Repository path: workshop/experiments/E-20260810z-idiom-reach/design.md · rendered 2026-09-09

E-20260810z — an unexercised permission and an absent device: telling them apart

Frozen 2026-08-10 (S155) before dispatch. ARM-idiom-reach step 1. The translation limb (T-botchan-R24-v1, its W5+ table and R24) was committed at 96ed0aa before this file existed; code.py was written and self-tested before it too; this file was committed at 95db9a2 before the pre-run critic saw it.

AMENDED after the pre-run critic, before any generation call. NEEDS-REDESIGN, 16 findings, 7 BLOCKING; 12 accepted, 4 accepted-in-part, 5 individual remedies overruled in writing. Amendments A1–A11 are in force and are marked in place below. The adjudication is critic.md; the pre-amendment text is git show 95db9a2:.../design.md.

What the critic changed, in one paragraph. The design compared a newly generated source-first +I arm against a −I arm generated at S150 in a different call, which conflates the permission with independent-pass variability (BLOCKING 1, 2, 11): a contemporaneous source-first −I arm Nsrc is added and the primary moves inside the resulting 2 × 2 (A1). The positive control was an arm instructed to maximise the measured outcome, which validates nothing (BLOCKING 5, 6, 14): it is replaced by the lead's own frozen R22 and R24 English at each site, a documented negative/positive pair (A4, A5). A loc answer was counted without checking the named marker was in the text (13): it now must be (A6).

1. Question

framework/v0.2 §7.2 refuses a register-carriage recommendation for the third time, and RS-20260810c §6 names the one thing that would move it:

is the located idiom unavailable in Japanese narration, or is it merely unreachable by minimal revision? The instrument that would separate them writes the +I arm from the source, not from the ∅, and pays the tie-rate cost that the minimal-pair procedure was adopted to avoid.

This is that instrument.

When two independent hands are given the located-idiom permission and write from the Japanese rather than revising their own placeless English, does the English they produce become placeable — and does it become more placeable than the same permission buys inside a revision?

(The second half is the critic's A1: the question is only answerable as an interaction, against a source-first −I arm dispatched in the same stage.)

2. What is carried from E-20260810c unchanged, and why that is the whole design

Everything except the one clause under test. The sites, the hands, the judges, the loc prompt, the item-batching rule and the respelling protocol are the previous run's, and three of the six model arms rated here are the previous run's frozen output, re-rated in this batch.

The one change is the procedure by which the +I arm is written. In E-20260810c the A cell was a minimal revision of the hand's own ∅. Here Asrc is the same clause, the same model, the same temperature and the same item format, written from the Japanese with no English in front of it.

3. What this run does NOT attempt, registered so a reader does not look for it

There is no REG stage and no register comparison. P1a — how much of English's downward movement each device buys — is not asked here and is not withheld here; it is simply not this run's question. §7.2's obstacle was the manipulation, and until a +I arm exists that actually differs from its baseline there is nothing to compare. This run is a reach run. If it passes, the register comparison is the successor and it needs a REG panel; if it fails, the register comparison has no arm to run on.

No new census, no new sites, no new criterion. Reusing a frozen list removes a degree of freedom the lead would otherwise have after having translated this span twice.

4. Arms

Six model arms per hand, plus two lead instrument controls — fourteen English renderings per site.

A1 — the design is a 2 × 2 and the primary lives inside it.

Nsrc and Asrc are dispatched in the same stage, to the same models, at the same temperature, over the same items, in the same format, and their prompt strings differ in exactly one clause — machine-checked before dispatch: the unified diff of the two is two lines and both are the clause. Nothing else about them differs, which is what loc(Asrc) − loc(Nsrc) therefore isolates.

A4 — NLEAD and PLEAD are instrument controls and enter G3b and nothing else. They are one rendering per site with no hand, drawn from two frozen lead translations of this exact span: NLEAD written under a rule that forbade located means, PLEAD under one that permitted them, with a per-site log naming the located item taken at 45 unprimed sites. The paragraph-to-site alignment is the lead's, made once, and code.assert_lead_sites() proves every span is a contiguous substring of the frozen translation — the boundaries are the lead's choice and not one word is. Three of the thirty spans are byte-identical across the pair (JA-11, JA-14, JA-27), which bounds G3b at 0.90 rather than 1.00 and is stated here rather than discovered.

Nsrc's prompt is E-20260810c's BASE + CLAUSE["null"] + TAIL, and Asrc's is the same with CLAUSE["A"], both with the REVISE block absent. The REVISE block is what makes A a revision of ∅; its absence is what makes Asrc a translation.

Lsrc's prompt is Asrc's with one sentence appended to the clause, printed here in full so the dose is on the record:

At every site where such an option exists, PREFER it: where English offers both a placeless wording and one a reader would place, choose the one a reader would place.

Lsrc is an instrument dose, not a translation regime, and no artifact is filed for it. Its role is narrowed by A6: it measures prompt-following reach — how far a hand told to prefer placeable English moves the placement rate — and it is not offered as a measurement of what the material affords. The critic's BLOCKING 5 is right that an arm instructed to maximise the measured outcome cannot validate the instrument, and A4's PLEAD/NLEAD pair does that job instead.

5. Roles, and no model does double duty

The lead is not in the run. T-botchan-R24-v1 is rated by nobody, enters no figure, and is reported only as a translator's-chair record beside the panel's (§9). The lead wrote it knowing the arm's question, which disqualifies it from every primary more strongly than any overlap measurement could.

6. The measure

loc, exactly as E-20260810c §7 defines it: for each ⟨site, hand, arm⟩ a judge returns a string naming the marker that would make a reader place the English, or the empty string. A non-empty string is a placement.

loc(arm, judge) = share of the 60 ⟨site, hand⟩ pairs at which that judge returned a non-empty loc string for that arm.

Denominator 60 = 30 sites × 2 hands. Reported per judge and, in the per-hand table, per hand. Absolute levels are never pooled across judges; every difference is read within one judge.

7. Gates

Failure criteria, and what each one does

• F1 — G3 fails on either judge → P1, P1b, P2 and P3 are all withheld. The batch's loc is not the instrument S145 and S150 measured with.
• F2 — ~~G2~~ deleted (A3).
• F3 — G1 fails on a hand-arm → the failure is reported and the primary is withheld, not recomputed on the surviving hand. A7, from the critic's finding 9: a denominator that moves after an arm-specific failure is observed can only make a result look stronger.
• F4 — G3b fails → P1's NULL reading is withheld and its positive reading is not. Asymmetric, and registered here rather than chosen later: an instrument that cannot see the device makes a null uninterpretable, while a positive result under a blind instrument is conservative. If P1 passes and G3b fails, P1 stands and the result says the instrument control failed.
• F5 — G4 fails after one re-dispatch → P1 is withheld and the run reports the returned cells with the shortfall and its distribution across arms, judges and sites named.

No bar here is loosened after it fires and none may be moved after any body is read. G3's 0.40 sits below both prior measurements of the same quantity on the same judges; P1's +0.20 is E-20260810c's G2 bar unchanged.

8. Predictions

Every rate is reported twice (A6): over all non-empty loc calls, and over TEXT-GROUNDED calls only — a marker counts as a placement only if code.grounded() finds it in the English the judge was shown. Where the two readings differ in sign, the grounded one governs and the result says so.

Supporting inference, registered now (A8). A two-sided exact site-level sign-flip permutation over the 30 site-level differences (Asrc minus Nsrc placements, summed over the two hands), per judge, α = 0.05. The site is the analysis unit; ⟨site, hand⟩ is the reporting unit. It is support, not a gate. The "both judges" rule is a conjunctive requirement, which is conservative, and is not offered as a multiplicity correction.

The primary is computed on the complete-case site set (A7) — sites at which every primary arm returned from every hand and both judges rated every one — fixed before the data are read. The add-excluded sensitivity of A2 is reported beside it.

What each outcome means, written before the data

All three are informative and the design is indifferent between them, which is the condition R24 was written under and is repeated here.

9. The lead's limb, and what it is for

T-botchan-R24-v1 answers the availability question from inside a translator, on the same span, under the same switch, and its W5+ table is a per-site record of TAKEN / AVAILABLE NOT TAKEN / NONE AVAILABLE. Its counts are computed by code.log_counts() and asserted by verify.py; none is typed by hand (note (bey)).

It is not evidence about the panel's behaviour and not a control for anything. It is quoted beside P2 because a translator saying a located option existed at this site and here it is and a model straining for one are two readings of the same question, and where they disagree that is worth seeing. The lead's two primed rows are excluded from every rate the log yields.

Contamination, measured after the freeze (materials/contamination.json, tools/dependence_check.py):

Two things are recorded here before any result is read. The +I rendering sits far closer to the published translation than the −I rendering of the same span by the same translator did — and the 14-token run is "i had studied for three years but to tell the truth i had no", which contains no located idiom at all. And the lead matches itself at 24 contiguous tokens across two sessions with the earlier rendering unopened, which is note (bhb) again and is why T-botchan-R24-v1 is not an independent second opinion of anything. Neither figure gates this run, because the lead's artifact enters no figure in it.

10. Procedure

1. python3 code.py — self-tests pass. Done before this file was frozen.
2. Pre-run critic, openai/gpt-5.6-terra, over this design, R24, code.py and the frozen W5+ table. Findings adjudicated in critic.md; every accepted amendment applied before any generation call; every overrule written down with its reason.
3. Generation — 6 calls: 2 hands × {Nsrc, Asrc, Lsrc}, temperature 0.0, cap 6,000, one call per ⟨hand, arm⟩ over all 30 sites, JSON array out.
4. Mechanical checks G1 and G2 computed from code.py before any rating call goes out; code.assert_lead_sites() must pass or nothing is dispatched. If F3 fires it fires here.
5. Rating — 4 blocks × 2 judges = 8 calls (A9); 14 arms per site (6 model arms × 2 hands, plus NLEAD and PLEAD), 420 items, ≤ 112 per call; item order shuffled across the whole block with seed 8155; all arms of a site in the same call (A10, and see §12.7).
6. analyse.py → analysis/report.json. verify.py recomputes every reported number from the raw bodies and the frozen inputs, and asserts the frozen-input hashes.

Every raw body is written to runs/ before anything is computed from it; a dead body is rotated into runs/discarded/ and never overwritten (note (bhd)).

11. Pre-flight cost estimate — built from max_tokens, not from expected output (note (abc))

Restated with arithmetic after A11 (critic finding 16), and after the critic itself was paid for at $0.02704075 actual.

UTC day 2026-08-10 stood at $2.236650490 of $5.00 across seven sessions; headroom $2.763349510, and $2.736308760 after the critic. The ceiling fits with $0.99 to spare. A run that does not fit is split, scaled down or deferred; nothing here needs to be.

12. Declared deviations and known defects of this design

1. Two hands, both LLMs, one work, one author, one language pair, thirty sentences — S150's limit 2, carried whole, including that the estimand is criterion-positive narration sentences of this span.
2. P2/G5 and P3/G3 are each one quantity in two roles. Declared here rather than discovered later. G3 is defensible because B is not one of the arms under comparison; G5 is not independent of P2, and F4 is written to make that harmless rather than to hide it.
3. Lsrc is a dose, not a regime. It is not a serious rendering policy and no claim about good translation rests on it. W8 (not parodic) still binds through BASE.
4. One post-freeze edit to T-botchan-R24-v1, made before this design existed and recorded here: row A62's item cell named the placeless wording before the located one, which would have put a placeless phrase into the frozen lexicon. The cell was reordered to name the located item first. No translation decision, code, count or coverage code changed; git diff 96ed0aa is the check.
5. loc is a judgement about English by two models whose competence is asserted by nobody (A10). Tier D is NOT PASSED and every figure here is panel-perceived and provisional.
6. The frozen ∅, A and B arms were generated in a different call from the one that rates them here. Their re-rating is a fresh measurement of frozen text, which is what makes P1b and P3 reproductions; it is not a re-run of S150 and cannot detect a generation-side defect.
7. There is no independent manipulation gate on P1, and A3 says so rather than hiding it. Byte divergence between two independently generated arms is guaranteed and gates nothing; a blinded item-level coding of whether a located option was taken is a third panel stage the budget does not hold. The mechanical corroborators are P4 (frozen lexicon, jury-free) and A6's grounding requirement; neither is a gate.
8. All fourteen arms of a site are rated in one call (A10, critic BLOCKING 7). Arm labels are never shown and item order is shuffled across the whole block, but a judge could in principle read the arms comparatively. The batching is RS-20260806g's rule and is why differences are read within a call at all. The empirical answer is that this exact batching returned loc(A) − loc(∅) = −0.0167 / 0.0000 at S150 — if comparative reading inflates differences, it inflated nothing there. It remains this run's largest instrument limitation.
9. NLEAD and PLEAD are written by a translator who knew the arm's question, and PLEAD's contamination against Morri 1918 is DEPENDENT? (§9). Neither matters for their one role: G3b asks only whether two judges can see a located idiom that a frozen log documents, in text with no respelling in it. They enter no comparison with any hand and no primary.