens 12,000 (note (abc)) | 0.195747000 | per-response usage.cost; opening key snapshot runs/snap/critic-open.json | finish_reason: length, 11,997 reasoning tokens, ZERO characters of content. Note (b), TWENTIETH firing, and the first time a cap raised because of this note was still not enough — 12,000 was chosen precisely because S073 lost the same seat at 8,000. Unlike S073's wasted dispatches, this one BILLED. | | 2026-07-31 | S074 | E-20260731f pre-run critic, re-dispatched to the DECLARED RESERVE — deepseek/deepseek-v4-pro | included above | 0.030371504 | per-response usage.cost | Provider Novita, stop, 135 s, 9,346 reasoning tokens of 10,390. Verdict NEEDS-REDESIGN, six findings, two BLOCKING, all six accepted, amendments A1–A9. The reserve is not a rater in this design, so the S053 role-collision fix is not breached. Note (bfc), working as designed for the first time on a critic fall-through this session. | | 2026-07-31 | S074 | E-20260731f stage 2b, independent reachability typing — 2 seats, added by amendment A5 to answer the critic's BLOCKING Finding 1 | 0.10 worst case for two seats from max_tokens 4,000 | 0.049995200 | per-response usage.cost | google/gemini-3.6-flash (Google, $0.03036) and qwen/qwen3.7-max (Alibaba, $0.0196352); both stop, both 41 items. Neither is a rater nor the critic. The two calls that made the session's reach figure honest cost 6% of the session. | | 2026-07-31 | S074 | E-20260731f stage 3 — own-passage second read of the residue, 3 passages × 2 raters, 8 dispatches for 6 cells | 0.32 worst case from max_tokens 6,000 | 0.164076600, of which 0.010324400 on a null retry | per-response usage.cost, re-summed from the .raw bytes by analysis/verify.py | P1 OpenAI throughout, P3 xAI throughout. One cell failed twice in two different shapes: s3-C-P3 first returned a 200 whose body was 2,167 bytes of pure whitespace and no JSON at all (new note (bgc), and it crashed the runner and aborted the two stages behind it), then returned finish_reason: stop with 176 completion tokens all reasoning and zero content (note (bfb)), then on a third dispatch returned all nine items. | | 2026-07-31 | S074 | E-20260731f stage 4 — the crossover, 4 cells × 2 raters | 0.42 worst case | 0.346179450 | per-response usage.cost | All stop, all items parsed, 0 missing across 8 cells. x-ai/grok-4.5 cost 3.3× openai/gpt-5.6-terra on identical prompts ($0.4364 against $0.1346 over nine comparable cells); note (x)'s list-price assumption held on every call and the difference is output length, not routing. | | 2026-07-31 | S074 | E-20260731f stage 5 — one cell re-sent BYTE-IDENTICALLY, 2 raters | 0.11 worst case | 0.060393400 | per-response usage.cost | Prompt bytes identical and prompt-token counts identical to the digit, both asserted by the verifier. Flip rate 5 of 34 cells, 14.7%, roughly double RS-20260730d's 7.2% on the same two seats. | | 2026-07-31 | S074 | the translation limb, both contamination gates, stage 2 and every analysis | 0.00 | 0.000000 | — | No API call. Kielland «Karen» ¶1–15 rendered twice from the Norwegian (799 source words → 862 + 950 English), the project's first Norwegian, under the new regime R10 against two different Tier 1 anchors, with two frozen logs and a frozen opportunity list; tools/dependence_check.py used unmodified for both gates; build_claims.py, checks.py, prompts.py, run.py, run_typing.py, analyse.py and the 130-check analysis/verify.py with five mutation tests are lead work or local computation. Nothing in tools/ changed this session. Lead translation is free and is never ledgered (charter §3, A4). | | 2026-07-31 | S075 | E-20260731g-provenance pre-run critic — qwen/qwen3.7-max (probed-but-not-selected), one dispatch after one abandoned | 0.18 worst case from max_tokens 12,000 (note (abc)) | 0.040804105 | per-response usage.cost; re-summed from the .raw bytes by analysis/verify.py; opening key snapshot runs/snap/session-open.json 37.474390751 | in 4,829 / out 8,738 of which 31,547 characters reasoning, stop, 158.9 s, provider Alibaba. 23% of the declared worst case. A reserve (openai/gpt-5.6-terra, a different lab) was declared before dispatch per note (bfc) and was not used; note (b) did not fire — the cap was left at 12,000 rather than raised, which is NEXT.md's reading that the remedy is the fall-through. Verdict NEEDS-AMENDMENT, seven findings, one BLOCKING, all seven accepted, none declined; two answered with a stronger remedy than proposed. Note (rr), twenty-eighth consecutive session. Its BLOCKING finding replaced the lead's self-written contamination floor with a deterministic non-lead null that then killed the session's own leading explanation. This session dispatched no other API call, so the S053 role-collision fix is vacuous rather than satisfied and is recorded as such — nineteenth session running. | | 2026-07-31 | S075 | one dispatch ABANDONED before the above, 120-second client timeout on a seat that took 158.9 s when re-run | (not separately estimated; the timeout created it) | unknown — this session cannot price it | — | Note (bfo), NEITHER CONFIRMED NOR REFUTED. Nothing was written to disk. The closing key snapshot read 37.474390751, identical to the opening, so the delta is $0.000000 against a per-request sum of $0.040804105 — note (bco)'s settling lag has swallowed the accepted call as well as any abandoned one, and no attribution is possible at hand-off. S069's twenty-minute hang billed nearly its full cap; S070's 120-second timeout billed nothing. This one is a third case and it is recorded as unresolved rather than assigned to either. | | 2026-07-31 | S075 | the archive sweep, the mutation harness, both translations, the contamination measurement, the anthology null, both length controls and every analysis | 0.00 | 0.000000 | — | No API call. 220 stored response bodies and 39 frozen verifiers swept; 54 verifier runs across four mutation passes; d'Annunzio «La fine di Candia» ¶1–24 rendered twice from the Italian (800 words → 971 and 848 English tokens) with 61 logged decisions and a frozen opportunity list; tools/dependence_check.py used unmodified; and sweep.py, mutate.py, measure.py, call.py, run_critic.py and the 80-check analysis/verify.py with six mutation tests are all lead work or local computation. Nothing in tools/ changed. Lead translation is free and is never ledgered (charter §3, A4). |

Declared worst case $1.30, RAISED IN SESSION to $1.60 with the reason written before the next dispatch — the wasted kimi-k3 dispatch ($0.195747), the reserve critic ($0.030372), and amendment A5's two extra calls. Actual $0.846762554 = 53% of the amended figure, above the 15–48% band this ledger has been running, because eleven of the twenty-one calls were prompt-dominated: each rater cell carries two passages of ~1,500 words. Note (abc); the estimate was built from max_tokens and not from an assumed output length.

THE KEY-USAGE CROSS-CHECK RECONCILES EXACTLY. Opening snapshot 36.437591897 (taken immediately before the first call, and equal to S073's settled close to the digit — no settling lag this session, unlike S073 and S070); settled close 37.284354451; delta 0.846762554. Per-request sum over the twenty-one billed bodies 0.846762554. Equal to 1e-9. Note (bet)'s rule was followed: only the session-level bound is asserted, because the per-stage snapshots are unreadable here — run.py stage3 ran twice after the (bgc) crash and overwrote its own stage snapshots.

| 2026-07-31 | S076 | E-20260731h-carryover pre-run critic — qwen/qwen3.7-max, one dispatch, declared reserve google/gemini-3.6-flash not used | 0.25 worst case from max_tokens 12,000 (note (abc)) | 0.053150150 | per-response usage.cost; opening key snapshot runs/snap/session-open.json 37.523401556 | Provider Alibaba, stop, 195.6 s, 10,775 completion tokens. Verdict NEEDS-REDESIGN — the second this project has taken since S021 — seven findings, three BLOCKING, three MANDATORY, one ADVISORY, all seven accepted, three answered with a stronger remedy than proposed and one literal remedy (raise temperature) declined in writing with a control offered instead. Note (rr), twenty-ninth consecutive session. 21% of the declared worst case. | | 2026-07-31 | S076 | E-20260731h subjects, seat S1 — openai/gpt-5.6-terra, 9 cells | included in the $0.70 subject line | 0.071335950 | per-response usage.cost, re-summed from the .raw bytes by analysis/verify.py | Provider OpenAI on all nine, all stop, largest completion 1,717 tokens against a 2,200 cap. The cheapest of the three seats by a factor of two. | | 2026-07-31 | S076 | E-20260731h subjects, seat S2 — x-ai/grok-4.5, 9 cells | included above | 0.166094800 | per-response usage.cost | Provider xAI on all nine, all stop. One cell returned 4,231 completion tokens against the declared 2,200 cap — max_tokens was not enforced for this slug, so note (abc)'s worst case was not in fact an upper bound. New note (bgk). | | 2026-07-31 | S076 | E-20260731h subjects, seat S3, RUN 2 — deepseek/deepseek-v4-pro at max_tokens 8,000, 9 cells + 1 reserve | 0.35 worst case, declared in amendment A8 before the re-dispatch | 0.135711755 | per-response usage.cost | Providers Together then StreamLake throughout. Eight of nine returned clean at 5,072–7,427 completion tokens; the ninth burned all 8,000 and returned zero characters, falling through to google/gemini-3.6-flash ($0.0329415) — which is a different model, so that cell is VOID as a same-seat repeat. Note (b), twenty-second firing; new note (bgl) on the cap-raise having worked eight times and failed once. | | 2026-07-31 | S076 | E-20260731h subjects, seat S3, RUN 1 — WASTED, and the fault is the lead's | (not separately estimated; it was the declared run) | 0.150118077 | per-response usage.cost, preserved under runs/void-run1-qwen/ | deepseek-v4-pro returned finish_reason: length with 2,200 tokens and zero characters of content on three cells (note (b), twenty-first firing, and the first ever on a translation SUBJECT rather than a critic or rater), and the reserve that answered them was qwen/qwen3.7-max — this design's own pre-run critic. That is the S053 role collision, written into run.py's reserve table by the lead and invisible to the critic, who was shown design.md and not the runner. All three renderings are VOID and the run is preserved entire. 26% of the session's spend. New note (bgj). | | 2026-07-31 | S076 | the four lead translations, the opportunity list, both comparator extractions and every measurement | 0.00 | 0.000000 | — | No API call. Mikszáth «Szent Péter esernyője» I.iii, 535 Hungarian words rendered four times (1,366 English words, 110 logged decisions), the project's first Hungarian and its fifteenth source language, with the R10 §6 opportunity list frozen before any of them and the arm order crossed within one work for the first time; tools/dependence_check.py and tools/build_index.py used unmodified; prompts.py, run.py, analyse.py, mutate.py, a multi-turn call.py and the 403-check analysis/verify.py with six mutation tests, six caught and a passing control are lead work or local computation. Nothing in tools/ changed. Lead translation is free and is never ledgered (charter §3, A4). |

S076's declared worst case was $0.45, RAISED IN SESSION to $0.95 with the reason written before the re-dispatch (amendment A8: the voided run 1, and S3's cap raised from 2,200 to 8,000). Actual $0.576410732 = 61% of the amended figure, above the 15–53% band this ledger has been running — and the reason is not estimation error but the $0.150118077 of voided work, which is 26% of the total. Without it the session is $0.426292655, or 45%.

THE KEY-USAGE CROSS-CHECK IS SHORT, WITH THE USUAL CAUSE. Opening snapshot 37.523401556 (taken immediately before the critic dispatch), closing 38.05968395, delta 0.536282394; per-request sum over the twenty-nine billed bodies 0.576410732. Shortfall $0.040128338 — note (bco)'s settling lag in the direction it usually runs, and unlike S075 there is no abandoned dispatch to attribute it to: every call this session returned a body with a usage row. CLAUDE.md makes per-request costs primary and the delta the sanity check, so $0.576410732 is the ledgered figure and note (bet)'s session-level bound is the only one asserted.

2026-07-31 day total: $2.476785798 of $5.00 — eight sessions (S069 $0.309273057, S070 $0.098154720, S071 $0.143625700, S072 $0.155200475, S073 $0.306554455, S074 $0.846762554, S075 $0.040804105, S076 $0.576410732). $2.523214202 headroom. 50% of the day's cap, and the first day in this ledger's history to carry eight sessions.

(S075's day-total line, superseded by the S076 line above and kept for the audit trail:) 2026-07-31 day total: $1.900375066 of $5.00 — seven sessions (S069 $0.309273057, S070 $0.098154720, S071 $0.143625700, S072 $0.155200475, S073 $0.306554455, S074 $0.846762554, S075 $0.040804105). $3.099624934 headroom. 38% of the day's cap. S075 is the cheapest session on this day by a factor of two and its principal unit cost nothing at all: an archive sweep, 54 verifier runs, 1,600 words of translation and every measurement were local computation, and the single call was the pre-run critic.

(S074's day-total line, superseded by the S075 line above and kept for the audit trail:) 2026-07-31 day total: $1.859570961 of $5.00 — six sessions. $3.140429039 headroom. 37%.

| 2026-08-01 | S077 | E-20260801-strata pre-run critic — qwen/qwen3.7-max, one dispatch after one lost to note (bgc); declared reserve moonshotai/kimi-k3 not used | 0.36 worst case from max_tokens 12,000 (note (abc)) | 0.053022120 | per-response usage.cost; opening key snapshot runs/snap/session-open.json 38.114613848 | Provider Alibaba, stop, 10,981 in / 11,121 out of which 36,511 characters reasoning. Verdict NEEDS-AMENDMENT, three findings, one BLOCKING, all three accepted, two answered with a stronger remedy than proposed. Note (rr), thirtieth consecutive session. 15% of the declared worst case. Its BLOCKING finding caught a synthetic control that would have voided the whole run through the design's own F1 — the third mis-built control in this line of work and the third caught only from outside (note (bgo)). | | 2026-08-01 | S077 | the session's FIRST dispatch, LOST — note (bgc) | (not separately estimated; it was the critic dispatch) | unknown — this session cannot price it | — | HTTP 200 with a body that is not JSON, which raised JSONDecodeError and crashed the runner before any guard existed. The bytes were not preserved; the guard built in response (call.NonJSONBody) preserved the session's second occurrence. No usage object, so the per-request sum is a lower bound, exactly as note (bfe) says. | | 2026-08-01 | S077 | E-20260801 C1 raters, seat P1 — openai/gpt-5.6-terra, 3 batches (one re-dispatched after a second note-(bgc) body) | 0.18 worst case for the seat from max_tokens 8,000 | 0.131503122 | per-response usage.cost, re-summed from the .raw bytes by analysis/verify.py | Provider OpenAI on all three, all stop, 663 / 664 / 814 characters — the terse seat, and the only one that needed no cap change. $0.111233650 of this is the three accepted primaries; the remaining $0.020269472 is the declared reserve deepseek/deepseek-v4-pro, which answered batch 2 after the note-(bgc) body with finish_reason: length and ZERO characters — note (b), and the primary was then re-dispatched and answered normally. | | 2026-08-01 | S077 | E-20260801 C1 raters, seat P3 — x-ai/grok-4.5, 3 batches (batch 1 re-dispatched to the primary after the reserve had answered it) | 0.27 worst case from max_tokens 12,000 | 0.114550800 | per-response usage.cost | Provider xAI throughout, all stop. One batch returned 1,331 bytes of pure whitespace — note (bgc), the same slug and the same shape S074 met — and the declared reserve deepseek/deepseek-v4-pro answered it for $0.002637600, which is included in the figure at left ($0.111913200 of accepted primaries plus the reserve). That would make the batch not a seat under the design's F5, so the primary was re-dispatched and answered normally; the reserve's answer is preserved unused under runs/void-p3-b1-reserve/ and no figure is computed from it. | | 2026-08-01 | S077 | E-20260801 C1 raters, seat P2, RUN 2 — google/gemini-3.6-flash at max_tokens 20,000 with effort: low, 3 batches | 0.24 worst case, declared in amendment A5 before the re-dispatch | 0.053184000 | per-response usage.cost | Providers Google and Google AI Studio, all three stop, all three first call. A third of the cost of the failed run at a cap two and a half times larger — note (bdl)'s shape, change the parameter rather than the seat. | | 2026-08-01 | S077 | E-20260801 C1 raters, seat P2, RUN 1 — WASTED, all three batches | (not separately estimated; it was the declared run) | 0.186061500 | per-response usage.cost, preserved under runs/void-p2-cap8000/ | Note (b), TWENTY-THIRD firing, and the first time a whole SEAT was lost to it. finish_reason: length on all three batches, 7,996 completion tokens each, returning a truncated prose thinking-summary rather than answers. 32% of the session's spend for nothing. It also exposed a defect in this session's own runner: the note-(bfb) fallback accepted the reasoning field because it had enough lines, and a line count is not a format check — new note (bgp), fixed as amendment A4 before the re-dispatch. | | 2026-08-01 | S077 | E-20260801 C2 byte-identical repeat (P1, batch 1) and C3 prose-only leak null (P2, 44 items) | 0.10 worst case | 0.042267900 | per-response usage.cost | Both stop. The repeat holds at 0.8947 cell-level (registered P3 ≥ 0.85). The null then beat the panel, 0.5294 to 0.4118 on the blind pool, which is the session's governing result and is not what its failure criterion was written to catch — new note (bgn). | | 2026-08-01 | S077 | the translation limb, both contamination gates, the item build and every analysis | 0.00 | 0.000000 | — | No API call. Multatuli «Max Havelaar» ch. I ¶1–4, 716 Dutch words into 796 English, the project's first Dutch and sixteenth source language, under R07 v1.0 with 78 logged sites, live options enumerated exhaustively before any code was assigned and both coverage tests applied at the moment of each decision; tools/dependence_check.py and tools/build_index.py used unmodified; parse_logs.py, build_items.py, prompts.py, run.py, analyse.py and the 130-check analysis/verify.py with six mutation tests, six caught, two of them cases the verifier must fail on a corrupted stored body, are lead work or local computation. Nothing in tools/ changed this session. Lead translation is free and is never ledgered (charter §3, A4). |

S077's declared worst case was $1.10 and it was NOT raised. Actual $0.580589442 = 53%, inside the 15–61% band this ledger has been running — and the notable thing is that it held despite $0.208968572 (36%) of work that produced no accepted cell. The estimate was built from max_tokens per note (abc); amendment A5 raised P2's cap from 8,000 to 20,000 and the seat then cost less than it had at the lower cap, which is note (b)'s remedy paying for itself for the first time in this ledger.

THE KEY-USAGE CROSS-CHECK IS SHORT, WITH THE USUAL CAUSE. Opening snapshot 38.114613848 (taken immediately before the critic dispatch; usage_daily 0.0110333, so $0.011 of the UTC day was already spent by something outside this session), closing 38.634664990, delta 0.520051142; per-request sum over the seventeen billed bodies 0.580589442. Shortfall $0.060538300 — note (bco)'s settling lag in the direction it usually runs. Two dispatches returned no usage object at all (note (bgc)'s non-JSON bodies), so the per-request sum is a lower bound in the sense note (bfe) established; it is nonetheless the ledgered figure, per CLAUDE.md, and note (bet)'s session-level bound is the only one asserted. The opening snapshot was overwritten once when the crashed first stage was re-run — both reads returned the same figure to the digit, so nothing was lost, and the runner defect is recorded as note (bgq).

2026-08-01 day total: $0.580589442 of $5.00 — one session (S077). $4.419410558 headroom. 12% of the day's cap. (The key's own usage_daily read $0.0110333 before this session's first dispatch and $0.531084442 at close; the $0.011 opening figure is not attributable to this project by anything this session can see, and is recorded rather than claimed.)

| 2026-08-01 | S078 | E-20260801b-census-author pre-run critic — qwen/qwen3.7-max; declared reserve moonshotai/kimi-k3 not used | 0.20 worst case from max_tokens 12,000 (note (abc)) | 0.071769075 | per-response usage.cost; opening key snapshot runs/snap/critic-open.json 38.69520329 | Provider Alibaba, stop, 222.3 s, in 11,964 / out 12,231 of which 11,055 reasoning. Probed-but-not-selected: not a rater and not the raters' declared reserve, so the S053 role-collision fix holds — eighteenth session running. Verdict NEEDS-REDESIGN, five findings, two BLOCKING, all five accepted, one prescribed remedy DECLINED with a written reason. Note (rr), thirty-first consecutive session. 36% of the declared worst case. Note (bgo) applied literally — the prompt opened with the control block — and it returned twice over: item-specific control verdicts instead of a blanket endorsement, and an ADVISORY finding that predicted the control failure which then happened at 3 of 3 seats. Its BLOCKING finding 2 rebuilt the payloads as a four-way rotation before any rater call and raised the declared worst case from $1.14 to $1.68, recorded rather than absorbed. | | 2026-08-01 | S078 | E-20260801b payloads R1–R4, the author-arm rotation — 3 seats each, 12 accepted + 2 wasted | 1.10 worst case for twelve seats from max_tokens 6,000, P3 priced at the worst plausible provider (S022 routing caution) | 0.852516100, of which 0.101253000 wasted | per-response usage.cost, re-summed from the .raw bytes by analysis/verify.py | P1 OpenAI throughout, P3 xAI throughout, all stop, all 30 of 30 lines. R4/P2 cost three dispatches: finish_reason: length twice at 5,996 completion tokens (note (b), twenty-fourth and twenty-fifth firings) before a third returned 30 lines. 78% of the twelve-seat worst case, and the overrun driver is P3, which alone accounts for $0.399 of the $0.853. | | 2026-08-01 | S078 | E-20260801b payload B, the elicitation-mode arms — 3 seats, 3 accepted + 2 wasted | 0.28 worst case for three seats from max_tokens 6,000; P2 re-dispatched at 20,000 under amendment A6, declared before the call | 0.364954400, of which 0.106525500 wasted | per-response usage.cost | P1 OpenAI and P3 xAI first call, both stop, both 46 of 46. B/P2 is the session's instrument finding: length at 5,996 tokens, then a retry that reported stop while ending mid-item-id at the same 5,996 tokens — new note (bgt). Caught by the registered F4 line count, not by finish_reason; both bodies preserved unread under runs/void-b-p2-cap6000/ and the verifier asserts the analysed body is not one of them. The re-dispatch at the raised cap cost $0.138051 against the failed call's $0.053262 — 2.6× MORE, which is the opposite of S077's result and is why note (b)'s remedy paying for itself is not a rule. | | 2026-08-01 | S078 | E-20260801b the byte-identical repeat — R1 re-issued to P1 unchanged | 0.10 worst case | 0.034111000 | per-response usage.cost | Provider OpenAI, stop, 30 of 30. The cheapest call of the session and the one that decided how its numbers may be read: it moved k by 0.333–0.667 in every arm, all in one direction, on a request that did not change. Note (bfz), second affirmative application. | | 2026-08-01 | S078 | the translation limb, all three contamination gates, the item build and every analysis | 0.00 | 0.000000 | — | No API call. Szymański «Stolarz Kowalski» opening, 759 Polish words into 1,045 English, two adjacent passages under two elicitation modes with the order fixed in an instrument frozen before the span was chosen; R06 drafts frozen in their own commits first. tools/dependence_check.py, tools/build_index.py and tools/check_balance.py used unmodified; mode-rule.md, build_items.py, call.py, run_critic.py, run.py, analyse.py and the 458-check analysis/verify.py with six mutation tests, six caught are lead work or local computation. Nothing in tools/ changed this session. Lead translation is free and is never ledgered (charter §3, A4). |

S078's declared worst case was $1.14, RAISED IN SESSION to $1.68 with the reason written before any rater call (amendment A2: the pre-run critic's BLOCKING finding 2 doubled the author-arm dispatches from six to twelve). Actual $1.323350575 = 79% of the amended figure and 116% of the original — the highest fraction in this ledger's history, and the amended estimate is the only reason it is not an overrun. $0.207778500 (15.7%) produced no accepted cell, all of it one seat truncating.

THE KEY-USAGE CROSS-CHECK IS EXACT. Opening snapshot 38.69520329 (taken immediately before the critic dispatch), closing 40.018553865, delta 1.323350575; per-request sum over the twenty-one billed bodies 1.323350575. Residual −0.000000000 — the first exact-to-1e-9 session-level reconciliation since S073, and every dispatch returned a usage object, so the sum is not a lower bound in note (bfe)'s sense. The key's own usage_daily reads 1.914973317 against a two-session per-request total of 1.903940017; the $0.011033300 difference is exactly the figure S077 recorded as already spent by something outside this project when it opened, and is recorded rather than claimed.

2026-08-01 day total: $1.903940017 of $5.00 — two sessions (S077 $0.580589442, S078 $1.323350575). $3.096059983 headroom. 38% of the day's cap.

| 2026-08-01 | S079 | the CTRL-NEG gate — 3 adjudicating seats, openai/gpt-5.6-terra / google/gemini-3.6-flash / x-ai/grok-4.5, reserve deepseek/deepseek-v4-pro | 0.102 worst case from max_tokens 3,000 (note (abc)); registered in gate/registration.md §7 before dispatch | 0.105042406, of which 0.087967971 wasted | per-response usage.cost; opening key snapshot gate/runs/snap/session-open.json 40.348182665 | P1 stop first call ($0.006756750); P3 stop first call ($0.006690400). Seat 2 cost six dispatches: gemini-3.6-flash length three times at cap 3,000 ($0.0257 each) and deepseek-v4-pro length twice ($0.0055, $0.0054) before a third deepseek dispatch returned ($0.003627285) — note (b), and the reserve answering on its third try to an identical payload at temperature: 0. A 3% OVERRUN, and the cause is not the token cap: the estimate was one dispatch per seat and the caller can dispatch 3 attempts × 2 slugs. Note (abc) amended; attempts capped at 2 for the session's main design. Verdict 2 of 3 for the stated criterion, and all three seats read the design text the same way before seeing the code. | | 2026-08-01 | S079 | E-20260801c-anchor-instance pre-run critic — qwen/qwen3.7-max (probed-but-not-selected) | 0.16 worst case from max_tokens 12,000 | 0.061464725 | per-response usage.cost | Provider Alibaba, stop, 220.6 s, in 5,242 / out 12,143 of which 11,152 reasoning. Not a rater and not the raters' reserve — S053 role-collision fix, nineteenth session running. Verdict NEEDS-AMENDMENT, four findings, two BLOCKING, all four accepted, one with a different remedy and the reason written. Note (rr), twenty-ninth consecutive session. 38% of its declared ceiling. Its BLOCKING finding 1 moved the primary denominator from 19 propositions to 17 and made the registered bar stricter. | | 2026-08-01 | S079 | E-20260801c stage A, the per-text attestation — 2 arms × 3 seats, 6 accepted + 4 wasted | 0.76 worst case for six seats from max_tokens 8,000, attempts capped at 2 | 0.313193796, of which 0.060159746 wasted | per-response usage.cost, re-summed from the .raw bytes by analysis/verify.py | All six accepted bodies stop, all 22 of 22 lines plus the END terminator. The waste is the session's own runner defect, not a seat's: the F4 item-count guard was carried forward from S078 with that run's item-id pattern hard-coded, so it counted zero answer lines in every body and rejected four dispatches — AP_P1 $0.028725500, AP_P1_gpt-5.6-terra_try2 $0.011105200, AP_P1_deepseek-v4-pro_try1 $0.014202419 (length, empty body — note (b)), AP_P1_deepseek-v4-pro_try2 $0.006126627. New note (bgw). And the bodies really were truncated at stop / native completed with 303 / 219 / 244 content tokens against a cap of 8,000 — note (bgt) at two labs in one stage. The re-dispatch was a single-seat probe, declared in amendments.md A5 before it ran. | | 2026-08-01 | S079 | E-20260801c stage F, the forcing probe — 3 seats, Swedish alone | 0.18 worst case for three seats from max_tokens 4,000, attempts capped at 2 | 0.077239900, of which 0.031260000 wasted | per-response usage.cost | P1 stop first call, 8.2 s, $0.006031; P3 stop first call, $0.0141204. P2 length at cap 4,000 (note (b)) then stop on the same slug at the same cap. All three returned 3 of 3 paragraphs, so F4 did not fire. 43% of the declared worst case, and it produced the session's largest finding: three independent seats given only the Swedish return longest runs of 17 / 24 / 22 against Stork 1923, against the lead's 24 on the same material. | | 2026-08-01 | S079 | an in-flight dispatch killed by the operator — google/gemini-3.6-flash, stage A | (not estimated; it did not exist as a planned call) | ≈0.066959998, INFERRED, no per-request record | key-usage delta minus the per-request sum | The process was killed while diagnosing the F4 defect, with a gemini dispatch open; the runner writes the .raw only after the read returns, so the call billed and wrote nothing. The successful AP_P2 at the same cap billed $0.06726 — the two agree to $0.0003 and no other candidate exists. Third case for notes (bfo)/(bfp): S069 an abandoned hang billed at full cap, S070 an abandoned call billed nothing, S079 an operator-killed call billed and left no record. | | 2026-08-01 | S079 | the translation limb, the contamination gate, the hand read, stage 0 and every analysis | 0.00 | 0.000000 | — | No API call. Söderberg «Pelsen» entire, 1,273 Swedish words into 1,298 English — the project's thirteenth source language and its first Swedish — under R10 against A-mchugh-presence, with a 32-decision log carrying R10 target codes and an opportunity list frozen before it. tools/dependence_check.py, tools/build_index.py and tools/check_balance.py used unmodified; E-20260731f/analysis/checks.py imported unmodified and run against the anchor's second stored text, which no instrument had read before. fetch.py, build_gate.py, build_items.py, run.py, analyse.py and the 287-check analysis/verify.py with six mutation tests, six caught and each asserted to change bytes (note (bgu)) are lead work or local computation. Nothing in tools/ changed this session. Lead translation is free and is never ledgered (charter §3, A4). |

S079's declared worst case was $1.262 — $0.102 for the gate, registered separately, plus $1.16 for the design. Actual $0.623900826 = 49%, inside the band this ledger runs. The gate itself overran its own $0.102 by 3% and the reason was structural rather than about tokens: an estimate of one dispatch per seat is not a worst case when the caller can dispatch a seat 3 attempts × 2 slugs. Attempts were capped at 2 for the main design in the same session, and the retry structure was written into its estimate.

$0.179387717 (28.8%) produced no accepted cell, and almost all of it is this session's own defects rather than the seats': $0.0880 in the gate's retry loop, $0.0602 in four dispatches rejected by a guard pointed at the wrong item ids, $0.0313 in one genuine length truncation, plus the killed in-flight call.

THE KEY-USAGE CROSS-CHECK IS SHORT BY EXACTLY ONE KILLED DISPATCH. Opening snapshot 40.348182665 (taken immediately before the gate's first dispatch), closing 40.972083491, delta 0.623900826; per-request sum over the fifteen recorded bodies 0.556940828. Gap $0.066959998, attributed above to the operator-killed gemini dispatch on the arithmetic that it matches the same seat's successful call at the same cap to $0.0003. The larger key-delta figure is what is ledgered, per this page's stated method and the S015 precedent.

2026-08-01 day total: $2.527840843 of $5.00 — three sessions (S077 $0.580589442, S078 $1.323350575, S079 $0.623900826). $2.472159157 headroom. 51% of the day's cap.

| 2026-08-01 | S080 | E-20260801d-name-ground-truth pre-run critic — qwen/qwen3.7-max; declared reserve moonshotai/kimi-k3 not used | 0.16 worst case from max_tokens 12,000 (note (abc)) | 0.050377150 | per-response usage.cost; opening key snapshot runs/snap/session-open.json 40.979552091 | Provider Alibaba, stop, 199.7 s. Probed-but-not-selected: not a rater and not the raters' declared reserve — the S053 role-collision fix, twentieth session running. Verdict NEEDS-AMENDMENT, five findings, one BLOCKING, all five accepted, one prescribed remedy DECLINED with a written reason. Note (rr), thirtieth consecutive session. 31% of its declared ceiling. Its BLOCKING finding was note (bgx) found in this session's own design text: the byte-identical repeat's threshold was an absolute 0.90 standing in for a relative claim. Amended before dispatch — and the relative leg then fired, which the absolute one would not have. | | 2026-08-01 | S080 | E-20260801d the three rater seats, 138 items each, one call per seat | 0.48 worst case for three seats × 2 attempts from max_tokens 6,000, payload 11,288 input tokens measured, P3 priced at twice list for routing (S022 caution) | 0.264432680, of which 0.145144680 wasted | per-response usage.cost, re-summed from the .raw bytes | P1 OpenAI stop first call, 138 of 138, $0.02758925. P3 xAI stop first call, 138 of 138, $0.0493544. P2 cost four dispatches and an amendment: gemini-3.6-flash returned stop with 22 of 138 lines (5,996 completion tokens, 5,757 reasoning) and then stop with 107 of 138 and no END — note (bgt) twice at the cap S078 met it at — and the reserve deepseek-v4-pro came back 137 of 138 with the terminator present, which was NOT accepted because F3 is registered as exactly one line per item and relaxing a criterion after seeing the body it would reject is note (bgx)'s disease. Its second attempt returned a non-JSON body (note (bgc)) that billed nothing. Amendment A6, declared before the call: primary only, no reserve, cap 20,000. It returned 138 of 138 for $0.106467 — less than the two failed attempts it replaced ($0.118), which is S077's result and not S078's. | | 2026-08-01 | S080 | E-20260801d the byte-identical repeat — P1's payload re-issued unchanged | 0.10 worst case | 0.018173500 | per-response usage.cost | Provider OpenAI, stop, 138 of 138, 27.3 s. The cheapest call of the session and the one that decided how its agreement figures may be read: 0.9638 against a mean pairwise 0.9734, so the relative leg of amendment A1 fires and κ = 0.9435 is descriptive only. P1 agrees with P3 (0.9928) more than it agrees with itself. Note (bfz), third affirmative application. | | 2026-08-01 | S080 | the translation limb, the contamination gate, the pre-repair sweep, the partial recomputation and every analysis | 0.00 | 0.000000 | — | No API call. Brennu-Njáls saga chapters 1–2 entire, 1,151 Icelandic words into 1,348 English — the project's fourteenth source language and its first medieval Scandinavian source — under R04 v1.0 with the draft committed separately as an R06 output before revision and a 41-row log frozen before the comparator was opened and tallied by a committed parser. tools/dependence_check.py, tools/ngram_overlap.py, tools/build_index.py and tools/check_balance.py used unmodified; the Gutenberg comparator and the Gutenberg #52319 refetch for the sweep are free. fetch.py, fetch_comparator.py, build_gate.py, build_items.py, prompts.py, run.py, analyse.py, sweep/*.py, parse_log.py and the 93-check analysis/verify.py with six mutation tests, six caught and each asserted to change bytes are lead work or local computation. Nothing in tools/ changed this session. Lead translation is free and is never ledgered (charter §3, A4). |

S080's declared worst case was $0.76, RAISED IN SESSION to $1.10 by amendment A6 with the reason written before the call. Actual $0.397105980 = 36% of the amended figure and 52% of the original — inside the band this ledger runs, and the amendment is the reason it is not reported against a figure the run had already outgrown.

THE KEY-USAGE CROSS-CHECK IS SHORT BY EXACTLY ONE OVERWRITTEN BODY, AND THAT IS HOW THE DEFECT WAS FOUND. Opening snapshot 40.979552091 (taken immediately before the critic dispatch), closing 41.376658071, delta 0.397105980; per-request sum over the seven parseable stored bodies 0.332877480. Residual $0.064228500 — and the first P2 dispatch's usage.cost was $0.0642285, exact to the cent. That body's .raw was overwritten by the A6 re-dispatch, because call.py labels a primary slug's first attempt with the bare tag: note (bdt)'s preserve-before-parse guard defeated by a label collision. New note (bgz); backlog row opened. The larger key-delta figure is what is ledgered, per this page's stated method. Nothing in the analysis depended on the lost body — it was a rejected seat — and the .meta.json survived, which is the only reason the reconciliation is exact rather than merely close.

$0.145144680 (36.6%) produced no accepted cell, all of it the P2 seat: two gemini bodies failing the registered item guard and one reserve body one line short of it.

2026-08-01 day total: $2.924946823 of $5.00 — four sessions (S077 $0.580589442, S078 $1.323350575, S079 $0.623900826, S080 $0.397105980). $2.075053177 headroom. 58% of the day's cap. (The key's own usage_daily reads 3.273077523 at close against a four-session ledgered total of $2.924946823; the $0.348 difference is not attributable to this project by anything this session can see — S077 recorded $0.011 of the same kind at its open — and is recorded rather than claimed.)

| 2026-08-01 | S081 | E-20260801e-lead-carryover pre-run critic — qwen/qwen3.7-max; declared reserve google/gemini-3.6-flash not used | 0.16 worst case from max_tokens 12,000 (note (abc)) | 0.049782725 | per-response usage.cost; opening key snapshot runs/snap/session-open.json 41.386920671 | Provider Alibaba, stop, 172.7 s. Probed-but-not-selected: not a subject and not the subjects' reserve — the S053 role-collision fix, twenty-first session running. Verdict NEEDS-AMENDMENT, five findings, one BLOCKING, all five accepted in substance, two prescribed remedies DECLINED with written reasons. Note (rr), thirty-first consecutive session. 31% of its declared ceiling. Its BLOCKING finding was the normalisation denominator: min(tokens(a), tokens(b)) moves between the two comparisons the whole design rests on, so Carry could have moved with no change in raw overlap. Amended to the fixed text's own 7-gram count before a word was translated, and the remedy went further than prescribed by giving guard L a declared direction. | | 2026-08-01 | S081 | E-20260801e the Floor rung — 3 passages × 3 seats, fresh contexts, 8 accepted + 3 wasted | 0.41 worst case for nine seats from max_tokens 3,000 (K, C) and 1,600 (O), attempts capped at 2, P3 at twice list and P5 at four times list (S022 routing caution) | 0.138043428, of which 0.013038552 wasted | per-response usage.cost, re-summed from the .raw bytes by analysis/verify.py | P1 OpenAI and P3 xAI stop on every first call, 6 of 6. P5 deepseek-v4-pro cost all three failures: length twice on pair K (providers StreamLake, Baidu) exhausting that seat's two attempts, and once on pair O before Ionstream returned. Note (b), three firings in one run. No amendment was made and no cap was raised: design §9.3 already said a passage with two live seats reports its Floor on two seats, labelled — so the failure was absorbed by a rule written before the data instead of by a repair after it. 34% of the declared worst case. | | 2026-08-01 | S081 | the three translations, all extraction, the ladder, both controls, the leak sensitivity, the contamination measurement and every analysis | 0.00 | 0.000000 | — | No API call. Kielland «Karen» ¶1–15 (799 Norwegian words), d'Annunzio «La fine di Candia» ¶1–24 (800 Italian words) and Tarchetti «Un osso di morto» opening (357 Italian words) — 1,956 source words rendered blind against archived pairs, each frozen in its own commit before the next was begun, two of them under R10 with their own opportunity lists written before translating and without consulting the earlier session's. tools/dependence_check.py, tools/build_index.py and tools/check_balance.py used unmodified; material/extract.py, prompts.py, run.py, run_critic.py, analyse.py and the 503-check analysis/verify.py with six mutation tests, six caught and each asserted to change bytes on disk are lead work or local computation. Nothing in tools/ changed this session, and one live defect in tools/ngram_overlap.extract() was found, reported and deliberately not repaired. Lead translation is free and is never ledgered (charter §3, A4). |

2026-08-01 day total: $3.112772976 of $5.00 — five sessions (S077 $0.580589442, S078 $1.323350575, S079 $0.623900826, S080 $0.397105980, S081 $0.187826153). $1.887227024 headroom. 62% of the day's cap. (Key-usage cross-check: opening snapshot 41.386920671, closing 41.544405927, delta $0.157485256 against a per-request sum of $0.187826153. The delta is SMALLER than the sum — the ordinary settling lag note (bcx) names, and the fourth session in which it has run in that direction. Per-request sums are ledgered, per this page's stated method.)

| 2026-08-01 | S083 | E-20260801f-tierD-run pre-run critic — qwen/qwen3.7-max; declared reserve moonshotai/kimi-k3 not used | 0.16 worst case from max_tokens 12,000 (note (abc)) | 0.068078625 | per-response usage.cost; opening key snapshot runs/snap/session-open.json 41.583192123 | Provider Alibaba, stop, 252.1 s, 11,748 in / 11,469 out of which 10,202 reasoning. Probed-but-not-selected: not a juror (P1/P2/P5) and not a juror's reserve — the S053 role-collision fix, twenty-second session running. Verdict NEEDS-AMENDMENT, seven findings, two BLOCKING, all seven accepted, two prescribed remedies declined with written reasons. Note (rr), thirty-second consecutive session. 43% of its declared ceiling. Its blocking pair killed two sentences the design would otherwise have published: that the repaired sham lets the run catch a length response, and a pre-registered reading calling Garnett and Hapgood "same-quality". | | 2026-08-01 | S083 | the control-arm-spec ratification vote — P3 x-ai/grok-4.5, the standing gate on that page, routed before dispatch | 0.05 worst case from max_tokens 6,000 | 0.024095600 | per-response usage.cost | Provider xAI, stop, 73.1 s. A panel member (so a non-Anthropic panel vote, charter §8) and deliberately not one of this design's jurors. Verdict RATIFY-WITH-AMENDMENT, A1 applied verbatim; the S051 backlog obligation is discharged. And it returned the citing design's own reading of R1–R4 INCORRECT, which added §6.8 before anything was dispatched. 48% of its ceiling, and the cheapest correction of the session. | | 2026-08-01 | S083 | E-20260801f stage 1, the prior positive control — 1 item × 2 orderings × 3 jurors | 0.2896 full-stage reservation (§10), reserved against $1.7950 headroom before the first call | 0.046363, none wasted | per-response usage.cost, re-summed by analysis/verify.py | Caps P1=3500, P2=6000, P5=10000, sized from S034's measured per-call maxima (886 / 2,839 / 5,893). One transport failure (Connection reset by peer) retried and billed nothing; 6 of 6 accepted, 0 failures. 16% of the reservation. The gate fired (2 of 3 units at +1, none −1), so stage 2 was entered. | | 2026-08-01 | S083 | E-20260801f stage 2, the rebuilt sham — 5 items × 2 orderings × 3 jurors | 1.3470 full-stage reservation, reserved against $1.7487 headroom | 0.252122, none wasted | per-response usage.cost | 30 of 30 accepted, 0 failures, 493 s. Providers routed across eight (OpenAI ×12, Google ×8, Google AI Studio ×4, DeepInfra ×5, GMICloud ×2, Ionstream ×2, Novita ×2, Cloudflare ×1) — note (x). 19% of the reservation, and the per-call mean $0.0084 against S034's $0.0121 on the same task, most of it P1's halved price (S061). Result: IN BAND. | | 2026-08-01 | S083 | the translation limb, the contamination gate, the extract() audit, the materials build, the Metric A pre-check and all analysis | 0.00 | 0.000000 | — | No API call. Korolenko «Лес шумит» opening, 276 Russian words into 396 English under R04 v1.0, draft frozen separately as an R06 output in its own commit first, with a decision log frozen before any evaluation was designed. tools/dependence_check.py, tools/ngram_overlap.py, tools/metric_a.py, tools/run_tierD.py, tools/build_index.py and tools/check_balance.py used unmodified; materials/build.py, analysis/rules.py, analysis/score.py, analysis/metric_a_precheck.py and the 723-check analysis/verify.py with six mutation tests, six caught and each asserted to change bytes on disk (note (bgu)) are lead work or local computation. The experiment's local call.py carries a one-line fix for note (bgz)'s label collision; nothing in tools/ changed this session. Lead translation is free and is never ledgered (charter §3, A4). |

S083's declared worst case for what it actually dispatched was $1.847 ($0.16 critic + $0.05 vote + $0.2896 + $1.3470). Actual $0.390659225 = 21%. The two later stages of the design were not entered: stages 3–5 reserve $2.795 and the day had $1.497 left, so the run stopped at the stage boundary its dispatch order exists to make survivable, with the deferral written into NEXT.md. Nothing was wasted: 36 of 36 dispatched calls were accepted, and no body failed a guard.

THE KEY-USAGE CROSS-CHECK IS EXACT TO 1.55 × 10⁻⁷. Opening snapshot 41.583192123 (taken immediately before the critic dispatch), closing 41.973851503, delta 0.390659380; per-request sum over the 36 juror bodies plus the critic and the vote 0.390659225. The residual is float noise, not a lost body — the first session since S077 whose two accounting paths agree to better than a cent, and it is worth naming why: every attempt in this session was uniquely labelled, which is note (bgz)'s repair applied at the point of use.

2026-08-01 day total: $3.503432201 of $5.00 — seven sessions (S077 $0.580589442, S078 $1.323350575, S079 $0.623900826, S080 $0.397105980, S081 $0.187826153, S082 $0.00, S083 $0.390659225). $1.496567799 headroom. 70% of the day's cap.

| 2026-08-01 | S084 | E-20260801g-purpose-row pre-run critic — qwen/qwen3.7-max; declared reserve moonshotai/kimi-k3 not used | 0.16 worst case from max_tokens 12,000 (note (abc)) | 0.035414750 | per-response usage.cost; opening key snapshot runs/snap/session-open.json 41.986007903 | Provider Alibaba, stop, 147.8 s, 5,638 in / 6,124 out of which 4,554 reasoning. Probed-but-not-selected, and the S053 role-collision fix is satisfied trivially rather than by design: this experiment dispatches exactly one call and has no raters at all, which is stated rather than claimed as a control — twenty-third session the fix has held. Verdict NEEDS-REDESIGN (not NEEDS-AMENDMENT), seven findings across six headings, four BLOCKING, all accepted, one prescribed remedy declined with a written reason, two accepted in a stronger form than prescribed. Note (rr), thirty-third consecutive session. 22% of its declared ceiling — the cheapest critic of the eight. What it bought, before a word was translated: the primary's confirming licence cut to nothing (an elastic typology makes residue = 0 nearly free), a co-primary added, all three coding controls replaced because they tested meta-decisions where the primary asks about rendering decisions, a frozen purpose clause rewritten (it was naturalness restated as a licence and would have manufactured a trivial trade at every site), the controls' "blind" withdrawn rather than weakened, and one interpretive claim struck. | | 2026-08-01 | S084 | the three translations, both gates, all coding and every analysis | 0.00 | 0.000000 | — | No API call. Lagerlöf «En julgäst» ¶1–3 (216 Swedish words), «De fågelfrie» ¶42 (218) and the matched R11 pair on «Gudsfreden» ¶5–8 (276 words rendered twice, 393 + 301 English tokens) — 926 Swedish words, the project's second woman author, each frozen in its own commit before the next was begun and both pair arms frozen before any coding. tools/dependence_check.py, tools/build_index.py and tools/check_balance.py used unmodified; material/fetch.py, material/gate.py, material/spec_check.py, analysis/controls.py, parse_logs.py, divergence.py, pair_sites.py, score_controls.py, score_pair.py and the 86-check analysis/verify.py with three mutation tests, three caught and each asserted to change bytes on disk (note (bgu)) are lead work or local computation. call.py was copied unchanged from E-20260801f. Nothing in tools/ changed this session. Lead translation is free and is never ledgered (charter §3, A4). |

S084's declared worst case was $0.16, for the one call it made. Actual $0.035414750 = 22%.

THE KEY-USAGE CROSS-CHECK IS EXACT. Opening snapshot 41.986007903 (taken immediately before the critic dispatch), closing 42.021422653, delta 0.035414750; the single recorded body's usage.cost is 0.035414750. Zero residual — the second consecutive session whose two accounting paths agree, and for the same reason as S083's: every attempt uniquely labelled, note (bgz)'s repair applied at the point of use. Nothing was wasted: one dispatch, one acceptance.

Two frozen discard rules fired and cost nothing but work. The design's V3 contamination gate fired on two successive candidate works (17 and 16 contiguous tokens against a 1899 comparator, against six null cells at 3–4), and both were discarded before the studied pair was built — the order note (bcd) prescribes, working as a selection gate rather than as a flag.

2026-08-01 day total: $3.538846951 of $5.00 — eight sessions (S077 $0.580589442, S078 $1.323350575, S079 $0.623900826, S080 $0.397105980, S081 $0.187826153, S082 $0.00, S083 $0.390659225, S084 $0.035414750). $1.461153049 headroom. 71% of the day's cap. ARM-tierD-run stages 3–5 reserve $2.795 and still do not fit; they were deferred a second time, on arithmetic rather than on judgment.

| 2026-08-01 | S085 | D-20260801-10 ratification — independent adversarial review (openai/gpt-5.6-terra, P1) + ratifying vote (google/gemini-3.6-flash, P2) | 0.148 (worst case from max_tokens) | 0.0354295 | per-response usage.cost | review 0.0147865 (in 7,625 / out 876, provider OpenAI) + vote 0.020643 (in 8,832 / out 986, provider Google). Both finish_reason: stop; max_tokens 8,000 with reasoning: {effort: low}, i.e. notes (b) and (abc) applied together rather than one of them. Verdict B-AMENDED, unanimous, against the opening session's default of no change. Landed at 24% of the worst case. | | 2026-08-01 | S085 | D-20260801-11 ratification — independent adversarial review (deepseek/deepseek-v4-pro, P5) + ratifying vote (x-ai/grok-4.5, P3) | 0.089 (worst case from max_tokens) | 0.0173426234 | per-response usage.cost | review 0.0052962234 (in 3,163 / out 2,542, provider Baidu — list price honoured, no Venice-style excursion on the seat that produced one at S022) + vote 0.0120464 (in 3,975 / out 719, provider xAI). Both stop. Verdict C, unanimous, retiring a sense the list has carried since run one. Landed at 19% of the worst case. | | 2026-08-01 | S085 | the whole principal unit — the second long work chosen, span 1 translated, the study limb | 0.00 | 0.000000 | — | No API call. The three-route contamination search, the 2,699-word Finnish→English span with its 23-decision frozen log, the ten-rule binding register, the 573-paragraph source preparation, and the 23-locus comparison against Hertzberg's 1886 Swedish are all lead work. Lead translation is free and is never ledgered (charter §3, A4). |

2026-08-01 day total: $3.591619074 of $5.00 — nine sessions (S077 $0.580589442, S078 $1.323350575, S079 $0.623900826, S080 $0.397105980, S081 $0.187826153, S082 $0.00, S083 $0.390659225, S084 $0.035414750, S085 $0.0527721234). $1.408380926 headroom. 72% of the day's cap, and the first day in this ledger's history to carry nine sessions. (The key's own usage_daily reads 4.158489928 at close against the nine-session ledgered total; the $0.567 difference is not attributable to this project by anything this session can see — S077 and S080 both recorded the same shape on this day — and is recorded rather than claimed.) ARM-tierD-run stages 3–5 reserve $2.795 and still do not fit; deferred a third time, on arithmetic rather than on judgment.

| 2026-08-02 | S086 | E-20260801f-tierD-run stage 3, the held-out arm — 2 items × 2 orderings × 3 jurors | 0.579 full-stage reservation (§10), reserved against $5.0000 headroom on a fresh UTC day | 0.161661344 | per-response usage.cost, re-summed from the stored bodies by analysis/verify.py; opening key snapshot runs/snap/session-open-S086.json 42.262070476, usage_daily 0 | Caps P1=3500, P2=6000, P5=10000, unchanged from S083. 12 of 12 accepted, 0 failures, 513 s. 28% of the reservation. The arm did not fire (3 to Garnett, 0 to Hapgood, 3 split) — the pre-registered outcome, and the one that leaves a detection claim alive. One call breached its declared per-call worst case: H-HA__o0__P5 on SiliconFlow, 12,237 completion tokens against a max_tokens of 10,000, finish_reason: stop, $0.04140828 against $0.038775 — note (bgk), second firing, first on this vendor. No budget consequence; the stage reservation absorbed it eight times over. | | 2026-08-02 | S086 | E-20260801f stage 4, targeted heavy — 4 items × 2 orderings × 3 jurors | 1.108 full-stage reservation, reserved against $4.8383 headroom | 0.189338748 | per-response usage.cost | 24 of 24 accepted, 0 failures, 370 s. 17% of the reservation. Detection fires at ceiling: 12 of 12 units at +1, none −1, all three jurors 4 of 4 taken separately. Specificity does not fire — drop(naturalness) +1.12 against §6.6's ≤ 0.75 — and that one number is the whole verdict. | | 2026-08-02 | S086 | E-20260801f stage 5, targeted light (drawn 3 sites) — 4 items × 2 orderings × 3 jurors | 1.108 full-stage reservation, reserved against $4.6490 headroom | 0.162968897 | per-response usage.cost | 24 of 24 accepted, 0 failures, 291 s. 15% of the reservation, and the cheapest of the three stages. Detection fires at ceiling and all three specificity conditions fire (drop(naturalness) +0.50). Not claimed as a pass: the light cell does not gate, and making it the primary after watching it pass is the move the discipline exists to stop (RS-20260802-tierD-verdict §3). | | 2026-08-02 | S086 | the verdict page, the four sensitivity analyses, the verifier extension and all scoring | 0.00 | 0.000000 | — | No API call. analysis/stages345.py (new), the stage-3–5 half of analysis/verify.py (1,664 checks, 0 failures, nine mutation tests, nine caught and each asserted to change bytes on disk, note (bgu)), the P5-excluded sensitivity, prediction 8, the F3 within-stratum check and the five-sense recomputation after literary-quality's retirement are all lead work or local computation. tools/run_tierD.py, tools/build_index.py and tools/check_balance.py used unmodified; nothing in tools/ changed this session. |

S086's declared worst case for what it dispatched was $2.795 (the three full-stage reservations). Actual $0.513968989 = 18%. The whole 96-call experiment, across S083 and S086, cost $0.812454145 against a design worst case of $4.432 — 18%, and its central estimate of $1.16 — 70%. Nothing was wasted: 48 of 48 dispatched calls were accepted, no retry was needed, no body failed a guard, and no .attemptN.json or .err file exists.

THE KEY-USAGE CROSS-CHECK LEAVES $0.025883017 UNACCOUNTED, AND NO PROJECT BODY IS MISSING. Opening snapshot 42.262070476 (taken immediately before the first dispatch, usage_daily 0 — a fresh UTC day with the full cap intact), closing 42.801922482, delta 0.539852006; the key's own usage_daily reads 0.539852006 at close, identical to the delta, so the whole of today's key spend is that figure. The per-request sum over the 48 stored bodies is 0.513968989, leaving $0.025883017. Unlike S080's residual this one has no in-session explanation: 48 dispatches, 48 stored bodies, zero discarded attempts, zero error files — checked by listing rather than assumed. Concurrent non-project key use is the standing attribution at this order of magnitude (S042, S045, S049) and settling lag from S085's 2026-08-01 spend is the other; both are recorded rather than one being asserted. The per-request sum is what is ledgered, per this page's stated method, which is also the more conservative direction for the project's own accounting.

2026-08-02 day total: $0.513968989 of $5.00 — one session (S086 $0.513968989). $4.486031011 headroom. 10% of the day's cap. The stages deferred three times on arithmetic at S083, S084 and S085 fitted on the first fresh day with $2.20 of the reservation to spare, exactly as RS-20260801f §8 predicted.

| 2026-08-02 | S087 | E-20260802-voice-warrant pre-run critic — deepseek/deepseek-v4-pro (P5), a non-rater seat | 0.079 worst case from max_tokens 8,000 (note (abc)) → raised to 0.158 in session, amendment A1 | 0.030478153 | per-response usage.cost; opening key snapshot runs/snap/session-open-S087.json 43.319309166, usage_daily 1.05723869 | Two dispatches, and the first returned nothing. At max_tokens 8,000 the seat spent 7,999 of 8,000 completion tokens on reasoning and returned zero characters, finish_reason: length, billing $0.014185606302 — note (abc)'s sharpened form, the cap bounds reasoning as well as visible output, firing on the session's first call. Cap raised to 24,000; second dispatch stop, 5,947 chars, $0.01629254655, provider StreamLake. Verdict NEEDS-AMENDMENT, eight findings, two BLOCKING, all eight accepted, one prescribed remedy accepted with a written modification. Note (rr), thirty-fourth consecutive session. Both BLOCKING findings were about the CONTROL, not the argument, and both were right: the critic prompt contained the design in full, whose §3.2 tabulates the item strata, so the home assignment taken alongside it was discarded and re-taken blind. | | 2026-08-02 | S087 | E-20260802-voice-warrant blind home assignment — deepseek/deepseek-v4-pro (P5), stateless, items + definitions only | 0.079 | 0.023229 | per-response usage.cost | Provider Parasail, stop, 79.8 s, 468 chars, 24 of 24 lines. Amendment A3. Matched the lead's strata on 9 of 24, against the same seat's 15 of 24 when it had the design in hand — the framing effect the critic predicted, measured. | | 2026-08-02 | S087 | E-20260802-voice-warrant second home assignment — moonshotai/kimi-k3 (P4) | 0.180 → 0.600 after amendment A4 | 0.5030868 | per-response usage.cost | THREE COMPLETED DISPATCHES, ZERO CHARACTERS OF CONTENT, AND THE MOST EXPENSIVE LINE IN THIS RUN BY A FACTOR OF TWO. 5,997 of 6,000 reasoning tokens ($0.101382), then 5,997 of 6,000 again ($0.0958524), then — after the cap was raised to 20,000 — 19,997 of 20,000 ($0.3058524, 682 s), finish_reason: length every time. A fourth attempt was stopped in flight and billed nothing. The design's amendment A4 had pre-declared the fallback and A5 took it: the agreed subset is two-way. The verdict does not depend on the seat — F3's floor is 16 and P5 supplies only 9. New note (bhf): measure a seat's reasoning appetite before a registered control depends on it; max_tokens is the wrong instrument for measuring it. | | 2026-08-02 | S087 | E-20260802-voice-warrant stage 1 — WARRANT and EXTENT, 24 items × 2 axes × 3 seats, order counterbalanced | 0.672 (3 seats × 2 calls × 2 attempts, max_tokens 8,000) | 0.156291550 | per-response usage.cost | 6 of 6 accepted on the first attempt, 24 of 24 answer lines each, all stop, 190 s total. P1 openai/gpt-5.6-terra 0.010361 + 0.01884925 (OpenAI); P2 google/gemini-3.6-flash 0.036036 (Google AI Studio) + 0.0434085 (Google); P3 x-ai/grok-4.5 0.0111344 + 0.0365024 (xAI). 23% of the declared worst case. No cap was breached — note (bgk) checked after the fact by the verifier, 10 of 10 bodies within cap. | | 2026-08-02 | S087 | the translation, both gates, the item set, all scoring and verification | 0.00 | 0.000000 | — | No API call. Heine «Die Harzreise» ¶124 (421 German words → 466 English, 30-decision frozen log) and the ¶118 gate span (203 → 226), the two-comparator seven-cell contamination gate, the 24-item set, analysis/score.py and the 131-check analysis/verify.py with two mutation tests (both asserting byte change, note (bgu); every written file snapshotted and restored, note (bhd)) are all lead work or local computation. tools/dependence_check.py, tools/build_index.py and tools/check_balance.py used unmodified; nothing in tools/ changed this session. call.py copied unchanged from E-20260801g. Lead translation is free and is never ledgered (charter §3, A4). |

S087's declared worst case, after amendments A1 and A4, was $1.68. Actual $0.713085503 = 42% — and the fraction is high for the worst reason: 71% of the spend went to a seat that returned no data. Against the original $0.70 declaration it is 102%, i.e. the run overran its first estimate, and note (bhf) is what that bought.

THE KEY-USAGE CROSS-CHECK IS EXACT — AND ONLY BECAUSE TWO DESTROYED BODIES WERE WRITTEN DOWN BEFORE THEY WERE DESTROYED. Opening snapshot 43.319309166, closing 44.032394668, delta 0.713085502. The per-request sum over the 10 surviving stored bodies is 0.597517897; two billed bodies were overwritten when amendments A1 and A4 re-invoked their stages and call.py restarted its attempt counter at 1 ($0.014185606302 and $0.101382, together $0.115567606). 0.597517897 + 0.115567606 = 0.713085503, agreeing with the key delta to 1e-9. Note (bgz) is amended for this: its S080 repair labels every attempt uniquely and does not cover a collision between invocations; workshop/experiments/E-20260802-voice-warrant/runs/OVERWRITTEN.md carries the record.

The $0.543 that is not this project's, and it reconciles too. The key's usage_daily reads 1.770324192 at close against a ledgered day total of 1.227054492. The difference is $0.543269701, and it decomposes exactly: $0.025883017 is S086's own unattributed residual (recorded there) and $0.517386684 was billed to the key between S086's closing snapshot and S087's opening one, by something no session can see. Concurrent non-project key use is the standing attribution at this order of magnitude (S042, S045, S049, S086) and it is recorded rather than asserted. The per-request sum is what is ledgered, per this page's stated method.

2026-08-02 day total: $1.227054492 of $5.00 — two sessions (S086 $0.513968989, S087 $0.713085503). $3.772945508 headroom. 25% of the day's cap.

| 2026-08-02 | S088 | E-20260802b-pole-uniformity pre-run critic — x-ai/grok-4.5 (P3), a seat taking no part in the coding stage | 0.20 worst case from max_tokens 24,000 (notes (abc), (bhf)) | 0.0673904 | per-response usage.cost; opening key snapshot runs/snap/session-open-S088.json 44.231851868 | One dispatch, stop, 7,927 chars, 175.2 s, 6,136 of 8,003 completion tokens on reasoning. Verdict NEEDS-AMENDMENT, nine findings, five BLOCKING, all nine accepted — note (rr), thirty-fifth consecutive session. Every BLOCKING remedy was taken as written; two of them (omission-codes-D, the asymmetric licensing rule) move against the result the lead expected, which is the only defence available to amendments taken after the renderings had been read. | | 2026-08-02 | S088 | E-20260802b independent coding, seat 1 — openai/gpt-5.6-terra (P1), stateless, site book + six renderings only | 0.15 | 0.023322 | per-response usage.cost | Provider OpenAI, stop, 40.2 s, 349 chars, 21 of 21 answer lines, first attempt. Never shown the design, the hypothesis or the lead's codes. | | 2026-08-02 | S088 | E-20260802b independent coding, seat 2 — google/gemini-3.6-flash (P2), same payload | 0.15 | 0.0619755 | per-response usage.cost | Provider Google, stop, 35.1 s, 349 chars, 21 of 21 lines, first attempt; 7,433 of 7,649 completion tokens on reasoning for a 349-character answer — the reasoning tax on a mechanical coding task, and the reason this seat cost 2.7× the other for identical output. | | 2026-08-02 | S088 | D-20260802-12 ratification (the gate, not the principal unit) — independent adversarial review deepseek/deepseek-v4-pro (P5) + routed review vote x-ai/grok-4.5 (P3) | 0.15 | 0.017744733 | per-response usage.cost | review 0.005760183 (provider Ionstream, 45.7 s) + vote 0.0119844 (xAI, 16.6 s). Verdict B, unanimous, both against S087's provisional default of A — the third consecutive ratification to go against a no-change default. Applied to wiki/goodness-senses.md the same session. | | 2026-08-02 | S088 | the two translations, the contamination gate, the site book, the lead's coding, all scoring and all verification | 0.00 | 0.000000 | — | No API call. Iliad XII.310–327 rendered twice from the Greek alone (R07 184 words, R08 165, 20-decision logs each), the XII.331–341 gate span and its two-comparator dependence_check (0 shared 7-grams, longest runs 6 and 5 against the published pair's own 4), the 21-site book, analysis/score.py and the 181-check analysis/verify.py with three mutation tests (all caught, each asserting the bytes on disk changed, note (bgu)) are all lead work or local computation. tools/dependence_check.py, tools/build_index.py and tools/check_balance.py used unmodified; nothing in tools/ changed this session. call.py copied unchanged from E-20260802-voice-warrant. Lead translation is free and is never ledgered (charter §3, A4). |

S088's declared worst case was $0.50 (design §10) and the ratification gate was budgeted separately at $0.15. Actual $0.170432483 — 26% of the $0.65 reserved. Five dispatches, five accepted on the first attempt, zero retries, zero discarded attempts, zero .err files, and no call breached its declared cap (note (bgk) checked after the fact).

THE KEY-USAGE CROSS-CHECK IS EXACT. Opening snapshot 44.231851868, closing 44.402284351, delta 0.170432483, against a per-request sum over the five stored bodies of 0.170432483 — agreement to 1e-9, on five calls across four providers, and the second exact cross-check in three sessions. Both snapshots written to disk before being read (note (bco)); no body was overwritten, so the S087 collision (note (bgz) as amended) did not recur.

2026-08-02 day total: $1.397486975 of $5.00 — three sessions (S086 $0.513968989, S087 $0.713085503, S088 $0.170432483). $3.602513025 headroom. 28% of the day's cap.

| 2026-08-02 | S089 | E-20260802c-regime-scoring pre-run critic (P3 x-ai/grok-4.5, 1 call) | 0.13 | 0.062304400 | per-response usage.cost | raw: workshop/experiments/E-20260802c-regime-scoring/runs/. Overran the design's own stated $0.08 line and the overrun is recorded rather than the line re-based. Verdict NEEDS-AMENDMENT, 10 findings, all accepted before dispatch | | 2026-08-02 | S089 | E-20260802c-regime-scoring scoring block: 10 items × 2 orderings × 3 jurors | 2.696 (full-stage reservation) | 0.515743035 | per-response usage.cost, summed | raw: .../runs/scores/. 60 of 60 accepted first attempt, 0 parse failures, 0 transport errors. Realised at 19% of the reservation, the same proportion as the S086 Tier D stage |

2026-08-02 day total: $1.975534410 of $5.00 — four sessions (S086 $0.513968989, S087 $0.713085503, S088 $0.170432483, S089 $0.578047435). $3.024465590 headroom. 40% of the day's cap, and the largest single-session spend of the four.

The cross-check is NOT exact and the gap is stated rather than absorbed. Opening 45.164290590, closing 45.896845469, delta $0.732554879 against a per-request sum of $0.578047435 — $0.154507444 unattributed. Per this page's stated method the per-request sum is primary; the residual is the same order and shape as the $0.543269701 of non-project drift this same UTC day already carries on S087's row, and it is recorded as unattributed rather than assigned.

| 2026-08-02 | S090 | E-20260802d-deference-carry pre-run critic (P5 deepseek/deepseek-v4-pro, 1 accepted body) | 0.15 worst case from max_tokens 24,000 (notes (abc), (bhf)) | 0.014593728 | per-response usage.cost | raw: workshop/experiments/E-20260802d-deference-carry/runs/. Provider DigitalOcean, stop, 10,820 chars, 38.6 s, 0 reasoning tokens. Verdict NEEDS-AMENDMENT, five findings, three BLOCKING, all five accepted — note (rr), thirty-sixth consecutive session. Two amendments (A2 kin-term code, A3 unframed arm) made the design able to refute the lead's own register rule, and one of them did | | 2026-08-02 | S090 | E-20260802d stage 1 — contrast-subject translations, 2 sites × 3 seats + 1 unframed arm | 0.52 (9 calls × 2 attempts, max_tokens 4,000) | 0.094170700 | per-response usage.cost, re-summed from the stored bodies by analysis/verify.py | 9 of 9 accepted on the first attempt, all stop, 0 failures, 0 retries, 0 .err files, no body above its cap (note (bgk), checked on every body by the verifier). Routing (note (x)): P1 OpenAI ×3, P2 Google ×2 / Google AI Studio ×1, P3 xAI ×3. P2 spent 2,053 / 2,498 / 2,882 completion tokens on 767–1,300-character translations — 4–6× P1's cost for the same passage, the third instance of the note (bhf) shape | | 2026-08-02 | S090 | span 2 of «Köyhää kansaa», the 13-locus Hertzberg comparison, all coding and all verification | 0.00 | 0.000000 | — | No API call. The 1,782 → 2,758-word span with its 23-decision frozen log, the two errata, the alignment of Finnish ¶77–134 to Swedish ¶74–131, the 13-locus classification and the 61-check analysis/verify.py with three mutation tests (all caught, each asserting the bytes on disk changed, note (bgu); every mutated file snapshotted and restored, note (bhd)) are lead work or local computation. tools/build_index.py and tools/check_balance.py used unmodified; nothing in tools/ changed this session. call.py copied unchanged from E-20260802b-pole-uniformity. Lead translation is free and is never ledgered (charter §3, A4) |

S090's declared worst case was $0.67 after the critic's amendments raised the subject count 6 → 9. Actual $0.108764428 — 16% of the reservation. Ten dispatches, ten accepted on the first attempt.

AN EXACT CROSS-CHECK, AND THE FIRST THIS LEDGER HAS ON A KILLED DISPATCH. The critic prompt was first sent in the foreground and killed in flight at a 2-minute tool timeout before the run was backgrounded; whether it had billed was open. Session-open snapshot 46.007974247, snapshot after the accepted critic body 46.022567975, delta 0.014593728 — the accepted body's per-request cost to 1e-9. The killed dispatch billed exactly zero, established rather than assumed, which is the question note (bcx) usually has to leave open.

Stage 1's own delta is 0.019216000 (46.022567975 → 46.041783975) against a per-request sum of 0.094170700: $0.074954700 still unsettled at close, the ordinary lag direction and not the (bet) excess direction. Per-request sums are ledgered, per this page's stated method. Between-session drift from S089's close (45.896845469) to this session's open was $0.111128778, consistent with S089's own $0.154507444 residual settling after its closing snapshot.

2026-08-02 day total: $2.084298838 of $5.00 — five sessions (S086 $0.513968989, S087 $0.713085503, S088 $0.170432483, S089 $0.578047435, S090 $0.108764428). $2.915701162 headroom. 42% of the day's cap.

| 2026-08-02 | S091 | E-20260802e-displaced-marking pre-run critic (P5 deepseek/deepseek-v4-pro, 1 body) | 0.20 worst case from max_tokens 24,000 (note (abc)) | 0.009172451 | per-response usage.cost | raw: workshop/experiments/E-20260802e-displaced-marking/runs/. Provider StreamLake, stop. Verdict NEEDS-AMENDMENT, four findings, two BLOCKING, all four accepted — note (rr), thirty-seventh consecutive session. A1 (an independent naturalness audit of the decoy) and A2 (a manual audit of all 126 census windows, and the population claim downgraded) are the two that could hurt the lead | | 2026-08-02 | S091 | E-20260802e stage 1 — 3 seats × 2 rotations grading 4 renderings × 6 sites | 0.55 | 0.220316800 | per-response usage.cost, re-summed from stored bodies by analysis/verify.py | 6 accepted, 4 rejected. P1 and P3 clean on first attempt. P2 gemini-3.6-flash returned finish_reason: length on all four attempts at max_tokens 4,000 (4,177–5,691 reasoning chars) and recovered at 16,000 under declared amendment A5. The four failed bodies billed $0.133473 | | 2026-08-02 | S091 | E-20260802e stage 1b — the decoy naturalness audit (P4 moonshotai/kimi-k3), critic amendment A1 | 0.25 | 0.298161000 | per-response usage.cost | 1 accepted, 1 rejected. The rejected body returned length with 16,803 reasoning characters and ZERO content characters for eighteen one-line ratings; it billed $0.062820. The A5 retry at max_tokens 20,000 routed to Fireworks rather than Moonshot AI and cost 3.7× the failed attempt — the S022 routing caution, live. This stage overran its own $0.25 line and the overrun is recorded rather than the line re-based | | 2026-08-02 | S091 | E-20260802e stage 2 — followability, 3 seats × 20 frozen log excerpts | 0.25 | 0.269987400 | per-response usage.cost | 3 accepted, 2 rejected (both P2, both length, $0.073197, recovered under A5). Result: pairwise κ 0.630 / 0.631 / 0.774, all above the pre-registered 0.60 bar | | 2026-08-02 | S091 | the census, the six forced re-renderings, the decoys, the positive controls, all classification and all verification | 0.00 | 0.000000 | — | No API call. 126 → 61 → 13 census, the 8 Class A / 8 Class B classification, the four renderings at each of six sites in four source languages, and the 62-check analysis/verify.py with three mutation tests (all caught, each asserting the bytes changed, note (bgu); every mutated file snapshotted and restored, note (bhd)) are lead work or local computation. tools/ unmodified. call.py copied unchanged from E-20260802d. Lead translation is free and is never ledgered (charter §3, A4) |

S091's declared worst case was $1.20, raised to $1.40 by amendment A5. Actual $0.797637651 — 57% of the reservation, the highest fraction of a declared worst case this project has spent, and the reason is on the record rather than in the noise.

$0.269490000 — 34% of the spend — bought nothing. Seven of eighteen bodies returned finish_reason: length and were rejected under note (b); all seven were hidden-reasoning overruns on two seats, and all seven recovered at a raised max_tokens with byte-identical prompts. Note (bhf) has fired three times on price; this is its first firing on capacity, and note (bhq) now records that max_tokens must be sized from a seat's measured reasoning appetite plus the answer, never from the answer alone.

2026-08-02 day total: $2.881936489 of $5.00 — six sessions (S086 $0.513968989, S087 $0.713085503, S088 $0.170432483, S089 $0.578047435, S090 $0.108764428, S091 $0.797637651). $2.118063511 headroom. 58% of the day's cap, and the largest single-session spend of the six.

| 2026-08-02 | S092 | E-20260802f-licensed-strangeness pre-run critic (P2 google/gemini-3.6-flash, a seat taking no part in the rating stage) | 0.19 worst case from max_tokens 24,000 (notes (abc), (bhq)) | 0.079264500 | per-response usage.cost; opening key snapshot runs/snap/session-open-S092.json 48.049987491 | Provider Google, stop, 5,452 chars, 39.3 s, 7,119 of 8,574 completion tokens on reasoning. Verdict NEEDS-AMENDMENT, five findings, three BLOCKING, all five accepted — note (rr), thirty-eighth consecutive session. Two could have voided the run: A2 struck a LICENSED edit that was a forced etymological gloss rather than a feature of the Greek, and A1 computed the design's own matching criterion and found it failing at 38.3% against a 30% ceiling before any datum existed — note (bhr) | | 2026-08-02 | S092 | E-20260802f condition N (no source) — 3 seats × 13 attribution items + 4 naturalness ratings | 0.38 | 0.060470664 | per-response usage.cost, re-summed from stored bodies by analysis/verify.py | 3 of 3 accepted on the first attempt, all stop, 17 of 17 answer lines each. P1 openai/gpt-5.6-terra (OpenAI) $0.014977; P3 x-ai/grok-4.5 (xAI) $0.0394764; P5 deepseek/deepseek-v4-pro (StreamLake) $0.006017263824 | | 2026-08-02 | S092 | E-20260802f condition S (the Greek supplied) — same 3 seats, same items, dispatched after every condition-N body had returned | 0.38 | 0.064910571 | per-response usage.cost | 3 of 3 accepted on the first attempt. P1 $0.01575975; P3 $0.0415484; P5 $0.007602421398. Six of the seven bodies in this session spent over 95% of their completion tokens on reasoning for a 143-character answer; no body came within 7,000 tokens of its cap, note (bgk) checked by the verifier on every body | | 2026-08-02 | S092 | the four arms, the edit sets, the contamination gate, all scoring and all verification | 0.00 | 0.000000 | — | No API call. Three derived renderings of Iliad XII.310–327 with a frozen edit log, the nine-cell dependence_check gate, analysis/score.py and the 81-check analysis/verify.py with two mutation tests (both caught, each asserting the bytes on disk changed, note (bgu); every mutated file snapshotted and restored, note (bhd)) are lead work or local computation. tools/dependence_check.py, tools/build_index.py and tools/check_balance.py used unmodified; nothing in tools/ changed this session. call.py copied unchanged from E-20260802e. Lead translation is free and is never ledgered (charter §3, A4) |

S092's declared worst case was $1.20. Actual $0.204645735 — 17% of the reservation, and 7 of 7 dispatches were accepted on the first attempt: no retries, no .err files, no finish_reason: length, nothing rejected. That is the direct answer to S091, where 34% of the spend bought nothing; the caps were sized from note (bhq) rather than from the answer length, and the reasoning tokens (2,056–7,119 for a 143-character answer) show the note was measuring the binding quantity.

One exact cross-check and one named residual. Session-open 48.049987491 → the snapshot taken when the rating stage opened, 48.129251991: delta 0.079264500, the critic body's per-request cost to 1e-9. Session-open → final close (48.247030804) is 0.197043313 against a per-request sum of 0.204645735; the residual $0.007602422 equals the last dispatched body's cost to 1e-9, i.e. that call had not settled when the closing snapshot was taken — the ordinary lag direction, note (bcx), and for once fully attributable. Per-request sums are ledgered, per this page's stated method. All snapshots written to disk before being read (note (bco)).

2026-08-02 day total: $3.086582224 of $5.00 — seven sessions (S086 $0.513968989, S087 $0.713085503, S088 $0.170432483, S089 $0.578047435, S090 $0.108764428, S091 $0.797637651, S092 $0.204645735). $1.913417776 headroom. 62% of the day's cap.

| 2026-08-02 | S093 | D-20260802-13 ratification (the gate, not the principal unit) — independent adversarial review, moonshotai/kimi-k3 (P4), first dispatch, REJECTED | 0.25 worst case from max_tokens 14,000 (notes (abc), (bhq)) | 0.221532 | per-response usage.cost; opening key snapshot session-open-S093.json 48.390045095 | finish_reason: length. 14,000 completion tokens, ZERO content characters. Provider Together. This bought nothing and it is 59% of the session's entire spend. Note (b) was written at S044 about this same model doing this same thing, and its prescribed fix — cap the reasoning effort, a generous max_tokens alone is not enough — was not applied on the first dispatch. The note is amended for it: on a seat with a recorded history, the cap is not advisory | | 2026-08-02 | S093 | same review, retry under note (b)'s own fix (--reasoning-effort low, max_tokens 16,000) | included above | 0.047242200 | per-response usage.cost | Provider Together, stop, 50.9 s, 2,392 completion tokens. Verdict C, eight findings, one BLOCKING — and the BLOCKING one is real: RS-20260802f's naturalness item quotes the sense minus the clause, so the run never varied the clause and "the clause is inert" is withdrawn. The retry cost 21% of the failure it replaced | | 2026-08-02 | S093 | D-20260802-13 routed vote — openai/gpt-5.6-terra (P1), a blind rating seat in E-20260802f that never saw the design, hypothesis or results | 0.10 | 0.019125500 | per-response usage.cost | Provider OpenAI, stop, 23.3 s. Verdict C, accepts all eight findings, id perceived-source-carriage, six conditions. Applied to wiki/goodness-senses.md the same session | | 2026-08-02 | S093 | E-20260802g-programme-divergence — 22-site rule-application coding, 2 blind seats | 0.20 worst case from max_tokens 14,000 (note (bhq)) | 0.086262800 | per-response usage.cost, re-summed from the stored bodies by analysis/verify.py | 2 of 2 accepted on the first attempt, both stop, 22 of 22 answer lines each, 0 retries, 0 .err files. P5 deepseek/deepseek-v4-pro (provider Alibaba) $0.0313644, 115.1 s; P3 x-ai/grok-4.5 (provider xAI) $0.0548984, 159.4 s. 43% of the reservation. Both registered controls fired clean | | 2026-08-02 | S093 | the complete reading, both regimes, both translations, the contamination gate, the site book, the lead's coding, all analysis and all verification | 0.00 | 0.000000 | — | No API call. The 438,198-character Arnold–Newman volume read entire; R12 and R13 written; Iliad XXIV.486–512 rendered twice from the Greek alone with 22- and 20-decision frozen logs; the XXIV.552–570 gate span and its two-comparator dependence_check; the 22-site book; the lead's own 22 codes; and the 32-check analysis/verify.py with two mutation tests (both caught, each asserting the bytes on disk changed, note (bgu); every mutated file snapshotted and restored, note (bhd)) are all lead work or local computation. tools/dependence_check.py, tools/ratify_vote.py, tools/build_index.py and tools/check_balance.py used unmodified; nothing in tools/ changed this session. Lead translation is free and is never ledgered (charter §3, A4) |

S093's declared worst case was $0.55 across the two stages. Actual $0.374162500 — 68% of the reservation, and the highest fraction any session has spent (the previous high was S091's 57%). The reason is one rejected body: $0.221532, 59% of the spend, for zero content characters. Excluding it, the session spent $0.15263 on five accepted bodies against a $0.55 reservation — 28%, the ordinary range. The lesson is not that the cap was too small; it is that note (b) existed, named the model, prescribed the fix, and was not applied.

THE KEY-USAGE CROSS-CHECK IS EXACT, AND FOR THE FIRST TIME THE RESIDUAL IS ZERO. Opening snapshot 48.390045095, closing 48.764207595, delta 0.374162500, against a per-request sum over the five stored bodies of 0.374162500 — agreement to 1e-9 with a residual of exactly 0.000000000, across five calls on four providers. No lag, no drift, nothing unattributed. Both snapshots written to disk before being read (note (bco)); no body was overwritten. Between-session drift from S092's close (48.247030804) to this session's open was $0.143014291, unattributed and consistent with the concurrent non-project key use this ledger has recorded at this order of magnitude since S042.

2026-08-02 day total: $3.460744724 of $5.00 — eight sessions (S086 $0.513968989, S087 $0.713085503, S088 $0.170432483, S089 $0.578047435, S090 $0.108764428, S091 $0.797637651, S092 $0.204645735, S093 $0.374162500). $1.539255276 headroom. 69% of the day's cap.

| 2026-08-03 | S094 | E-20260803-a4-set pre-run critic (P4 moonshotai/kimi-k3, a seat taking no part in any other stage) | 0.19 worst case from max_tokens 16,000 (notes (abc), (bhq)) | 0.070260 | per-response usage.cost; opening key snapshot runs/snap/session-open-S094.json 49.914103845 | Provider Moonshot AI, stop, 133.9 s, 6,136 chars. Note (b)'s prescribed fix — reasoning: {"effort": "low"} — was applied on the FIRST dispatch to the seat the note is about, against S093, where it was not and the first dispatch billed $0.221532 for zero content characters. Verdict NEEDS-AMENDMENT, eleven findings, three BLOCKING, all eleven accepted — note (rr), streak restored after S093's breach. One of the three accepted BLOCKING amendments then introduced note (bhs)'s defect | | 2026-08-03 | S094 | E-20260803-a4-set edit-neutrality check (P3 x-ai/grok-4.5), critic amendment A8 | 0.04 | 0.010834 | per-response usage.cost | Provider xAI, stop, 19.9 s, 11 of 11 answer lines. Verdict "quality-neutral" with a predicted direction (naturalness up, voice/style-correspondence down) that the run's measured floor agrees with on sign. A check, not a gate: the edit list was frozen and was not revised on it | | 2026-08-03 | S094 | E-20260803-a4-set log-reading stage — P3, six frozen translator's logs, dispatched in full before the first rating call | 0.22 | 0.095542 | per-response usage.cost | 6 of 6 accepted on the first attempt, all stop, 19.6–38.9 s. The seat saw no source and no translation | | 2026-08-03 | S094 | E-20260803-a4-set rating stage — 3 jurors × 8 items × 2 passes, one text per call, strictly sequential | 1.80 across both passes | 0.472781 | per-response usage.cost, re-summed from the stored raw bodies by analysis/verify.py | 48 accepted, 1 rejected. rate_PAN_P5_p1 returned finish_reason: length and billed $0.016288625 for nothing — 2.5% of the session, against S093's 59%; it recovered on a byte-identical retry to the same slug. P5 routed across ten providers in 32 calls with no price excursion | | 2026-08-03 | S094 | the contamination gates, the fresh translation, the null's edit list, all scoring and all verification | 0.00 | 0.000000 | — | No API call. Three gate runs of tools/dependence_check.py (two firing, one re-checking a published figure and reproducing it), the 220-word Daudet and 233-word Sand gate translations, the 503-word T-clos-des-ames-R04-v1 and its ten-edit paraphrase, analysis/score.py and the 640-check analysis/verify.py with three mutation tests (all caught, each asserting the bytes on disk changed, note (bgu); every mutated file snapshotted and restored, note (bhd)) are lead work or local computation. tools/dependence_check.py, tools/build_index.py and tools/check_balance.py used unmodified; nothing in tools/ changed this session. Lead translation is free and is never ledgered (charter §3, A4) |

S094's declared worst case was $2.21 across four stages. Actual $0.649417427 — 29% of the reservation, and 54 of 55 dispatches were accepted on the first attempt.

THE KEY-USAGE CROSS-CHECK IS THE CLOSEST THIS LEDGER HAS RECORDED. Opening snapshot 49.914103845, closing 50.563521268, delta 0.649417423, against a re-sum over all 55 stored raw bodies (including the rejected one) of 0.649417427 — a residual of −4 × 10⁻⁹, which is floating-point and not lag. S093 was the first exact zero on five calls; this is fifty-five calls across five slugs and thirteen providers. Both snapshots written to disk before being read (note (bco)). Between-session drift from S093's close (48.764207595) to this session's open was $1.149896250, unattributed and larger than the usual order of magnitude for the concurrent non-project key use this ledger has recorded since S042 — noted rather than explained, since a whole UTC day elapsed between the two sessions.

2026-08-03 day total after S094: $0.649417427 of $5.00. $4.350582573 headroom.

S095's worst case was $1.94 as first frozen and $0.94 after the second critic pass halved the run. Actual $0.263880726 — 28% of the revised reservation, and 20 of 20 dispatches accepted on the first attempt.

The key-usage cross-check closes at a residual of −2 × 10⁻⁹, the closest this ledger has recorded. Opening snapshot 50.567528368, closing 50.831409092, delta 0.263880724, against a re-sum over all 20 stored raw bodies of 0.263880726. Between-session drift from S094's close (50.563521268) to this session's open was $0.004007100 — two orders of magnitude smaller than the gap at S094's own open, and these two sessions ran on the same UTC day.

2026-08-03 day total: $0.913298153 of $5.00 — two sessions (S094 $0.649417427, S095 $0.263880726). $4.086701847 headroom. 18% of the day's cap.

S096 — 2026-08-03 (UTC), E-20260803c-occupied-slot

S096's declared worst case was $2.10 across three stages. Actual $0.679379608 - 32% of the reservation, in the ordinary range.

The key-usage cross-check closes at a residual of exactly zero. Opening snapshot 51.112336122, final 51.791715730, delta 0.679379608, against a per-request re-sum over all 25 stored raw bodies of 0.679379608 - agreement to 1e-9, across 25 calls on ten providers, with the four rejected bodies included in both figures. Between-session drift from S095's close (50.831409092) to this session's open was $0.280927030 - an order of magnitude above the S094->S095 gap and consistent with the concurrent non-project key use this ledger has recorded since S042; it is not attributed to this project.

2026-08-03 day total: $1.592677761 of $5.00 - three sessions (S094 $0.649417427, S095 $0.263880726, S096 $0.679379608). $3.407322239 headroom. 32% of the day's cap.

S097 — 2026-08-03 (UTC), E-20260803d-class-uniformity

S097's declared worst case was $0.84, revised to $1.10 by amendment A1 after the first critic dispatch failed. Actual $0.520948932 — 47% of the revised reservation — and $0.260301 of it, exactly half, bought nothing. Without note (b)'s firing the session would have cost $0.26.

Five rejected bodies are preserved on disk and are billed in both the per-request re-sum and the totals above, per note (bdt).

The key-usage cross-check does NOT close this session, and the residual is recorded rather than reconciled. Opening snapshot 51.925936255, closing 52.624133459, delta 0.698197204, against a per-request re-sum over all 13 stored raw bodies of 0.520948932 — residual $0.177248272. Between-session drift from S096's close (51.791715730) to this session's open was a further $0.134220525. Both fall in the concurrent non-project key use this ledger has recorded since S042; per-request costs are primary and the delta is the sanity check (CLAUDE.md), so the figures above stand on the stored bodies.

2026-08-03 day total: $2.113626693 of $5.00 — four sessions (S094 $0.649417427, S095 $0.263880726, S096 $0.679379608, S097 $0.520948932). $2.886373307 headroom. 42% of the day's cap.

S098 — 2026-08-03 (UTC), D-20260803-14 ratification gate + E-20260803e-purpose-index

S098's declared worst case was $1.31 — $0.17 for the ratification gate and $1.14 for the experiment, both built from max_tokens (note (abc)). Actual $0.347714326, 27%. No dispatch failed, nothing was retried, and no body billed anything for nothing — the first session since S092 of which that is true.

The key-usage cross-check CLOSES this session exactly. Opening snapshot 52.811520842, closing 53.159235167, delta 0.347714325, against a per-request re-sum over all 14 stored raw bodies of 0.347714326 — agreement to 1e-9, and the first exact cross-check since S095. Both snapshots were written to disk before being read (note (bco)). Between-session drift from S097's close (52.624133459) to this session's open was $0.187387383, the concurrent non-project key use this ledger has recorded since S042; it is not attributed to this project.

2026-08-03 day total: $2.461341019 of $5.00 — five sessions (S094 $0.649417427, S095 $0.263880726, S096 $0.679379608, S097 $0.520948932, S098 $0.347714326). $2.538658981 headroom. 49% of the day's cap.

S099 — 2026-08-03 (UTC), D-20260803-15 ratification gate + E-20260803f-craft-carriers

S099's declared worst case was $1.09 — $0.30 for the ratification gate and $0.92 for the experiment, less the reserve double-count, both built from max_tokens (note (abc)). Actual $0.801520290, 74% — the closest this project has run to its own reservation, and the reason is not the estimate.

$0.264163127 — 33% of the session — bought nothing. Thirty of fifty-seven stored bodies returned finish_reason: length. Note (b) fired twice on P5 and its prescribed effort: low was not applied on the first dispatch to that seat, which is the identical omission S097 paid $0.260301 for; the two other causes were caps sized for the shape of an answer rather than its length. Every rejected body is preserved on disk and is billed in both the per-request re-sum and the totals above, per note (bdt).

The key-usage cross-check CLOSES this session. Opening snapshot 53.640545655, closing 54.442065942, delta 0.801520287, against a per-request re-sum over all 59 stored bodies of 0.801520290 — residual −3 × 10⁻⁹, the second exact cross-check in two sessions. Between-session drift from S098's close (53.159235167) to this session's open was $0.481310488, the concurrent non-project key use this ledger has recorded since S042; it is not attributed to this project.

2026-08-03 day total: $3.262861309 of $5.00 — six sessions (S094 $0.649417427, S095 $0.263880726, S096 $0.679379608, S097 $0.520948932, S098 $0.347714326, S099 $0.801520290). $1.737138691 headroom. 65% of the day's cap.

S100 — 2026-08-03 (UTC), E-20260803g-address-axis (one adversarial pass, ARM-atelier-cycle span 4)

$0.0641798 of $0.0975582 — 66% of the session — bought nothing: both P1 bodies and the orphan. The declared worst case was $0.42 and the actual is 23% of it, so the reservation held; what did not hold is the yield, and the two causes are one known note firing again and one new one.

The key-usage cross-check DOES NOT CLOSE, and the gap is the finding. Opening 54.776232642, closing 54.873790842, delta 0.097558200; the per-request re-sum over the three stored bodies is 0.077852200; residual +0.019706. That residual is not drift and is not non-project use: it is a fourth request, dispatched to P1 and billed at roughly 1,831 in × $1.25/M + ~2,300 out × $7.50/M ≈ $0.0196, whose client was killed by a shell timeout mid-flight. The ledger takes the key-usage delta, $0.097558200, as this session's actual — the re-sum is the number that is wrong here, which is the reverse of this ledger's usual ordering and is why the cross-check exists. Note (bid).

2026-08-03 day total: $3.360419509 of $5.00 — seven sessions (S094 $0.649417427, S095 $0.263880726, S096 $0.679379608, S097 $0.520948932, S098 $0.347714326, S099 $0.801520290, S100 $0.097558200). $1.639580491 headroom. 67% of the day's cap.

S101 — 2026-08-04 (UTC), E-20260804-displaced-marking-fr (ARM-framework-v01 step 3)

Declared worst case $1.50 after amendments ($1.10 before). Actual $0.252177609 — 17%. The margin is not luck: nothing was re-dispatched except the one re-dispatch the amendments required.

$0.0586302 of $0.252177609 — 23% of the session — bought a second opinion nobody asked for, and it was the best-value call of the run. The accidental duplicate returned the harsher verdict and three findings the first pass missed, one of which (the undefined 3-of-7 band) would have left the primary result to be interpreted after the data existed. It is recorded as waste in the accounting sense and as the opposite in every other.

The key-usage cross-check CLOSES this session EXACTLY. Opening snapshot 54.880387742, closing 55.132565351, delta 0.252177609; the per-request re-sum over the ten stored bodies is 0.193547409, plus the $0.058630200 recovered by hand for the overwritten pass-1 body (note (bih)) = 0.252177609, residual −0.000000000. Between-session drift from S100's close (54.873790842) to this session's open was $0.006596900, the concurrent non-project key use this ledger has recorded since S042; it is not attributed to this project.

| 2026-08-05 | S109 | stage 0 — the note-(bhf) seat probe, first paid use (P1, P2, P5; P2 re-probed at the run ceiling) | 0.01 | 0.002335 | per-response usage.cost | FIRED. P2 returned finish_reason: length with empty content twice at max_tokens 64 — S108's failure mode, caught for $0.000188. At the run ceiling P2 answered normally, at $0.001188 against P1's $0.000094 for the identical 24-character reply | | 2026-08-05 | S109 | pre-run adversarial critic, one pass (z-ai/glm-5.2) | 0.36 from max_tokens 24,000 | 0.03466904 | per-response usage.cost | NEEDS-REDESIGN, six findings, three BLOCKING, all accepted. Its accepted F5 amendment is what broke the retention instrument, and the canary caught it — see RS-20260805 §4 | | 2026-08-05 | S109 | stage 1 — French line-end specificity, 3 seats × 2 calls | 0.30 | 0.181251 | per-response usage.cost | 6 accepted bodies and 3 length failures costing $0.069, because A7 raised only stage 2's ceiling; A8 then raised every stage to 12,000 | | 2026-08-05 | S109 | stage 2 — flavour retention, 3 seats × 4 calls + 1 re-dispatch | 0.72 | 0.553 | per-response usage.cost, re-summed by analysis/verify.py from the stored raw bodies | 12 accepted, 4 seat failures. P5 group 2 needed max_tokens 20,000 and 192 s | | 2026-08-05 | S109 | stage 3 — rhyme scheme read + group re-coding, 1 seat × 3 calls | 0.11 | 0.053565 | per-response usage.cost | 3 of 3 accepted, and the bodies are unusable: 61.7% agreement with the written rule, wrong in both directions on textbook rhymes. FC3 fires | | 2026-08-05 | S109 | 56 lines of Heredia rendered under R16/R17, the contamination gate, the F3 pixel verification, the rhyme census, site selection and verification | 0.00 | 0.000000 | — | No API call. Lead translation is free and is never ledgered (charter §3, A4) |

S109 total, summed per-request over all 34 stored billed bodies: $0.860694702, against a declared worst case of $1.50 — 57%. Eight seat failures cost $0.197815245, 23% of the spend — against S108's 60%, and the difference is the stage-0 probe.

The key-usage cross-check does NOT close, and the residual is recorded rather than explained. Opening snapshot 60.091117829, closing 61.021052140, delta 0.929934311 against a per-request sum of 0.860694702 — residual +$0.069239609, 8.0%. Stable across two reads eight minutes apart (note (bil)'s lag is not the cause), and 34 raw bodies, 34 metas, zero transport failures, so it is not an unrecorded call. The larger figure is ledgered. Every prior session that reconciled closed to 1e-9; this one does not, and no explanation is offered.

| 2026-08-05 | S110 | stage 0 — the note-(bhf)/(bit) seat probe at the run's own ceiling (P1, P3, P5, max_tokens 12,000) | 0.01 | 0.000456344 | per-response usage.cost | 3 of 3 alive and cheap: 6, 16 and 35 completion tokens. Ran before anything else; no seat later returned a length body or an empty one | | 2026-08-05 | S110 | pre-run adversarial critic (z-ai/glm-5.2; two billed bodies, providers Together and Sail Research) | 0.36 from max_tokens 24,000 | 0.0760597 | per-response usage.cost | NEEDS-AMENDMENT, six findings, four BLOCKING, all six accepted. F4 corrected the design against its author and is what built arm 3. $0.0341885 of this is a second body the runner's format guard rejected and nobody opened until after the run — it held two findings the first did not. Note (biv) | | 2026-08-05 | S110 | the kinship-line run, 3 arms × 3 seats (P1 OpenAI, P3 xAI, P5 Alibaba/DigitalOcean/StreamLake) | 0.82 from max_tokens 12,000 with routing margin | 0.0670119895 | per-response usage.cost, re-summed by analysis/verify.py from the stored raw bodies | 9 of 9 accepted on first dispatch, zero retries, zero length bodies. P5 routed to three different providers across three identical-shaped calls | | 2026-08-05 | S110 | ¶376–572 collated against the 1886 first edition, span 6 translated whole, the whole-novella herra census, the second-person grid, and verification | 0.00 | 0.000000 | — | No API call. Lead translation is free and is never ledgered (charter §3, A4) |

S110 total, summed per-request over all 14 stored billed bodies: $0.143528033, against a declared worst case of $1.40 — 10%. The key-usage cross-check CLOSES: opening 61.02627934, closing 61.169807373, delta 0.143528033, residual −$0.000000001.

Note (bit) paid again, and note (biv) is new and cost $0.0341885. The probe cost $0.000456 and every one of the nine seat dispatches came back clean on the first attempt — the second session running with zero seat-failure waste. What was wasted was a critic body: the runner's answer-line guard required a bare VERDICT, the critic wrote **VERDICT:**, and a valid review was declared a seat failure and re-dispatched. A guard rejects a format, never a content.

| 2026-08-05 | S111 | stage 0 — the note-(bhf)/(bit) seat probe at the run's own ceiling (P1–P5, max_tokens 12,000) | 0.01 | 0.002097622 | per-response usage.cost | 5 of 5 alive, 3 chars each. It proved liveness and NOT appetite — P2 later died twice on the real payload at the same ceiling. Note (bhf) fires a sixth time | | 2026-08-05 | S111 | pre-run adversarial critic (z-ai/glm-5.2, one body) | 0.30 from max_tokens 24,000 | 0.02569214 | per-response usage.cost | NEEDS-REDESIGN, nine findings, two BLOCKING, all nine accepted. Finding 1 changed the admission gate's selector before dispatch; finding 2 replaced a control that could not fail with the arm that produced the run's headline | | 2026-08-05 | S111 | stage A — yardstick A1/A2, NEUTRAL, PARA, FORCEDM (5 calls) | 0.30 | 0.123283845 | per-response usage.cost | 5 of 5 accepted on first dispatch, 10 of 10 items each | | 2026-08-05 | S111 | stage A pass 2 — amendment A8, the quotation re-request (A1, A2) | 0.10 | 0.023514485 | per-response usage.cost | Uniform all-or-nothing re-request; the quotation leak went to 0 of 10 and two sites left every denominator | | 2026-08-05 | S111 | amendment A6 — WRONG plausibility rating (P4) | 0.08 | 0.080799 | per-response usage.cost | Mean 2.12 of 5, with two spans at 4. The control is not a straw man at the two sites where it mattered | | 2026-08-05 | S111 | stage B — grading, 3 seats × 2 orderings + 2 dead P2 bodies + the A9 re-dispatch | 0.72 from max_tokens 12,000 | 0.443922926 | per-response usage.cost, re-summed by analysis/verify.py from the stored raw bodies | 6 accepted bodies, 56 of 56 judgements each. $0.174 of this is two P2 bodies that returned finish_reason: length with zero content; amendment A9 raised that seat's ceiling to 24,000 and it returned in 36 s | | 2026-08-05 | S111 | «Kamizelka» translated whole twice (R06 + R04), the census, FORCED, REAIM, POSITIVE, WRONG, the coding and all verification | 0.00 | 0.000000 | — | No API call. Lead translation is free and is never ledgered (charter §3, A4) |

S111 total, re-summed per-request over all 20 stored billed bodies: $0.699309918, against a declared worst case of $1.55 — 45%. The key-usage cross-check CLOSES to −0.000000000 — opening 61.169807373, closing 61.869117290, delta 0.699309917 against a per-request sum of 0.699309918.

S113 — 2026-08-05 (UTC), E-20260805e-naturalness-wording (ARM-tierP step 3, arm closed)

Pre-flight, written before the first dispatch. 24 payloads × 3 jurors + a blind anchor stage + a pre-run critic; worst case built from max_tokens (note (abc)), per-stage reservation checked before each stage is entered. Declared worst case $2.747 against $2.927599105 of headroom on the day.