Repository path: wiki/archive/method-notes-S032-S243.md · rendered 2026-09-09

Standing method notes

What this is for. These are the project's accumulated procedural lessons, each one paid for by a defect. They lived as a single unbroken paragraph at the bottom of NEXT.md, re-typed every session, and had grown to roughly fifty items — carried in full every session, unsorted, with no way to tell a live note from a dead one.

The wall was already losing items. Note (ii) — a self-test fixture must be shaped like the target material — was created in S025, fired in S026 and again in S027 ("third session running"), and had vanished from the list by S031. Nothing recorded its removal. It is restored below. That is the argument for this page: an append-only paragraph re-typed by hand loses things silently, and the loss is invisible precisely because the thing that would notice is the list itself.

Ids are permanent. The letter ids are cited verbatim in frozen result and design pages, so they are never renumbered, reused, or tidied. A discharged note keeps its id and is recorded as discharged — (n) is the example.

One id was issued twice, and it stays that way. (bha) names two notes: the audit-is-a-write note (S080, §Live below) and the front-matter/senses note (S081). Both are cited by that id in frozen pages, so renumbering either would break a citation, which is the one thing this rule exists to prevent. Decided at S084 rather than tidied: a citation of (bha) is resolved by its subject, not by its number — audits and execution-in-place resolve to the S080 note, senses: front matter to the S081 note — and both entries carry a pointer to the other. No third note may take the id.

Compaction rule

• A note fires when a session's work is changed by it. Record the session in last fired.
• A note discharged by a permanent structural fix (a checklist, a load-time refusal, a tool gate) moves to §Discharged with what discharged it. It is not deleted.
• NEXT.md does not reproduce this list. It cites at most the two or three notes bearing on the next unit, by id.

Design and pre-registration

| (bdm) | When a category label will not reproduce between readers, test whether it is asking two questions before concluding the phenomenon is fuzzy. RS-20260729 found the drift window's four-class scheme reproducing at 0.654 and read it as evidence that "no overlap at all" is not a line two readers can share. It is partly that and partly something cheaper: the scheme runs semantic distance and register down one axis, so duguð → doughty drew total from one reader ("does not mean retainers") and marked from the other ("still means valiant but is archaic"), and both were right about different questions. Splitting the label into two graded axes at S054 raised three-rater reliability from α = 0.51 (nominal, four-class) to α = 0.78 (semantic) and α = 0.89 (currency). The diagnostic is one question — "could a reader answer this label two ways and be right both times?" — and it costs nothing to ask at design time. | S054 | S054 | | (bds) | Two failure criteria whose antecedents overlap will both fire, and one of them will rest on a measurement the other has disabled. E-20260729d registered F3 (does this candidate enter the inventory?) and F4 (is the project's standing zero a property of translation decisions?) with antecedents sharing a term, plus F1, a reliability floor over the pass F4's antecedent counts come from. All three fired. The adjudication is available — F3's antecedent needs only an occurrence plus a second, reliable pass; F4's needs a count from the failed pass — but it is an adjudication made after seeing the numbers, which is exactly what pre-registration exists to avoid. Registering criteria is not enough: their antecedents must be checked for overlap, and any criterion resting on a quantity another criterion can void must say in advance what happens when it is voided. | S056 | S056 | | (bew) | When one agent writes a coding rule, applies it and reports the statistic, the rule's definition and its application come apart silently — and the cheapest check is to re-derive its clearest cases from the definition's own text. R07 §5 defines D as a rule names the feature and only ONE live option satisfies it. Across three runs a D was in fact recorded whenever a rule excluded an option. Three sites were forced into E-20260730e as a positive control because the lead judged them its clearest cases; at two of the three, F4 excludes one option and two English renderings remain, so the correct code is P. Three independent readers returned P there. The control failed because the control was wrong, and the readers were right. The defect is invisible to any amount of care in the individual judgments, because each one felt decided. Rule: for any coding scheme the project both defines and applies, take three cases the coder is most confident of and re-derive them from the definition's wording before reporting a rate over them. | S064 | S064 | | (bey) | Read a long table through a truncated view and you get a wrong denominator; parse tables, do not read them. E-20260730e's pre-registration quoted T-osso-di-morto-R07-v1 at 21 logged sites; it has 44, and its own tally section says so correctly. The first 21 rows were taken for the whole table. The same parse then found T-postmaster-R07-v1's stated split (P 8 / S 2) wrong at P 7 / S 3. With T-osso-di-morto's own note that its hand tally of 7/28/9 was wrong, three of five R07 tally lines were wrong on first writing and every one was caught by parsing rather than rereading — and every error was in the P/S split, never in a D count. Rule: any count over a table in this repository is computed by a script that reads the file, and the script's row count is asserted against the table's own stated N. | S064 | S064 | | (bib) | "Tier D does not bind this run, because nothing here judges quality" is a true premise that does not reach an operational rule change — and a design that leans on it will have its motion refused. D-20260803-15 proposed amending naturalness on descriptive markedness coding by an uncalibrated panel, arguing correctly that Tier D governs quality verdicts and this run made none. The ratifying vote accepted the premise and refused the motion: the proposed clause would govern how future evaluations score markedness, which is operational and consequential whether or not it is itself a quality claim. Two further grounds it named, both usable in advance: a difference between two prompted rating procedures is not a demonstrated property of the target language or of readers — say which you measured, in the sentence, every time — and an effect whose magnitude nearly doubles with item order supports no threshold-based decision, however clean the pre-registration. Rule: any motion to change a sense's operational wording on model-rater evidence must pre-register validation capable of distinguishing model prompt-following from the property claimed, and must pre-register what order sensitivity would void the threshold. Registering the threshold is not enough if the estimate is not stable under the design's own robustness arm. | S099 | S099 | | (bhr) | [FIRED S162, on an amendment the lead adopted from its own critic. E-20260811h accepted A10 — a minimum of 24 decided judgments before any equivalence verdict — and in the same session cut three control contrasts to one order per passage, giving them a maximum of 24 judgments. Two equivalence verdicts were therefore unreachable by construction and returned INCONCLUSIVE at 21 and 17. The arithmetic needed no seat, no call and no result. An amendment is a failure criterion too, and adopting one does not exempt it from this note. RS-20260811h §5.] Every failure criterion computable from the frozen materials must be computed BEFORE dispatch, not at analysis time. E-20260802f registered FC3, a matching check on two arms the lead had already written — and the pre-run critic computed it and found it failed at 38.3% against a 30% ceiling, which would have withheld the run's second primary before a single datum existed. Nothing about that check needed a seat, a call or a result; it needed arithmetic over two frozen strings. A criterion that can only fire after the money is spent is a criterion the design chose not to test. The rule: at freeze time, partition the failure criteria into those that depend on returned data and those that do not, run the second set, and put their values in the design. This is not the same as note (abc) (estimate from the cap) or note (bfc) (declare a reserve); it is about the design's own logic**, not its cost. | S092 | S092 |

| (bio) | A mechanical check on a frozen specification catches literal strings and nothing else, so the pre-run critic must be asked for the paraphrase BY NAME — and when it finds one, the specification does not get rewritten, the CLAIM does. R11 requires that a purpose specification not be written in the vocabulary of wiki/goodness-senses.md, and records that at S084 a paraphrase got through the check and would have manufactured a trade at every site. S107 repeated the failure with the check passing. materials/check_specs.py cleared both frozen specs against 35 strings; the critic, asked for the paraphrase by name in the design's own brief, found each spec restating one half of the definition the run existed to test. By then both arms were rendered against those specs. The repair is not to rewrite a freeze — that is worse than the defect — but to re-register, before any measurement call, what the run is allowed to conclude: S107 downgraded its primary from the halves are separable to the halves are separable by a translator pointed at these two readers, promoted the null to the strong outcome, and raised its own motion bar. Rule: on any run whose specification could contain its own answer, (i) write the paraphrase question into the critic brief, (ii) run the critic BEFORE the material is produced if the specification governs production, and (iii) if it fires late, amend the claim and not the freeze. Companion to (rr) and (big). | S107 | S107 | | (biq) | A blind is not a property of the materials; it is a property of the materials AT A GRAIN, and a design that reports "the seats recognised the translators" without naming the unit has not measured what it thinks. RS-20260804c put whole loci of «Певцы» to three seats and all three named Garnett and all three named Hapgood, and the run concluded — in the strongest sentence in it — that "a documented human ranking of two published translations of this period cannot be tested blind on this instrument, because the instrument is not blind." E-20260804i put the same story, the same pair, the same three seats one to three sentences at a time, gave them a false attribution and asked them to name the translators independently of it: P1 followed the false label at 24 of 24 cells and the true text at 0. The blind that failed at locus level holds at site level, and an arm step written to go and find new materials did not need to. Rule: every recognition probe reports the unit it was run on, and no blind claim generalises across grains without being re-measured at the grain the design actually uses. The corollary is cheap and worth having: when a blind fails, shrink the unit before abandoning the materials. Companion to (bfq) and (bdq) — reachability is a property of the reader — with the addition that it is a property of the reader and the window. | S108 | S108 | | (bir) | An amendment that changes a design's shape silently invalidates the failure criteria written for the old shape, and the pre-run critic will not catch it because the criteria were clean when it read them. E-20260804i registered FC2, a degeneracy check excluding any seat that returns the identical ranking at ≥ 11 of 12 sites. It was written for a four-arm stage. The same critic pass that approved it produced BLOCKING finding F4, which abolished the four-arm stage in favour of two-arm blocks — and with two arms, "the same arm first at 11 of 12 sites" is not degeneracy but exactly what a real dimensional difference looks like. Applied literally after the run, FC2 would have deleted the whole of stage 2's accuracy data because both its seats were consistent. The run reported both readings and declared the defect rather than quietly dropping the criterion, which is the only honest move left once the numbers exist. Rule: after accepting an amendment that changes arms, stages, or units, re-read every registered criterion against the NEW shape and re-freeze it before dispatch — an amendment is a redesign, and a criterion inherited across it is inherited untested. Sibling of (bds), which is the same failure between two criteria rather than between a criterion and an amendment. | S108 | S108 | | (bip) | A no-effect control site must be chosen so that the artefact could show up in it — a control whose ceiling is below the effect's floor cannot fail. E-20260804h registered a localization control on two no-fork sites, and the pre-run critic showed arithmetically that it could not fail: one of the two was twenty words long and nearly unflattenable, capping the no-fork rate near 0.50 while the fork rate could reach 1.00, so the control would have passed with the artefact present. Four sites were added, giving six and six, and the criterion was restated as a margin rather than a comparison. It then returned 0.778 against 0.278 and the added sites were where the run was most informative — they split in both directions, which is what showed the translator's hand was not the explanation. Rule: before registering a control-vs-treatment comparison, compute the control arm's attainable range under the artefact you fear; if the artefact cannot move it past the bar, the control is decoration. Companion to (rr): a critic that does this arithmetic is worth more than one that reads the argument. | S107 | S107 | | (bin) | Byte-identity is not neutrality: a repair that holds material fixed in order to vary one thing must first read what it is holding fixed. E-20260804g was designed to re-run six sites with only the yardstick changed, and proposed to hand the yardstick seat the same frozen literal glosses the graders get. The pre-run critic went and read them: five of the six state the relation outright — little-dear-earth, your-servant, daughters rather than sons ... feminine, you[plural] — and at one site the literal gloss and the filed rendering are the same English sentence, so the design would have shown its "independent" seat the very translation it promised on its own face that seat had never seen. Rule: when a design's argument is we changed only X, enumerate what the held-fixed material actually contains before dispatch, and say which of it could produce the result on its own. The held-fixed inputs of the new role are not covered by the freeze that protects the old comparison, and inheriting them unread is how a control leaks. Companion to (rr): the critic is the mechanism that caught it, and it caught it by reading the materials rather than the design. | S106 | S106 | | (bll) | A manipulation that produces a MIXTURE of two arms is tested against a composition-matched arm, never against an interpolation between them. E-20260809i registered its primary as a dip below the linear interpolation of an all-domesticated and an all-foreign arm at the drifting arm's own foreign-token proportion. The pre-run critic's BLOCKING 1 showed the null is unearned: jury scores are bounded ordinal judgements, and any convex response to that proportion produces the registered dip with no drift effect present at all — and can do it on one sense and not another, so the cross-sense interaction the design called immune is not immune either. The remedy was cheap and had to be built rather than argued: an item-uniform arm carrying exactly the same number of foreign tokens in every block (12/16, 13/19, 14/19), drawn from as nearly as possible the same items, so the contrast is a direct one between two texts of equal composition. Rule: where an arm is a mixture, build the matched arm; a chord residual is an inference about a curve nobody has measured. The general form: an additive or linear null asserted over a scale whose shape has never been measured is an assumption presented as a control. RS-20260809i-terminology-drift §6. | S147 | S147 | | (blw) | A registered count is untestable unless it says how a case with two causes is counted, and a rate computed from a list drawn after the data are open is selection on the outcome. RS-20260810x registered P-B — sites where a rule bound after S132 makes the new English differ — inside a frozen translation log, before the comparator was opened, which is the hard part of pre-registration and was done right. It then adopted the precedent's exclusive-cause precedence after the run, under which the count is 5 and the 6–15 band fails; under P-B's literal wording it is 9 and the band passes. Both readings are legitimate and the prediction cannot be scored, so it is recorded INVALID AS REGISTERED rather than as a pass or a fail — choosing between them after seeing the answer is the thing pre-registration exists to prevent. The same page's first draft reported 23 of 24 named decision sites agree; the independent critic's BLOCKING 5 showed the 24 were named after both texts were open and were memorable because they matched. Replaced by a frame fixed inside the freeze — the 22 decisions the translator's own log names — on which the figure is 12 agree, 10 differ. Two rules. (i) Any registered count states its counting rule, including how an item with more than one cause is assigned, before dispatch. (ii) Any agreement or convergence rate is computed over a frame that existed before the comparison did, and a hand-picked list of matches is published as examples with no denominator. Sibling of (bkm), which is the same failure on a null; companion to (bkc), which is about what a count of divergences can mean at all. RS-20260810x-legend-again §6, §7, §12. | S154 | S154 | | (bnj) | In a census of what published translators did with a source device, a locus that is not in a translator's own source text is not a locus he dropped — code it separately, never fold it into the loss count, and where the copy-text is not the text the hands worked from, say so on the verdict. E-20260814 censused eight rhymed-prose loci over Lane 1839 and Burton 1885. Locus 8 turned out to be absent from both, and the reason is neither translator's: both of their Arabic texts carry a vizier-and-letter episode that the copy-text compresses into eleven words, so the join the locus names does not exist in either. Lane abridged and Burton expanded, but neither can explain material present in both of them and missing from the copy-text. Had absent been scored as lost, Lane's figure would have read 8 of 8 lost and Burton's 5 of 8 — both wrong, and wrong in the direction that flatters the finding. The design registered the four-value scale and a power floor (≥ 6 of 8 loci locatable in both hands) before either text was opened, which is why the ambiguity is reportable rather than fatal. The rule generalises past unstable recensions to any census over hands whose source texts are not demonstrably identical to the copy-text — which is most of them, and had not come up before only because the project's first eighteen source languages all had single authorial texts. RS-20260814-saj-carriage §6a, limit 3. | S180 | S180 | | (bnl) | An effort pin is a request, not a control: disable reasoning outright where a seat judges nothing, and where it judges, size the cap from MEASURED reasoning tokens before dispatching the stage. E-20260813f proved both halves in one session. Half one: reasoning: {"effort": "low"} was ignored by qwen on Alibaba (note (bnk)'s second firing), while reasoning: {"enabled": false} on the same seat and prompt returned a complete body with 0 reasoning tokens for $0.001081 — a 40× cost fall and the failure mode gone. z-ai/glm-5.2 behaved identically; google/gemini-3.6-flash rejects enabled: false with HTTP 400 and needs {"effort": "minimal"}, so the disable idiom is per-provider and must be probed, not assumed. Half two: the design's §8 caps of 500–700 were built from assumed answer length, and one probe per stage measured what the judges actually spend before answering — P1 0–403 reasoning tokens, P3 326–430, P2 up to 1,043 on one stage. Every judge stage would have truncated. Fires at: every stage of every run on a seat that reasons, which on this panel is all of them. Remedy, and the line it draws: disable reasoning on seats that write documents or code gates — they judge nothing and the project already wanted minimal effort there — and never on a seat whose output is a judgment, because whether a judge thinks before answering is part of the instrument; for those, spend one probe per stage on the real prompt and set the cap from the number it returns. RS-20260814c-plain-yardstick §9 and E-20260813f §12 A6. | S182 | S182 | | (bnm) | A prompt that embeds model-generated text cannot be asserted free of a word that the generating prompt requires — assert the ban on the scaffolding, measure the contamination in the block, and never claim a blind the design cannot hold. E-20260813f §10.3 asserted that no GS prompt contains original, Japanese, source, effect, comparable or translation — the whole point of GS being a target-only style question. But a GS prompt embeds a YP document, and YP's own instruction two stages earlier requires the word: "what the passage does to a reader of Japanese". The assertion was unsatisfiable the day it was frozen. Measured: 24 of 90 GS prompts carry a banned word inside the document block, across 5 segments, so in a quarter of its cells the seat could see that the two passages are translations from Japanese. Fires at: every design with a blindness or leakage assertion over a prompt containing generated material — every yardstick design this project has run. Remedy: split the assertion in two. The scaffolding is fixed text and the ban holds over it absolutely; the generated block is data, and the verifier counts the leak and the result page reports it as a limit. And read the generating prompt against the ban at freeze time — here the two prompts sat four pages apart in one document and contradicted each other. RS-20260814c-plain-yardstick §9. | S182 | S182 | | (bnn) | A dispatcher's usability test must require the reply to PARSE and carry its stage's key; "non-empty content" passes bodies that answered a different question. E-20260813f's runner implemented F6's truncation guard as empty, or unbalanced braces. yp__S14 returned 1,200 tokens of bare prose with no JSON anywhere and finish_reason: "stop" — a parse failure F6 explicitly covers — and the test called it usable. It was caught only because a downstream count came back 14 where 15 was expected. Fires at: every runner in this project, all of which key deadness off content being non-empty. Remedy: one table from stage to expected key, checked inside the dispatch loop so the re-dispatch ladder actually runs; and mark the stored record usable: false so the analysis and the F5 count key off the same fact rather than re-deriving it. RS-20260814c-plain-yardstick §9. | S182 | S182 | | (bnk) | On a reasoning seat, max_tokens caps the answer PLUS the hidden thinking, and the thinking will take all of it — cap the reasoning, not the total, or budget for buying nothing. E-20260814b's pre-run critic (P4 moonshotai/kimi-k3) was dispatched at max_tokens 4,000 and then 14,000; both returned finish_reason: length with empty content after spending 3,997 and ~14,000 tokens on reasoning, and a third attempt was billed and lost to a network error. $0.526018 — 47.4% of the run's total spend — bought no output token at all. The fourth attempt, identical except for reasoning: {"effort": "low"} and max_tokens 6,000, returned a full NEEDS-AMENDMENT verdict with eight findings in 91 seconds for $0.083293. Fires at: every dispatch to a seat that reasons before answering — on this panel P4 above all — and every pre-flight estimate that prices such a call. Remedy: send reasoning: {"effort": "low"} (or an explicit reasoning cap) on any call whose value is the written answer rather than the depth of the thinking, and treat note (abc)'s worst case as a floor rather than a ceiling on a reasoning seat: (abc) prices the cap, and this note says the cap can be consumed without producing anything. RS-20260814b-honorific-hands §8. | S181 | S221 | | (bod) | "Sentence 24" is not a reference unless the design says whose sentences, and any property a selection rule claims must be ASSERTED BY THE SCRIPT rather than described in prose. E-20260816 numbered the Arabic source's sentences in its frozen figure inventory and the English rendering's sentences in its decoy-selection rule, called both sentence N, and the two sequences collide at 24, 35 and 45. The pre-run critic read the collision as a BLOCKING defect — three decoy loci allegedly containing enumerated source figures — and it was wrong on the facts (those decoys render Arabic sentences 19–20, 31 and 41, against figures at 24, 35 and 45) and right that nothing in the design could settle it. A reader who cannot check a claim has to disbelieve it, and a critic who disbelieves the wrong claim spends its findings there instead of on the design. Fires at: every design that indexes two texts — which is every design in this project with a source and a rendering, and the collision has simply not been noticed before. Remedy, two parts. (i) Name the text in every index (Ar-24, En-24), never the bare number. (ii) Where the prose asserts a structural property of the materials — no decoy span overlaps a figure span, the arms differ only here — put it in the cell builder as a check that fails the run, so it is verified rather than believed; E-20260816's cells.py now asserts pairwise locus disjointness and the class multiset. Corollary worth keeping: a critic finding that is factually wrong BECAUSE the design was ambiguous is a finding about the design, and is recorded as accepted-as-defect rather than refused — refusing it teaches the next session nothing. RS-20260816-answering-figure §4; E-20260816 §11 finding 1. | S191 | S191 |

Controls, gates and critique

| (bgy) | The rule an audit validates can be the rule the audit's own repair left broken, and only a ground truth can tell you which. ARM-figure-audit spent three sessions establishing that published figures stand under the repaired name_tokens, and every one of those verdicts is a statement about stability. E-20260801d put 138 ever-capitalised types to three seats as a question of fact and got recall 1.000, precision 0.830 type-level and 0.679 token-weighted: one excluded token in three is not a name, the error is entirely one-sided over-exclusion, and 11 of the 19 false positives come from R1 — the rule S031's defect A repaired — against 1 from R2, which is the opposite of what RS-20260726c reported for its own seven cells. The mechanism is a symmetric blind spot: ends_sentence() strips closing delimiters from the previous token and never looks at an opening delimiter on the current one, so every quotation opened after a comma yields a false name. Rule: an audit that recomputes a figure under a rule has tested the figure, not the rule; when the rule is a heuristic, one ground-truth cell is worth three sessions of recomputation, and it is cheap because the candidate set is finite and enumerable. Corollary, from amendment A3: report a type-level precision and a token-weighted one — the gap was 15 points here, and the type figure alone understates the effect on n-grams by half. | S080 | S080 | | (bgz) | The preserved-bytes guard is defeated by a label collision, and only the ledger can see it. Note (bdt) writes a response's raw bytes to disk before any parse, which is what makes a rejected body inspectable. call.py labels the primary slug's first attempt with the bare tag, so a re-dispatch of the same seat — amendment A6's cap raise on P2 — wrote over the failed body's .raw. Nothing in the runner, the analysis or the verifier noticed: the accepted body is present and correct and every check passes. It surfaced only in the spend reconciliation, as a key-usage residual of $0.064228500 against an overwritten body whose usage.cost was $0.0642285 — exact to the cent, and sayable at all only because the .meta.json had been read before it was clobbered too. Rule: a preservation guard must be keyed on something monotonic, never on a tag that a retry can reuse — and, more generally, the cost ledger is a second, independent detector of missing artifacts and should be reconciled before the run record is called complete. Sibling of (bgu): both are guards that pass while guarding nothing. AMENDED 2026-08-02 (S087) — the S080 repair does not cover the case that actually recurred. S080 labelled every attempt uniquely and fixed the collision between a primary and its reserve inside one invocation. call.py numbers attempts from 1 per process, so a stage re-invoked after an amendment reuses ..._try1. S087 raised a max_tokens cap twice and re-ran two stages, overwriting three billed bodies worth $0.211820146. Nothing detected it: the verifier's cost check sums the bodies that survive, so it reconciles perfectly against a smaller truth. The rule is now: a label must contain something a re-invocation cannot reproduce — a stage-invocation counter persisted to disk, or the declared cap — not merely the attempt number; and the ledger figure must be the sum over bodies plus anything a re-invocation destroyed, stated as two numbers. | S080 | S080 · S087 | | (bha) | Running another session's analysis script in its own directory is a WRITE. The ARM-figure-audit sweep tried to recompute three pre-repair experiments by invoking their analyse.py in place. Two failed on missing arguments and materials; the third succeeded and rewrote E-20260726-period-control/runs/cell.json, a stored output of a session five sessions earlier. It was caught by git status and restored, and the diff turned out to be the documented S031 name-set change, so nothing was lost — which is luck, not method. Rule: an audit reads; it does not execute the thing it audits in the thing's own directory. Reimplement the specific figure from the stored materials in the auditing experiment's own tree, or run the script against a copy. And: git status is part of a verification pass, not part of committing. ⚠ Id collision: (bha) also names the S081 front-matter/senses: note; resolve a citation by its subject (S084). | S080 | S080 | | (bhk) | A forced-choice null control between two identical texts manufactures the slot bias it is then used to police, and its floor measures identity detection rather than resolution. E-20260802c presented three null items — one Japanese, one Japanese held out, one Russian — as the same text in both slots, and required a non-tied preference. All eighteen cells recognised the identity and said so ("both translations are identical, so I arbitrarily prefer A"), scored the two slots the same, and produced a floor of exactly 0.0000 on all six senses. Two consequences, and both bit. (i) A floor of zero is not a resolution measurement: a juror that has recognised two texts as the same has told you nothing about the smallest difference it could register between texts that genuinely differ, and every "above the floor" statement built on it is vacuous — the design's own registered prediction that accuracy would sit below the floor became passable only on an exact zero. (ii) Told to choose anyway, jurors defaulted to slot A (P2 5 of 6, P5 6 of 6), which pushed their registered slot-A rates from 0.643 and 0.714 up to 0.750 and 0.800, outside the pre-registered band, so the null control triggered the exclusion of two jurors' preference data on the strength of cells where preference is arbitrary by construction. Rules: a null control must be a pair the panel cannot separate while believing they differ — a meaning-preserving micro-paraphrase with a frozen edit list, not byte-identical text; and a forced-preference instrument must exempt cells whose two arms are identical, or exclude them from every slot-preference statistic computed from it. RS-20260802c-regime-scoring §§3, 5. | S089 | S089 | · DISCHARGED S094 — E-20260803-a4-set built the meaning-preserving micro-paraphrase null this note demanded; it moved the jury by 0.139 against a 0.233 retest floor, and it is not co-presented, so recognition cannot occur. | (big) | The statement a rendering is scored against is an arm of the experiment, and if the author of the rendering also wrote it the comparison is closed. E-20260802e established R1 by asking three seats whether an English rendering conveys a stated relation — and the lead wrote both the rendering and the statement. Two independent pre-run critic passes at S101 raised this as their first BLOCKING finding, in the same words: the grading question then measures whether the grader can detect the clause the author embedded to match the relation the author wrote. Repaired in place: a seat shown the source, a gloss and the construction and no English of any kind authored all eight statements, and the lead's were discarded. The result reversed. Against an independently authored yardstick, the filed close translation scored 0.810, an uninvolved seat's plain translation 0.810, the forced re-rendering 0.738 — and a negative control stating a different relation sat at 0.071, so the readers were discriminating and the flatness is not instrument failure. Rule: whenever a design scores prose against a written statement of what the source conveys, that statement is authored by a party that has seen no candidate rendering — and any arm frozen against the old statement (a positive control, a gloss) must be re-authored from the new one, which S101 failed to do and paid for through its own F1. Sibling of (bhz): the leak is not always in the text shown, sometimes it is in the question asked. | S101 | S101 |

| (bis) | A positive control can convict a repair the pre-run critic asked for, and this one did. E-20260805's critic (NEEDS-REDESIGN, three BLOCKING) found, correctly, that coding "does the specific sense survive anywhere in the poem" scores kept the word but moved it off the rhyme as a full save — when that displacement is the cost the hypothesis is about. The accepted repair added a fourth label, ELSEWHERE. ELSEWHERE became an absorbing state: 13 of 14 canary positions, 59 of 82 Taylor positions, 38 of 80 Egan positions — and it swallowed two of the four seeded generalisations, firing FC4 at 2 of 4 and withholding every registered prediction in the run. Two rules. (i) Adding a label to a forced-choice code is a change to the instrument and needs its own control run, not the old one — the canary was designed against the three-way code and was the only thing that noticed. (ii) A critic's finding can be right about the measure and its fix wrong about the instrument; accepting a BLOCKING finding does not license accepting its prescribed change unexamined, and a design that accepts one should say which of the two it is accepting. Sibling of (bhr), from the other direction: there a control computed before dispatch saved the run; here a control computed after dispatch convicted the repair. RS-20260805-rhyme-or-flavour §4. | S109 | S109 |

| (biu) | A positive control must be registered as a DIFFERENCE against the null arm. An absolute threshold the null arm also meets is not a control, and it will pass while demonstrating nothing. E-20260805b froze FC2 as "if fewer than 2 of 3 seats in arm 2 separate the pairs, H1 is withheld" — an arm-2-only threshold. Arm 2 separated at 3 of 3 and FC2 passed; arm 1 also separated at 3 of 3, so the loud inserted marker ("Mr Holpainen") moved no letter and no ranking, and the run ended with no demonstrated ceiling and therefore no licence to say what the target language could carry. The criterion could not have failed for the reason it existed, because it never looked at the arm it was controlling for. Rule: write every control criterion as a comparison between arms — arm2 > arm1 by a stated margin — and never as a threshold one arm can satisfy alone. Companion to (bis): there a critic's prescribed repair broke the instrument and a control caught it; here the control was written so that nothing could catch it. RS-20260805b-address-or-action §7.1. | S110 | S110 | | (biv) | A dispatch guard rejects a FORMAT, never a CONTENT. Read every billed body it rejects, before proceeding. E-20260805b's pre-run critic wrote **VERDICT: …** with markdown emphasis; the runner's line_re required a bare line start, counted zero answer lines, declared a seat failure and re-dispatched. Both bodies were valid, full adversarial reviews and both were billed — $0.0418712 and $0.0341885. The first was read straight off disk and its six findings became the run's amendments; the second was never opened until after the nine seats had run, and it contained two findings the first did not, one of which (the scene-content confound) is the reason the run's headline is a null. So the cost of the unread body was not its $0.034 but a design change that arrived after dispatch instead of before it. Rule: when a guard rejects a body, open the body. A guard's verdict is about whether the runner can parse the answer, and this project's guards have now twice (with (bgw)) rejected work that was correct. RS-20260805b-address-or-action §7.2. | S110 | S110 | | (biw) | A leak screen that checks TERMINOLOGY does not check QUOTATION, and quotation is the same leak. E-20260805c inherited E-20260804g's banned-lexicon screen, which fired on 0 of 10 yardstick statements. Five of the ten nevertheless quoted something out of the material: Polish tokens (sprzedano, takich państwa), an English word choice (strapping), and at one site the string "your husband" — the unmarked English rendering of that very site, verbatim. A grader given that statement is given the answer, and the frozen screen could not see it. Rule: any screen protecting a blind judgement must check for quotation as well as for vocabulary — no source token, no quoted string, no naming of a target-language word the arms might use. And when the fix is found after reading the output, make the trigger mechanical and the response uniform: E-20260805c re-requested all ten statements on any single hit, so no statement could be selected for re-request after being read, and it declared that the lead had seen the first set. Companion to (biu): there a control could not fire; here a control fired correctly on the wrong property. RS-20260805c-no-loss-to-repair §9. | S111 | S111 | | (bix) | If every arm of a recovery design preserves the passage's content, the design measures CONSISTENCY WITH the relation, not whether any marking carries it. E-20260805c graded seven arms against an independent statement of what a Polish passage conveys. WRONG — the one arm that asserts different content — was recovered in 4 of 48 judgements, so the instrument plainly discriminates. But PARA, a free paraphrase written by a seat that was told nothing about any relation, any study, or anything to preserve, was recovered at 8 of 8 sites, 0.854 per judgement — level with the filed translation and above both R1 re-renderings. The six content-preserving arms sat between 0.667 and 1.000 and the content-changing arm sat at 0.083. Rule: before reading a recovery rate as evidence that a marking was carried, the design must contain an arm that varies the marking with the content held fixed. Without it, a high rate is a fact about the sentence's subject matter. This is the standing explanation now available for E-20260804's FR→EN null and for E-20260805c's empty gate alike, and it bounds what RS-20260802e's 5-of-6 can be read to have shown. RS-20260805c-no-loss-to-repair §4. | S111 | S111 | | (biy) | An unbriefed arm is not thereby a NULL arm: ordinary machine rewriting has directions of its own, and one of them is towards the knowing ironist. Note (bix) prescribed an arm that varies the thing under test with the content held fixed, and E-20260805d built one — a persona shift, hand-written, against a paraphrase from a seat told only "say this a different way". Signed to the five declared directions the shift was aimed at, the deliberate change of narratorial person moved source-blind readers 0.800 and the aimless rewrite moved them 0.800: identical to three decimals, ratio 1.000 against a withholding bar of 0.75. The paraphrase raised the narrator's rated self-awareness from 2, 2, 7 to 6, 6, 7 without being asked to touch it. So (bix)'s remedy does not by itself rescue a design: an unbriefed control tests whether the measure moves for free, and when it does the primary is withheld, but the control is not a floor and cannot be read as one. Before registering an unbriefed arm, say which direction you expect its default to lie in and whether that direction is the treatment's. RS-20260805d-two-persons §3. Companion to (bix), which it extends rather than replaces. | S112 | S112 | | (bja) | A seat probe bounds a seat's token appetite on the PAYLOAD SHAPE IT PROBED and on no other. Note (bit) bought the probe and E-20260805d ran it: qwen/qwen3.7-max passed at max_tokens 5,000 on the rating payload — one 600-word text, eight integers out — and then returned finish_reason: length with ~30,000 reasoning characters twice on the parity payload at a cap of 8,000, $0.0757 for zero content, 25% of the run's spend. The two payloads differ in kind: rating is judge one text on a fixed scale, parity is enumerate the differences between two texts, and enumeration has no natural stopping point. Probe the shape whose failure would cost the most, not the shape you happen to dispatch first, and where a run has two task shapes, probe both or price the unprobed one at its cap. Note (bhf)'s eighth firing and its sixth distinct slug. RS-20260805d-two-persons §10. | S112 | S113 | | (bjb) | When a design delegates a judgement to a seat, the seat's own count governs the criterion — including when the lead would have counted differently. E-20260805d's F3 withholds the primary if the substantive divergences between two renderings exceed three, and it delegated substantive to the screen seats. One returned NONE; one returned COUNT: 4. Had the lead adjudicated one of the four as a wording difference the count would have been 3 and the criterion would not have fired. The rule decided and the margin was printed, and the alternative — the author of both texts ruling on how many ways they differ — is the thing a delegated criterion exists to prevent. Rule: a delegated judgement is not reopened by the delegator after the answer arrives; if the delegation was wrong, that is a finding about the design and is recorded as one. | S112 | S112 | | (bjm) | A signed scale is not a two-sided scale until you have counted how often its negative half is used — and the count belongs in the design, as a gate, not in the discussion. E-20260806b asked three seats to score five renderings against their Italian on three signed −3…+3 scales, positive = the deformation Berman predicts. The headline came out clean and monotone — Strettell +1.833 on vernacular effacement, Dole +1.500, the 2026 arms lower, the resistancy control −0.167. Then the registered gate F1′ counted the negatives: 4 of 186 codes on explicitness (2.1%) and 9 of 186 on popular speech (4.8%), against 29 of 186 (15.6%) on register. A mean of +1.8 on a scale nobody scores below zero is compatible with a one-sided instrument, so two of the three primaries were withheld and the two best-looking numbers in the run went with them. The one-sidedness is not noise and cannot be argued away as sloppiness: the same seats returned 0.000 test–retest on a duplicated arm, 18 of 18, and 77 of 78 extreme codes quote something actually present in the span they refer to. They are precise, reproducible and textually accurate, and on two scales of three they never looked below zero. Rule: any design using a signed scale registers, before dispatch, the minimum rate of negative-half usage below which directional claims on that scale are withheld — and it does not substitute a control arm's sign for that count, because a control that fails to go negative is ambiguous between a broken scale and a true universal. That ambiguity is what the pre-run critic caught in this run's original F1, which would have withheld the primary exactly when the hypothesis was supported. RS-20260806b-berman-occurrence §4, §7. | S118 | S118 | | (bjo) | A materials diagnostic must be computed the way the analysis will consume the material, or it will be true of a corpus the run never analyses. E-20260805f chose narration-only as its primary analysis on the ground that "dialogue share is higher in the translation in all four cells", and published the sentence in the design and in RS-20260805f §2. Its analysis blocks one work at a time, choosing that work's dominant quotation convention per work; the diagnostic was computed over an arm's works concatenated, with one convention chosen for the pair. Machen's original arm is two books with different printers' conventions, so the concatenated figure came out 13.7 where the per-work figure is 37.5 — and the cell's asymmetry runs the opposite way to the published claim (0.290 translated against 0.375 original). The statistics were unaffected, because the choice the diagnostic justified is defensible either way; what was wrong was a sentence, published twice, about the corpus. Rule: a diagnostic that justifies a design choice is computed by the same code path that will later read the material — same unit, same normalisation, same per-file decisions — and where it cannot be, the divergence is stated in the design rather than discovered by the next run. RS-20260806c-source-grammar §9; E-20260806c design §2.4 and amendment A7. | S119 | S119 | | (bkf) | An "unlicensed strangeness" arm must be checked edit by edit against the source's own morphology and syntax AT THAT SITE, and a preposition swap almost never survives the check. E-20260807f built a factor out of eighteen markedness edits declared to answer to nothing in the French, and the pre-run critic found all six of one kind — odd prepositional government — traceable to the French preposition at the site: saturated of ← saturé de, upon them ← sur eux, unto ← à twice. A third of the factor was source-licensed, which is the one thing it may not be, and the design's flat sentence "none is a calque" was false. The kinds that did survive are the ones the source language cannot form — Germanic noun-compounding and English predicate fronting, neither available in French — and the replacement kind, broken / unresolved construction, survives because the passage completes every construction it starts. Rule: for each edit in an unlicensed-markedness arm, name the source feature at that site and say why it cannot motivate the edit; prefer kinds the source language structurally lacks; treat function-word substitutions as licensed until shown otherwise. Caught before dispatch and cost $0.03. RS-20260807f §1; E-20260807f amendment A3. | S130 | S130 | | (bkg) | Write down an exact test's minimum attainable P before the run, because a permutation over the wrong units cannot reach significance no matter what the data say. E-20260807f §9 registered an exact permutation "relabelling the items' factor assignment"; the first implementation relabelled the four pooled cell means instead, which has C(4,2) = 15 arrangements and a floor of P = 0.0667 — so every effect in the run, including several that were unanimous across 3 of 3 seats and 6 of 6 segments, came back at P = 0.3333 or 0.6667 and would have been reported as null. Recomputed on the six segments as paired units (2⁶ = 64 sign patterns, floor P = 0.03125) the same effects sit at the floor. Rule: state the exact test's permutation units and its floor in the frozen design, and check the floor is below the alpha the design intends to use. The general form: a test whose smallest possible P exceeds your threshold is not a conservative test, it is a broken one. Caught before any number was reported. RS-20260807f §1; analyse.py docstring. | S130 | S130 | | (bkh) | The empty-body-with-all-tokens-on-hidden-reasoning failure can be PROMPT-shaped, not seat-shaped, and when it is, neither standing remedy works. Note (bhf) has fired eighteen times and its two remedies are raise the cap (note (bhq)) and change the seat. E-20260807g dispatched three sibling generation prompts in one batch to one model; IND came back with zero content and finish_reason: length while INDC and DIRF completed cleanly at the same cap — then did exactly the same under a raised cap of 30,000, then did exactly the same under a different model at $0.212442 for nothing, whose siblings again completed. Three failures, two models, one prompt. The distinguishing feature of the failing prompt is that it asked for a mechanical transformation under a conjunction of prohibitions (keep every word grammar allows, change the speech act not at all, add nothing), which is a specification a model can deliberate about without bound. Rule: when a batch of sibling prompts has one that fails (bhf) and the others succeed, stop applying (bhq) and change the PROMPT — loosen or split the conjunction of constraints — or accept a cheaper hand. Cost of learning this the other way: $0.319654560, 39% of the session. RS-20260807g §8; config/budget.md S131 waste row. | S131 | S131 | | (bki) | A between-hand divergence or agreement rate measured on ONE rendering per hand is not a stable quantity, and the instability is larger than the effects such designs go looking for. E-20260808a took two draws from each of two models at temperature 1.0 — the only reason it knows — and the same ten loci, judged by the same three seats, came back 0.900 different in one draw and 0.200 in the other. No class difference anywhere in that run was a third of that. The mechanism is visible in the prose and needs no statistics: an ordinary locus is a one-word synonym choice (earth against soil, hardship against misery) that resamples. Rule: a design whose statistic is how often two hands part must take at least two draws per hand, or state on the result page that its number is a single draw and that its spread is unmeasured. This bites backwards as well as forwards — RS-20260807c's divergence rates and RS-20260806d's 234 sites are single-draw numbers, and ES-20260806 §3.2 now says so. | S132 | S132 | | (bks) | A slug with no failure record in this project is probed on the SMALLEST call of a stage before the batch, and the fallback is named in the frozen design. E-20260808f swapped its gate checker to qwen/qwen3.8-max on a critic's BLOCKING finding — a lab new to the project, so note (bkh-corr)'s check the project's own failure record had nothing to check. The design therefore registered the probe and the fallback in advance: dispatch the 27-item RU call first, and on a dead body take a named model with a record and restore the limitation the swap had deleted. The probe fired on the first call: finish_reason: length, 14,000 reasoning tokens and zero visible content, $0.087184 — for one call rather than five, and the fallback was taken without deliberation because it had been written down. The rule generalises past new slugs: in the same run, a model with a good record in this project (nvidia/nemotron-3-ultra-550b-a55b) died the same way on the Japanese call at a 20,000 cap while returning 59 Italian items on the same cap in 5,322 characters — so the source language, not the batch size, drove the reasoning chain; splitting the call in two produced both halves for $0.0334. Rule: probe-smallest-first, name the fallback in the frozen design, and treat a per-language cap as unproven until that language has returned once. RS-20260808f-placeless §7. | S137 | S137 | | (bkh-corr) | [FIRED S136 NOT HONOURED, $0.305211000; HONOURED S137 — E-20260808f's frozen design named moonshotai/kimi-k3 as not-to-be-dispatched before any call went out, and it was not. See note (bks) for what replaced the missing-record case.: moonshotai/kimi-k3 was chosen as E-20260808e's second generation hand for MODEL-FAMILY diversity, and returned finish_reason: length with zero content twice — the exact failure this note records for this exact slug, three sessions earlier. The note's remedy was never reached because its precondition, check the project's own failure record for a slug before dispatching it, was not written down. It is now: a slug this project has recorded returning empty length bodies on generation prompts is not selected for a generation stage, whatever else recommends it.] Note (bkh)'s the failure is prompt-shaped, not seat-shaped does not hold, and the cheap remedy is the first one. At E-20260808b stage 2, moonshotai/kimi-k3 returned zero content at finish_reason: length on three of four generation bodies — $0.365763 for nothing — while the fourth body, same model, same prompt shape, same batch, came back clean; so the same prompt shape both succeeded and failed on one seat. Changing the seat to mistralai/mistral-medium-3-5 produced all four arms cleanly for $0.042, an eighth of what the failures cost. Rule: when a generation prompt (not a rating block) burns its allowance on hidden reasoning, change to a cheap non-reasoning seat before raising the cap a second time, and regenerate every arm on the new seat so the within-hand contrast is not split across two authors — the one clean body from the failing seat is discarded deliberately, and the discard is recorded. | S133 | S133 | | (bkj) | Zero content characters with finish_reason: stop is NOT note (bhf), and neither of (bhf)'s remedies is the first move. (bhf) is a length body: the seat spent its whole allowance on hidden reasoning, and the remedies are (bhq)'s raised cap or a changed seat. E-20260808a produced a third shape — x-ai/grok-4.5 returned zero characters with stop on the primary block, $0.023064 for nothing, while the same seat completed three other blocks of the same shape in the same dispatch. A straight re-dispatch of the identical prompt to the identical seat returned 8,334 characters. Rule: on a zero-content stop, re-dispatch once unchanged before touching the cap or the seat, and preserve the dead body as always. Changing the seat here would have spent a second model's money to fix nothing and would have put one block on a different hand from its siblings. | S132 | S132** |

| (bkk) | A judgement that will define an admission condition must be shown to reproduce on a second text in a second language BEFORE the condition is written into a release. RS-20260807d measured the concordant/discordant call at Fleiss κ = 0.786 on one Ōgai courtroom scene, and ARM-marking-work step 2 was scoped to write it into a v0.2 as prediction 1′'s admission term. RS-20260808b put the identical three-question instrument to three readers of two whole Chekhov stories chosen because they are made of the phenomenon, and it came back at κ = 0.0584 — six utterances named between three readers, none named by all three, one reader finding it nowhere — while the markedness half of the same call from the same seats reproduced at 0.7214. The cause is not noise: discordance lives in a text's trajectory and the instrument asks about an utterance, and S128's κ was earned on a scene that restates the standing in every line. Rule: a κ from one text licenses this text; a condition a release will quantify over needs a second text, in a second language, before the wording is drafted. Corollary: check whether the high-agreement measurement came from a text where the property is locally restated — that is the shape that generalises worst. | S133 | S133 | | (bkl) | A rating scale that asks how a speaker presents his own standing will be answered, on a polite isolated sentence, as though politeness meant lowness — and the error lands on the SOURCE side, where there is no other cue. E-20260808b's two largest transfer discrepancies were both the general in «Смерть чиновника»: Russian seats scored «Вам что угодно?» and «Ах, полноте…» at −0.33 (naming «вы»/«вам» as the marking form) and English seats scored the same utterances at +2.00 and +1.67 — and the English seats were right. вы marks distance and respect toward the addressee; it does not claim lowness for the speaker. Neither the design nor the pre-run critic caught it, and it inflated every all-marked mean in the run (0.368 → ≈0.20 with the two sites removed). Rule: any standing scale must separate deference shown to the addressee from rank claimed by the speaker, as two questions, or it cannot be used on a symmetric-polite pronoun system. | S133 | S133 | | (blm) | Showing a seat every arm of a passage side by side turns a rating task into spot-the-odd-one-out, and one arm per call costs about the same. Every jury run in this project from S020 to S146 presented arms together, and E-20260809i's pre-run critic named the consequence: where one arm is built to be visibly anomalous, a seat can infer the experimenter's contrast from the comparison set and mark the odd member without ever consulting the rubric, so the run measures discrimination in a simultaneous comparison rather than how a translation scores when read as a translation. Reverse-ordering does not touch it — it presents the same contrast backwards. The run was rebuilt with one arm per call, no seat told that another rendering existed: 128 small calls for $0.740, against 54 large ones, because the per-call instructions and source repeat but the arms do not. Rule: where an arm is constructed to be conspicuous, score it alone; where arms are presented together, the run says so and reports the finding as a discrimination result. Companion to the halo rule (wiki/goodness-senses.md usage rule 6), which governs several senses in one call; this one governs several arms. RS-20260809i-terminology-drift §1, §6. | S147 | S147 | | (bln) | Choosing an obscure text does not buy an unrecognising panel, and the successor design that needs one must measure recognition as an admission gate rather than assume it. RS-20260809d §5b recorded 11 of 12 bodies naming "The Cask of Amontillado" through an anonymisation, and set its successor the requirement of a text the panel does not recognise. E-20260810-source-asymmetry went as far the other way as this project can reach — Bret Harte, "Brown of Calaveras", an 1870 magazine story by an author out of print in every language he was translated into, every personal and place name substituted, 794 words of dialogue and no narration — and 5 of 5 seats named Harte; 2 of 5 named the story, one at confidence high. Rule: a design whose validity turns on the panel not knowing the text dispatches a recognition probe FIRST, as a gate that can stop the run, and does not treat obscurity as evidence of anything. This is the third measured failure of the obscurity premise, after S065 and S079 measured it for translator independence — same premise, different consequence: obscurity is not protective against recall for contamination, and it is not protective against recognition for judging either. Companion to the closed measurement programme on contamination (CLAUDE.md), which this does not reopen: this note is about designs, not about the rule on lead independence. RS-20260810-source-asymmetry §4. | S148 | S148 | | (blo) | A design may register what a null WILL MEAN, but it may not register a destructive consequence for one — non-rejection is a failure to found, never a refutation, and a p-value above α may not delete work. E-20260810-legend-lexis's frozen decision table, and the translator's frozen log D47 that had already promised to honour it, both said: if the study limb finds no Károli phrasing, V12 is struck and erratum 1 flattens two spans of translation. The independent pre-run critic (BLOCKING 9) refused it before the run: non-rejection at α = 0.05 shows the effect was not found, not that it is absent; the design's "power floors" were sample-size thresholds with no registered minimum effect and no equivalence margin behind them; and a population-level association cannot license a change to five particular paragraphs in any case. The primary then did come back null (+0.0087, p = 0.4392) and the erratum did not fire. Rule: a pre-registered consequence is itself a claim and gets critiqued like one; where the consequence of a null is destructive, register instead (i) what estimate and interval would license the act, (ii) an equivalence margin if absence is to be asserted, and (iii) a unit-level analysis if the act is unit-level. The honest consequence of a second failed founding is that the artifact says so on its face — the register now labels the marking an explicitly unsupported translator's judgment — which is a stronger record than a deletion would have been. RS-20260810b-legend-lexis §4, §9; E-20260810-legend-lexis §11 A9. | S149 | S149 | | (blr) | No reader-side question yet found turns perceived-source-carriage into evidence that anything was carried — including asking the reader to point. Three runs had shown the seven-point rating cannot separate carried form from free oddity (RS-20260807f, RS-20260808d, and RS-20260809h's word-scramble above five of six foreignizing hands). The live repair was that a rating is permissive and a quotation is a commitment: ask the reader to quote the German words a marked English span carries, or answer NONE. E-20260810d did that at ten sites of one Arnim paragraph, four versions per site byte-identical outside one marked span. The competence gate PASSED at 0.90 — 27 of 30 cells quoted the registered locus where something really was carried — and the same seats attributed source-carriage to a word-scramble in 0.60 of cells and to ordinary faithful English in 0.4333, with licensed-minus-free at +0.200, exact P = 0.156 and the licensed arm 1.37 points stranger on the run's own scale. Rule: a design that needs to know whether a translation carries a source feature measures it on the two texts; a reader-side judgment is evidence about the reader, and no reformulation of the question has repaired that. Corollary for uptake and execution measures: RS-20260809h's bar — never without a matched oddity control in the same run — stands, and the pointing task does not satisfy it. What is NOT licensed: this is a failure to demonstrate discrimination, not a demonstration that licence is invisible; no equivalence claim is registered and none may be made. RS-20260810d-carriage-naming §3, §5. | S151 | S151 | | (bls) | A body whose finish_reason is length did not finish, and a parse rule that can read its first line will hide that. This project's dead-body practice (notes (bhf), (bhq), (bkw)) is written for the zero-content length case, and E-20260810d's registered F5 inherited it: MALFORMED or missing bodies are retried once. Eighteen bodies came back length; ten were unreadable and caught, and seven were readable — truncated part-way, with an intact first line that the design's own parse rule accepts as a verdict. They passed every check silently. Re-dispatching them at a raised cap changed six of eight verdicts, all six in the same direction, so the truncation was not cosmetic. Rule: finish_reason is checked before the parser runs, and a length body is dead whatever it contains; a run's cell-return criterion names the finish reason, not the parseability. Corollary, because the repair was decided after the direction was visible: where a protocol deviation is discovered post hoc, compute both datasets and report both — here every gate verdict and every primary reading was identical across them, which is the only thing that made the deviation harmless. RS-20260810d-carriage-naming §6.5. | S151 | S151 | | (bmh) | A binary same/different verdict is not evidence until its reason has been read, because paraphrase drift elsewhere in the passage counts as signal. E-20260811g measured whether a one-word source edit reaches an independent English hand, with the arbiter returning {verdict, what_differs}. Of 24 bound-morpheme cells, exactly 2 came back DIFFERENT — and reading what_differs showed neither was the manipulation: one seat had seen ghost against image, another the horses against a horse, both ordinary drift at other points in the same window. Counted, the bound arm arrived at 2 of 24; read, it arrived at 0 of 24, and a 0.083 arrival rate would have been published as a small real effect. The rule: any design whose statistic is a binary difference verdict collects a free-text reason in the same call and codes it, and the result page reports arrival-by-reason beside arrival-by-count. A verdict without its reason may be dispatched but may not be a primary. Sibling of (bkt): a control that cannot be read is not a control. RS-20260811g-carrier §4. | S161 | S161 |

Sources, texts and provenance

| (bif) | Three freely reachable copies of a public-domain text were one transcription of one printing, and the count of sources that carry a reading is not evidence that the reading is attested. Project Gutenberg 13976, Projekti Lönnrot 99 and Finnish Wikisource all reproduce «Köyhää kansaa» — and all three reproduce the same corruptions character for character, because W2's own front matter names the same transcriber and the same Otava 1917 volume that W1 carries. A page scan of the 1886 first edition was one search away, is public domain, is free, and reads Ei where all three e-texts read EL. Rule: before a source page or a translation treats an e-text as its text, establish what PRINTING it descends from and whether a different printing is reachable — and prefer an edition published in the author's lifetime to a posthumous reprint. The reprint here had silently standardised the eastern-dialect layer the translation was separately recording as a loss (collation.md §3a), which is the failure mode that makes this more than hygiene: a copy-text that pre-normalises the thing you are recording as a loss makes your record false in your own favour. | S100 | S100 | | (bjs) | A contamination gate's verdict is a property of the PROBE as much as of the pair, and a short probe in the wrong register understates it — sometimes by enough to reverse the verdict. S121's mandatory gate ran on a 163-word narration span and returned lead~King 8 twelve-grams / run 19 (DEPENDENT?), lead~Oxenford clean, King~Oxenford clean. S122 re-ran it on the 1,421-word dialogue scene the run was actually about: lead~King 33 twelve-grams / run 23, lead~Oxenford 13 twelve-grams / run 19, and King~Oxenford 6 twelve-grams / run 15. Two of the three clean verdicts did not survive, including the one on which RS-20260806e §2 rested the claim that the two published translations are independent of each other — which is what its admission condition needed. The design's own pre-run critic had registered the worry in words (amendment A6(iii), "the dependence verdict covers narration; CLOSE and FORCE are dialogue and may be King-tinged") and nobody measured it. So: the gate's probe must match the material at issue in LENGTH and in REGISTER — dialogue for dialogue — and a clean from a short out-of-register probe is reported with its probe's size and kind attached, never as a bare verdict. This is a rule about how to run the standing gate, not a reopening of the measurement programme CLAUDE.md closed at S082. RS-20260806e §6 limit 4; T-kohlhaas-R20-v2 §Dependence measurement. | S122 | S122 | | (bsk) | _djvu.txt throws away font information, and where typography IS the signal that is the whole finding. The Internet Archive derivative this project has read since S222 — and the *.xml DjVuXML alongside it — carry word coordinates and no font attributes; _abbyy.gz carries <formatting italic="true"> and exists for every scan the shelf uses. E-20260831's central result — that Eastwick marks Sa'di's Arabic by italic at 11 of 12 unframed loci and Ross at 8 of 10 — is invisible in _djvu.txt, and E-20260822c had already coded Arnold's verse FORM from line-break patterns rather than italics for exactly this reason without naming the cause. Rule: before coding anything typographic off an Internet Archive text, fetch _abbyy.gz; and establish the book's BASELINE italic use before reading any single italic as a marking — Arnold 1899 sets all his verse in italic (24.6% of the chapter) and Gladwin 1806 uses italic only for running heads (1.8%), so the identical observation means opposite things in the two volumes. | S235 | S235 | | (bsl) | A "does not cover this span" exclusion is a claim about a text and must be verified in the text, not from the title. E-20260827-declared-play declared "Arnold 1899 (frompersianguli00arnogoog) is a verse selection and does not cover this span. Declared, having been checked, not assumed" — and that volume, the first four bábs, prints باب دوم whole from its tale I. The right identifier was named and the wrong book's scope was attached to it: gulistanbeingro00arnogoog, a different Arnold, is a verse selection. The study lost a hand it had, and A-gulistan-hands §5 carried the false statement for eight sessions. Rule: an exclusion on grounds of scope is recorded with the location where the span's own first sentence was looked for and not found — a page or line number, the same evidence an inclusion carries. Sibling of (cde): a title is not an identifier, and an identifier is not a table of contents. | S235 | S235 | | (bsm) | A cap probed on a small batch is not a probe of the batch you will run. E-20260831b probed openai/gpt-5.6-terra on ten matching items at max_tokens 1500 — clean, 254 completion tokens — and then dispatched the same template at twenty items per call at a cap of 3000. Five of the first ten calls returned finish_reason: length with nothing parseable, at $0.060 each: $0.301 spent for no data. The visible answer scales linearly with the batch; the hidden reasoning does not, and on this seat it grew faster than the answer. Note (bsf) says a cap is a per-seat, per-task-shape measurement; this says the batch size is part of the shape. Fires at: any run that probes a stage on a few items and then batches it. Remedy: probe at the batch size you will actually dispatch, and treat a probe at a smaller batch as evidence about nothing but that batch. Kin to (bsf), (bsh), (brx). RS-20260831b-radif-hands §9. Second firing, S237, and it is the note working as intended: opening 167.955767498, closing 171.717925748, delta $3.762158250 against a per-request sum of $3.465563850. The session reported the pair as an observation and ledgered the per-request sum, without claiming either agreement or disagreement — which is all this note ever asks. RS-20260901-inversion-habit §8. | S236 | S237 | | (bsn) | Before a retrieval or matching stage, ask what the seat can see of the outcome — and mask it. E-20260831b matched Payne's 199 English odes to a Persian census by showing each ode's opening couplet. Two independent critics put the same finding first: the maṭlaʿ's second hemistich carries the first instance of the very repeated tail the study measures, so the match could be made partly on the outcome. Hiding the precomputed code was not enough — the raw text carried it. The last three words of each hemistich were removed and accuracy did not move: 10 of 10 held-out anchors masked, 10 of 10 unmasked. The mask cost nothing and removed the objection entirely. Fires at: any stage where an instrument is pointed at material that also contains the dependent variable — matching, retrieval, alignment, deduplication. Remedy: enumerate what the prompt's raw text reveals about the outcome, mask it, and probe the masked version against known items before adopting it, because a mask that destroys the signal is worse than the leak. Sibling of the shared-rater rule the S234 critics raised. RS-20260831b-radif-hands §2. | S236 | S236 | | (bso) | The OpenRouter key-usage endpoint lags, so a same-session delta is not a cross-check. RS-20260829-radif-hands reconciled its key delta against its per-request sum exactly, and the practice of reporting that as a check has stood since. At S236 the delta read $5.674416850 against a per-request sum of $1.947485275. A control call costing $0.000086 moved the endpoint by $0.000000000, so the endpoint is not counting something extra — it is behind; and the $3.727 residual is within $0.12 of the two preceding sessions' ledgered totals. Fires at: any session that reports the key delta as confirmation. Remedy: the per-request sum with "usage": {"include": true} is the ledgered figure and the only one, exactly as CLAUDE.md says; report the delta as an observation with its snapshots and never as agreement or disagreement, because an endpoint that can be days behind can produce either by accident. A second, smaller rule from the same session: write raw bodies to a tag unique per call — a fill pass that re-used its first pass's tag overwrote two bodies and $0.056 of the record. RS-20260831b-radif-hands §9. | S236 | S236 |

Translation practice

| (bie) | A rule that says "count it as you go" produces the number late and produces it wrong, because the sites already behind the counter are not counted. register.md V18a made a rule about a grammatical feature the translation cannot carry and instructed spans 4–8 to record every further occurrence so the craft report could state the total. The register's own running count said two sites. A whole-text search, run in one command the moment the question was asked, returned four — and the one it had missed was in span 1, already translated, behind the counter. The rule would have delivered the figure at span 8 and delivered it halved. Rule: when a feature becomes the subject of a register rule, count it over the WHOLE source text immediately, not span by span. This does not collide with the no-rule-for-an-unread-site rule (register.md V17): a count of where a feature occurs is not a decision about how to render it, and V17 forbids only the second. The same command also found a third verb in a family this project had twice declared closed at two (raahtia/hennoa/raskita, D74). | S100 | S100 | | (bim) | A mechanical diff over an OCR layer is a finding aid and never a witness — and it manufactures plausible substantive variants at about one in seven. S105 collated 5,372 words of a 1917 reprint against an 1886 first edition's OCR text layer: 62 rows survived noise filtering and nine dissolved on the page image, the printings agreeing exactly. Three of the nine would have been published as substantive divergences — laidallahan for laidalla hän, tuollaiseen for tuonaiseen, and murjottivat (sulked) for muljottivat (goggled), which would have changed an English word. The same run's four real corruptions were all confirmed on the image, so the discipline costs nothing it does not buy back. Rule: no reading enters a collation record, an erratum or a copy-text on the strength of a text layer; every divergence a diff reports is adjudicated against the page scan or is recorded as unadjudicated by name. | S105 | S105 | | (bjn) | Two machine renderings of one source from two different labs are not two independent renderings, and the gap is an order of magnitude. On 740 tokens of Verga, tools/dependence_check.py measured the lead's R04 against an unbriefed deepseek-v4-pro single pass at 90 shared 12-grams and a longest common run of 30 tokens — against 11 and 17 for Strettell 1893 versus Dole 1896, two published human translators three years apart. No shared prompt, no shared instruction, different labs. Note (bhb) said the lead matches itself at up to 37 tokens; this extends it across labs, so a panel model's rendering is not an independent second opinion on a lead rendering of the same source, and any design that needs a second modern translator must either measure the pair or say it did not. Two companions from the same run, both of which bound how the tool may be read: nine of ten pairs came back DEPENDENT?, including S~D, the one pair independent by construction if any is — on a passage this constrained a close rendering produces long runs whoever writes it, the recall floor of (bez)/(bgf)/(bgi) met again — and the informative cells were the zeros: the resistancy arm shared 0 twelve-grams with either published rendering. Read the table as a ranking against the published pair's own figure, never as a verdict. RS-20260806b-berman-occurrence §3. | S118 | S118 |

| (bry) | A second rendering of one text under a DIFFERENT declared policy, made with the first rendering open, is pulled toward it far enough to stop being a different arm — and the size of the pull is a fact about composition ORDER, measured. «المقامة الحلوانية» has two arms by one hand under two policies written to differ: restraint (R43) and Preston's balanced period (R50). On «المقامة الصنعانية», where the balanced arm was written last, they share 19 twelve-grams and a 15-token run. On «الحلوانية», where it was written first and was open on disk, they share 216 twelve-grams and a 48-token run — "and the glittering of his show, I looked hard into his features, and let my eye run over his marks, and there he was — our shaykh of Sarūj". Eleven times the overlap from reversing the order and nothing else. This is note (bhb)'s hardest firing and the first on two texts the regimes were written to make different. Fires at: every multi-policy design in this project — the shape that produced RS-20260825b, RS-20260825c and this run. Remedy: run tools/dependence_check.py on every pair of the lead's own arms before designing an evaluation on them, exactly as the contamination rule already requires against published hands; treat a DEPENDENT? pair as a near-duplicate control rather than as two arms, and drop it from the design by the measurement — which is what E-20260827b did to its BAL arm. And where the order can be chosen, write the arms in the order that puts the instrument arm first. RS-20260827b-shown-or-told §10; workshop/translations/maqamat-hulwan/R43-v1/translation.md §Contamination. | S227 | S229 | | (brz) | A de-chime must be checked for the chimes it CREATES, not only for the ones it removes — and a hand-checked one will ship both. R49 rule 2 says change the colon-end word of one colon per graded chime. Applied to «الحلوانية», the first pass left two chimes standing (examination ⁄ consideration, men ⁄ woven) and created two that were not in the treatment text at all (meet ⁄ brought, well ⁄ people) — a control arm that chimes where the treatment does not is worse than no control. dechime.py caught all four before dispatch because it re-grades every consecutive colon-end pair in both texts and fails on either direction; it also enforces rule 3 by failing if any colon outside the substitution table moved. Seven loci additionally needed a second substitution, because R49 rule 2 is written for two-colon loci and this maqāma has runs three, four and eight deep. Fires at: every subtractive control this project builds — the R49 family, and any same text with feature X removed arm. Remedy: make the control's construction a script that re-measures the whole text and exits non-zero on a survivor or a creation; extend the rule explicitly where the source's runs are deeper than the rule assumes, and say so on the artifact. workshop/translations/maqamat-hulwan/R48D-v1/translation.md. | S227 | S227 | | (bsa) | A compound prediction must be checked for ALGEBRAIC ENTAILMENT against what the design itself says was already visible, before it is registered. E-20260827c registered P4 in two clauses — Chappelow's mean clause within 20% of Chenery's, and his clauses per Arabic colon at least 50% above Chenery's — and presented the second as independent confirmation of a mechanism. Within a panel the Arabic denominator is shared, and total words = mean unit length × unit count, so with the word totals the design's own §6 declared already visible (ratios 2.11 and 2.65), the first clause forces the second to at least 1.76 and 2.20. Both pre-run seats derived the identity independently and both called it BLOCKING. The clause could not fail. Fires at: any registered prediction with two or more clauses over quantities linked by an identity, and especially where the design has a what was already visible section — that section is exactly the input the entailment runs through. Remedy: write the identity out before registering; if one clause is implied by another plus a known total, delete it and report the quantity as arithmetic. RS-20260827c-declared-page; E-20260827c-declared-page/critic-response.md. | S228 | S229 | | (bsb) | A reproduction check on a SIMULATED statistic cannot demand exactness, and one that does is a control set to fail. E-20260827c's C1a required the pipeline to reproduce RS-20260826's published figures exactly, and the design halted the whole run on a C1a failure. Its deterministic half reproduced exactly — 107 and 123 printed lines, longest 14 and 13 words, CV 0.198 and 0.177 — but the six shuffle ratios are Monte-Carlo estimates over 10,000 draws and a different call order gives a different stream, so they landed 0.0001 to 0.0017 away. The pre-run critic named this as a MINOR finding about tie tolerances and it arrived in execution as a control criterion instead. Fires at: every extraction check, replication check or figure audit that re-runs a permutation, bootstrap or shuffle null. Remedy: split the criterion — exact for deterministic figures, within a stated tolerance derived from the number of draws for simulated ones — and state both in the design, not after the run. RS-20260827c-declared-page §8. | S228 | S228 | | (bsc) | Before declaring that a comparator volume does not contain a work, open its contents. A-hariri-hands §4a stated that Chappelow 1767 does not contain al-Ḥarīrī's second Assembly. It does: his Assembly II is headed HULWANENSIS, and his six are al-Ḥarīrī's first six in order. The claim was inferred at S223 from the shape of his page — running prose, no colon marking, so he had been excluded from that session's measurement — and then written down as a fact about his book. It cost the panel its second declarer for a session, and the second declarer turned out to carry the arm's headline. Fires at: every anchor and every comparator table that records what a published volume does and does not contain. Remedy: a claim of absence about a printed book is a claim about its contents page; check it there, and record where it was checked. Note that the check is free — the volume OCR was already downloaded. RS-20260827c-declared-page §1. | S228 | S228 | | (bsd) | A preference item whose two members differ by a few words gives a seat nothing to be about, and position fills the vacuum — build the item from two WHOLE renderings and the same seats become order-consistent. Note (brs) recorded the failure at S224: on four line-end arms of one text, differing in three words, the same three seats returned the same label under an order swap at 0.548, one of them below chance, and every registered quantity was withheld. E-20260828 put the same question to the same three seats — which of these two do you prefer to read as English verse? — on two complete renderings of the same ghazal by one hand, and order consistency ran 0.852 at baseline and above chance in all five conditions, with a pooled first-position rate of 0.5296 inside its bar, 0 void of 270, and a 12-of-12 operational floor. So the remedy note (brs) proposed — give the judgment something to be against — is not the only remedy, and on this evidence not the first one to reach for: giving it something to be between worked, and the disclosure conditions that were supposed to supply the criterion moved nothing. Note the direction of the effect on the instrument: the condition that told the seats exactly what the source does had the lowest consistency of the five (0.593) and the most undecided windows. Fires at: any A-vs-B preference task built by editing one text into two. Remedy: where the question is about a policy rather than a word, render the policy twice and whole; a minimal pair is the right instrument for a minimal claim only. RS-20260828-forced-half §4; RS-20260826b-radif §5. | S229 | S229 | | (bse) | Ask whether a translator supplied a word and you must ask it against the HALF-LINE the English is rendering, not against the couplet — at couplet grain the answer saturates at 95% anchored and the measurement dies. E-20260828 set one English word against the Persian bayt its block renders and asked three seats whether the word answers to anything in it. The seats were not the problem: 12 of 12 on known-answer calibration items, 19 of 24 decoys called supplied, zero UNSURE in 104 study items, pairwise agreement 0.94–0.97, and the same low rate in each seat separately (0.029, 0.038, 0.087). The problem is the haystack: fifteen to twenty-five Persian words in two hemistichs offer something for nearly any English word to be pointing at, so the pooled supplied rate came to 0.048 and the run's own saturation gate withheld both primaries. The design had chosen couplet grain deliberately, to make supplied a conservative verdict; it bought the conservatism with the whole primary. Fires at: any design that codes addition, omission, or correspondence between a translation and its source. Remedy: align to the smallest unit the source itself divides — the hemistich, the colon, the clause — and pay the alignment cost up front; and where the alignment cannot be had, expect a base rate too low to test and say so in the pre-flight rather than after. Kin to (bsd), which is the same lesson at the other end of the instrument: there the item gave the seat nothing to be about, here it gives it too much. RS-20260828-purchased-figure §4. | S230 | S230 | | (bsf) | A max_tokens cap probed on one item does not verify it for the next, on a model whose hidden reasoning is variable. E-20260829 bought 27 short, uniform, structured classifications at a cap of 1200 and, as note (brt) requires, checked the first three returns for truncation before dispatching the rest. They passed. P2 then truncated on the seventh call — same task shape, same prompt template, a different item — having spent 1,123 of 1,196 completion tokens on hidden reasoning. The visible answer needs about seventy tokens; the reasoning is what varies, and it varies by item, not by shape. So (brt)'s remedy — probe the shape — is necessary and not sufficient. Fires at: any run buying a fixed-format answer from a reasoning model at a tight cap. Remedy: size the cap for the reasoning budget rather than the answer, which on this panel means several times what the visible output needs; and structure the prompt so the load-bearing fields come first, because a truncated body that already carries every field it is scored on is re-parsed at $0 under (brx) rather than re-bought — which is exactly what saved this call. RS-20260829-radif-hands §8. Fired again S232 and the second firing is instructive: at a cap of 300, 92 of 270 rating bodies from the same panel returned finish_reason: length, almost all P2 — but 86 of them had already delivered the scored line and were re-parsed at $0 under (brx), and only six lost it and had to be re-dispatched at cap 1200, for $0.016. So the failure mode is not "truncated" but "truncated past the answer", and only a finish_reason check plus a tolerant parser distinguishes it from a dead call. The remedy below is confirmed, and its second half — put the load-bearing field first — is what turned a 92-call problem into a 6-call one. RS-20260829b-inversion-price §11. | S231 | S243 — eighth session running; S239 was the seventh; S237 was the sixth; see the tail of the note. Earlier: S234 — fourth session running, and now on four seats at once. At max_tokens 2500 on a nine-line answer, google/gemini-3.6-flash and qwen/qwen3.7-max both returned finish_reason: length on the first item probed, and z-ai/glm-5.2 returned an empty body after 6000 completion tokens; at 6000 the two panel seats came back clean and six of thirty-eight stage-S bodies truncated anyway. The cap that works is now a per-seat, per-task-shape measurement and nothing else. RS-20260830b-rhyme-family §7. Fired again S236, fifth session running, and this time it cost money: the cap was probed on this shape for P2 (3000 truncated, 8000 clean) and not re-probed for P1, which had returned a ten-item probe of the same template clean at 1500. P1 then truncated on five of ten batch calls at 3000, at $0.060 each for nothing — $0.301 that bought no data. See the new note (bsm): the batch size is part of the shape. RS-20260831b-radif-hands §9. Fired again S237 — sixth session, and the first run that honoured the note in full and was bitten anyway. Every seat was cap-probed at the batch size it would actually be dispatched at, as (bsm) demands; the probes cost $0.297 and were worth it (4000 truncates P1 and P2 at 24 items; 12000 clears P1 and still loses six of P2's 24; P2 needs batch 12). Then one P2 batch in 44 truncated at the probed cap and the probed batch size, losing 12 items. The note cannot be discharged by probing: probing bounds the loss, it does not remove it. What remains is to parse what came back, count what did not, and report it. RS-20260901-inversion-habit §8. Fired again S239, seventh session running, and in a direction the note does not cover — see the new note (bst), where a LARGER cap made the answer worse. Two of 54 stage-G bodies returned EMPTY after spending the whole 2500-token cap on hidden reasoning and one stopped mid-list; the three dead bodies cost $0.092926 and the re-buys at cap 6000 cost $0.062950, so the abandoned calls were the more expensive half — a body that spends its whole cap and returns nothing bills for the whole cap. RS-20260902-rhyme-slot §8.2. Fired again S243, eighth session running, and in its sharpest form yet: google/gemini-3.6-flash truncated 6 of 12 batches at cap 2000 on a task whose visible output is eight lines of the form LI.11/GAR KEPT, pinning at exactly 1,996 completion tokens each; the identical items at four per batch and cap 6000 finished 4 of 4. Batch size is the variable — (bsm) confirmed from the other direction. config/budget.md 2026-09-04. |

| (bsg) | Never gate a primary on a pooled main effect when the hypothesis predicts an INTERACTION — the gate fires exactly when the hypothesis is most true. E-20260829b v1 registered a manipulation check M1 (pooled inversion penalty ≥ 1.0) and made it a withholding gate on both primaries. Its pre-run critic showed that the design's own best case — a large penalty in one register and none in the other — drives the pooled penalty down toward the gate, so the strongest possible confirmation would have withheld the result. Fires at: any crossed design that gates on a marginal quantity the interaction is defined as varying. Remedy: gate on controls only — a positive control, a compression or floor rule, a completeness rule — because those are properties of the instrument and are independent of which way the hypothesis falls; report manipulation checks as descriptive and let them withhold nothing. RS-20260829b-inversion-price §6, critic-response.md P3-1. | S232 | S232 | | (bsu) | When you count whether a source rhyme can be answered in English, count what the rhyme is MADE OF — repetition crosses every language boundary, a grammatical ending crosses none, and pooling the two makes a class that behaves two ways. ARM-gulistan span C explained a nought — 0 end-rhymes in 44 bayts — by counting 19 of 44 as rhyming on a Persian grammatical ending, and filed radif (a word or phrase repeated after the qāfiya) in that same class. Span D counted its own 38 bayts before translating, registered the prediction, and the prediction half failed: the count was exactly right about rhyme (1 taken, in a lexical-qāfiya bayt, at no cost) and blind to chime, because every line-end chime the English took came from repetition — the radif «گو مباش» becomes say, let there be none and both couplets end identically in English exactly as in Persian, while the radifs made of «را» and «کن» carry nothing because English has no word there to repeat. A lexical radif is free; a grammatical radif is unanswerable; and a grammatical ending is unanswerable. Fires at: every census of could this rhyme be carried — the ghazal work, ARM-radif, ARM-radif-hands, §7.42 and §7.49 — and at any classification whose category is a position (the line end) rather than a material. Remedy: classify by material, not by position: split radif from ending, and record separately how often the target chimes by repeating a word rather than by finding a rhyme, because that column is the one a translator can actually act on. spanD-qafiya-precount.md; T-gulistan-bab2-R05-v1 D47, D48; register V18. | S240 | S240 | | (bsv) | When you code whether a hand marks a foreign quotation, code the quotation's type AGAINST the type around it, not "italic / not italic" — a hand whose surroundings are already italic must mark by setting the quotation ROMAN, and an italic-only measure scores that as unmarked. E-20260903-latin-in-italian coded 21 Latin loci in «Vita Nova» across four English hands. Three of the four set Dante's divisioni in italic as a body and Martin extends it to the whole of chapter XXV, so 10 of 84 cells sit inside italic. On the italic measure the retained runs are marked 62/70 = 0.886 and §7.45.5's registered prediction fails on two of three marking hands; on the contrast measure they are marked 69/70 = 0.9857, and each of those two failures was the same single cell. Fires at: every marking census on a printed book — §7.45, §7.50, A-gulistan-hands, and any future bilingual source — and generally at any measure whose category is an absolute property of the marked span rather than a relation between the span and its context. Remedy: record the setting as <span type>-IN-<context type> and derive distinguished from it; report the direction of the contrast alongside the rate, because the direction is a fact about the volume's typography and the rate is a fact about the hand. The Persian material had no italic environment, which is why §7.45 could not see this and why the note is not a criticism of it. RS-20260903-latin-in-italian §2, §3; E-20260903-latin-in-italian/coding.md. | S241 | S241 | | (bsw) | Read the price of every seat a design dispatches from the API, in the session that dispatches it — config/models.md is a cache and nothing re-reads it. On 2026-09-03 GET /api/v1/models returned $2.00 / $12.00 per M for openai/gpt-5.6-terra against the $1.00 / $6.00 the table had carried since S106: the first upward move this project has recorded, and the first that can breach a declared ceiling rather than merely over-price one. The three earlier corrections (S061, S106, S182) were all downward and all went unnoticed for weeks precisely because a downward error is invisible — the run comes in under. Fires at: every pre-flight estimate, not only the ones that dispatch a reserve seat, which is where the S182 gate stopped. config/models.md §2026-09-03; E-20260903b-line-end-order §11. Honoured S243, and this is what honouring it looks like when nothing has moved: all four seats re-read from /api/v1/models before the ceiling was set, all four unchanged from the S242 re-read, ceiling built on the read and not on the cache. A no-op reading is the note working, not the note idle. config/budget.md 2026-09-04. | S242 | S243 | | (bsx) | Separate enjambment from inversion before counting line-end word order — they are confounded at the line end by construction. When a marked line's clause continues into the next line, its last word is not syntactically final, so a “would-it-be-last-in-prose” coding calls it displaced whether or not the hand displaced anything. In E-20260903b-line-end-order's fresh 60-item gold, all 21 enjambed items were coded INVERTED, against 21 of 39 among complete-clause items. Two hands who enjamb at different rates would then show an inversion-rate difference that is really an enjambment-rate difference. Fires at: any count of line-end word order (§7.47, §7.51, ARM-line-end-order and any successor) — add a clause-completeness field and report the primary on complete-clause items, or stratify. RS-20260903b-line-end-order §5. | S242 | S242 | | (bsy) | Buy a blind second coder on the one dimension a quoted passage can settle — it is cheap, and the coder who wrote the rule is the coder who breaks it. Two pre-run critics made it BLOCKING that E-20260904 had one coder who was also the predictor, the locus classifier and one of the translators. The repair — 96 cells to one masked seat, hand identity hidden, no prediction and no framework shown, five codes — cost $0.119 and agreed at 92 of 96 (0.958). It also cost a registered prediction: one of the four disagreements flipped P4 and is a lead error, at a cell where the coding instruction the lead wrote and handed the seat verbatim lists Monsieur among the French forms that count as kept, and the lead had coded one hand's Monsieur Pierre as englished while coding another hand's M. Pierre as kept. Fires at: any run where one reader codes every cell of a comparison — which is most of this project's coding studies. Remedy: send the quoted cells out blind on the single dimension the quote settles (retention, presence, direction), register the disagreement gate and the flip rule before dispatch, and report a flipped prediction as unresolved rather than correcting the cell into a pass. It is not a jury and it does not check location — the seat sees the passage the lead extracted — so it checks classification only, and that is worth saying on the result page. RS-20260904-french-in-russian §6. | S243 | S243 |

API, cost and panel

| (bfa) | A void criterion must not conflate "the instrument failed" with "the phenomenon is real". E-20260730g's F1 voided condition B when the 1:1 paragraph alignment covered 51.4% of the source against a registered 90%. Two entirely different states produce that number: the aligner drifted, and the translator did not preserve paragraph boundaries — and only the first is a reason to void. Both were in fact present: the arithmetic was consistent (69 + 7 − 1 = 75, i.e. seven genuine splits) and hand-verification found the aligner wrong at three of eight non-1:1 blocks. A coverage threshold cannot tell them apart, so a design that voids on coverage must also register a check that separates them — here, hand-verifying every non-1:1 block, which the design did require and which is what made the verdict readable instead of merely negative. | S066 | S066 · S071 — its own repair registered as F1a/F1b and DISCHARGED: the coverage threshold was unreachable by construction (a perfect alignment reaches 0.568) while the instrument failed independently, and splitting the criterion is what made both readable | | (bfb) | (fired again S074 — the same cell that had just produced note (bgc)'s whitespace body returned this shape on its retry; a plain third dispatch to the identical slug returned complete output.) A third dispatch failure shape: finish_reason: stop, content: null, and the whole answer — one line of thirty-seven — inside message.reasoning. x-ai/grok-4.5 billed 466 completion tokens, did not truncate, and did not return nothing. Note (b) is about length; note (beq) is about a payload-free socket; neither covers this. The remedy is neither fall-through nor blind retry: dropping reasoning: {"effort": "low"} and re-issuing to the same slug returned a complete, correctly formatted answer at 429 tokens. That is note (bdl)'s shape — change the parameter, do not change the seat — applied to a failure note (bdl) did not describe. Read message.reasoning on every null-content body before deciding what failed. | S066 | S066 · S077 — its fallback LAUNDERED a bad body; see (bgp) | | (bfc) | Declare a reserve for EVERY dispatch stage, not for the critic alone, or a stage that fails makes you choose a seat after seeing the failure. E-20260730g declared a reserve for its critic and for one reader seat. Note (b) then fired on a third stage that had none, and the fall-through seat was picked after the failure — which is the one thing a declared reserve exists to prevent, because the choice can be made to suit the result. The seat chosen (P1, whose other task could not leak into this one) was defensible and the point stands: defensible-after-the-fact is not the same as declared-in-advance, and the difference is not visible in the output. Cost of the fix: one line per stage in the design. | S066 | S158 — and in the place the note does not yet name: the stage that overran was a GATE. E-20260811c declared $1.80, spent $1.846625863, and the $0.047 overrun is exactly the parity re-dispatch. A failed judge stage can be scaled down; a failed PARITY control cannot, because a run with no parity data is a run about damage. A gate's reserve is not discretionary and must be priced at the full cost of dispatching it twice. | | (bfe) | A FOURTH dispatch failure shape — HTTP 504 "The operation was aborted", no body, no usage object — and it is NOT free. moonshotai/kimi-k3 at max_tokens 12,000 returned it twice; max_tokens 8,000 with reasoning: {"effort":"low"} returned a complete verdict from the same slug at the same provider. That is note (bdl)'s remedy — change the parameter, not the seat — on a failure neither (b) (truncation), (beq) (payload-free socket) nor (bfb) (answer inside message.reasoning) describes. The expensive half is the billing. Key usage was identical to the digit immediately after the second 504, and by session close the delta EXCEEDED the per-request sum by $0.0437483996 — the first excess this ledger has recorded, where every previous cross-check ran short (note (bco)). S061 built exactly this check — assert the delta as positive and not exceeding the sum, which would also catch a billed call with no stored body — and recorded that no excess appeared. It appears now. So: an error response is a lower bound of zero, not a cost of zero, and the per-request sum is a lower bound whenever a dispatch returned no usage object. | S067 | S067** |

| (bgb) | A recognition YES is not evidence that recognition occurred, and on non-canonical material it is usually the tradition rather than the work. E-20260731f asked both raters to identify a lead translation of Kielland's «Karen» (1882, Norwegian, no canonical English). Three YES answers, all three wrong — twice "The Poet, Isak Dinesen", once "Himmerland Stories by Johannes V. Jensen" — both Danish, both roughly the right literary neighbourhood, neither the author. A fourth, "Karen by Johannes V. Jensen", looks like a near-miss and is not: the name Karen appears eleven times in the passage, so the title is readable off the page. RS-20260730d §5 established that stripping authorship from a lead translation of a canonical work leaves the raters knowing the work; this is the other failure mode, and a design that scores recognition as a binary would have counted all four as hits. Score recognition against the truth, never as a YES/NO rate. | S074 | S074 | | (bgc) | A FIFTH dispatch failure shape: HTTP 200 whose body is nothing but keep-alive whitespace — no JSON object at all. x-ai/grok-4.5 returned 2,167 bytes of spaces and newlines, which raised JSONDecodeError in the caller and aborted a chained multi-stage run, losing the two stages behind it. Not note (b)'s truncation, not (beq)'s payload-free socket, not (bfb)'s answer-inside-message.reasoning, not (bfe)'s 504. Two remedies, both taken: write the raw bytes before parsing (already standing, note (bdt)) and retry a non-JSON body up to three times, preserving each bad body; and never chain stages with && when any link can raise, because the cost of the crash is the stages that never ran, not the call that failed. The same cell then failed a SECOND time in a DIFFERENT shape — (bfb)'s null content with the tokens inside reasoning — and a third dispatch to the identical slug returned complete output. One cell, one seat, two shapes, two retries. | S074 | S074 · S077 — fired TWICE in one session, the first time that has happened: 1,584 bytes to openai/gpt-5.6-terra and 1,331 bytes of pure whitespace to x-ai/grok-4.5. It crashed the runner on the session's FIRST dispatch, before any guard existed, and that body was lost; the guard built in response cost the second occurrence one cell rather than one stage · S102 — qwen/qwen3.7-max at Alibaba, 924 bytes of whitespace on a source-only stage; the retry guard held and the run continued, which is the first time this shape cost nothing but one dispatch | | (bgd) | When a claim's truth depends on a tokenisation choice, print both counts rather than picking one silently. A-mchugh-presence asserts 1,511 words. Whitespace tokens return 1,511, exactly; a letter-token regex returns 1,506, the difference being numerals and P&G. Run 1 of the check used the letter rule and would have reported the anchor wrong. The rule that reproduces a published figure is evidence about which rule produced it, and the check now declares the rule it uses and prints the alternative beside it. | S074 | S074 | | (bge) | An instrument repaired after seeing its failures must ship its first run beside its second, or a zero yield is unreadable. E-20260731f's stage-2 checks failed 3 of 18 claims and 20 of 53 quotations on first writing; every failure was the instrument (substring matching that fired ere on were and hell on Hello?; the anchor pages' own trailing commas; ellipsis-elision; a two-text anchor). After four repairs, yield 0 of 18 and quotations 53 of 53 — and "I repaired the checker until the failures went away" is indistinguishable from "the pages are clean" unless run 1 is stored. Each repair must be justified by a convention of the audited page, stated as a fact about the page, and run 1 kept as a file. | S074 | S074 | | (bgr) | A control item's qualification does not transfer between runs, and note (bfy)'s own prescribed remedy is what failed. (bfy) says "a control item is qualified by DATA — a prior run in which it behaved as declared." E-20260801b did exactly that: six items reproduced verbatim from E-20260730i's payload, read out of the frozen prompt file rather than retyped, where P1 scored 6/6 and P3 scored 6/6. The same two seats scored 3/6 and 4/6, and the failures are item-specific rather than diffuse — X2 (British-1890s register, options car / automobile / truck / carriage) went 0 of 5 payloads for both, having been correct at S068. What differed between the runs is small and enumerable — standing register American → British, surrounding items Portuguese → Polish, item numbers — and none of it touches X2, which carries its own register note. Rule: a control block must be scored on the run that uses it and reported there; a prior run's pass is provenance, not qualification. A design that leans on inherited controls has to be able to survive their failing, which means F1 must be written before the run and obeyed after it. | S078 | S078 | | (bgs) | k is a mean of the unstable judgment over a set fixed by the stable one, and nobody had reported it with that structure named. On a byte-identical repeat of one rater in one session, which candidate bears reproduced at 30 of 30 cells (Jaccard 1.000) and how many options it excludes did not — 0.6333 exact agreement, and the second pass excluded +0.467 options per cell more than the first. This is the reverse of RS-20260730i §4's ordering, which measured the cross-rater version and found BEARS unreliable at 0.21–0.51 on positive cells and the exclusion judgment reliable. The two are compatible because they are different quantities, and the pair is the point: the denominator of k is set by the reproducible step and its numerator by the irreproducible one. Rule: any statistic that is a mean over a selected set must report the reliability of the selection and of the value separately — a single agreement figure for "the instrument" hides which half is broken. | S078 | S078 | | (bgt) | A SIXTH dispatch failure shape: finish_reason: "stop" on a body that is unambiguously truncated. google/gemini-3.6-flash returned a body ending …I30 \| C3:- and then the two characters I3 — mid-item-id — with stop, at 5,996 completion tokens, the identical count to two bodies from the same seat in the same session that reported length. Note (b) is about length; (bga) is stop with content: null and the answer in reasoning; (bgp) is a line count is not a format check. None of them covers a stop that lies. What caught it was the registered failure criterion F4 — fewer answer lines than the payload has items — applied in the analysis; the runner's guard (finish_reason == "length") let it through and wrote it to disk as an accepted seat. Two remedies, both taken mid-run as amendment A6: move the item-count check into the runner so a short body is a seat failure whatever the provider says, and raise the seat's cap rather than change the seat (note (bdl)) — 6,000 → 20,000, which returned 46 of 46. Unlike S077, the higher cap cost MORE ($0.138051 against $0.053262), so (b)'s remedy paying for itself is not a rule. Rule: finish_reason is provider testimony, not a measurement; the only truncation detector a design can trust is its own expected item count. | S078 | S078 · S079 — THREE more, at two labs in one stage: bodies stopping mid-list at stop / native completed, 303 / 219 / 244 content tokens against a max_tokens of 8,000. No cap explains it. The remedy that worked was a required END terminator plus an explicit line count, both enforced in the runner: 6 of 6 seats returned 22 lines afterwards | | (bgu) | A mutation that does not mutate is a passing test that tests nothing, and a verifier that reads the first non-empty body reads a REJECTED one. E-20260801b's verifier failed three of its own checks on first run and all three were the verifier's or the analysis's fault, not the run's. (i) A mutation declared as "turn every C1 dash into two exclusions" rewrote the literal string C1:-, which the target body does not contain, so it changed nothing and the harness reported MISSED — it was reporting honestly only because a paired identity (must NOT change) mutation existed to give the comparison meaning. (ii) The verifier resolved a seat's stored body by taking the first .raw with non-empty content, which for a seat whose first attempt failed with a non-empty payload is the rejected body; the analysis had the mirror defect and silently dropped that seat entirely. Rules: assert that each mutation actually changes the bytes it claims to change, and resolve a seat's body by the artifact the runner marks as accepted (here: the label that has a .txt beside it), never by scanning for the first one that parses. | S078 | S078 · S079 — the inert-mutation assertion is now code in the verifier's harness rather than a rule in its author's memory; 6 of 6 mutations changed bytes and 6 of 6 were caught · S081 — 6 of 6 mutations changed bytes and 6 of 6 were caught, with the byte-change assertion inside the harness | | (bio) | A mutation test must target the finest-grained stored quantity, not the headline: thresholded counts absorb the very defect the test exists to catch. E-20260804g's third mutation flipped one graded answer inside a raw billed body, and the verifier did not notice — recovery is a >=2-of-3 threshold over a both-orderings rule, and one flip moved a site from 3 seats to 2 without moving any reported number. It caught only after a per-site, per-condition seat-count check was added alongside the totals. Rule: for every reported aggregate, the verifier must also recompute the un-aggregated quantity it is built from, and the mutation test must be aimed at the latter. Otherwise a verifier certifies that the arithmetic on top of the data is right while remaining blind to the data. Companion to (bgu) — that note makes a mutation actually mutate; this one makes the mutation observable. | S106 | S106 | | (bhd) | A mutation test that re-runs a scoring script rewrites the results file the next check reads, so the mutation survives the restore. S084's verifier restored each mutated input TSV and asserted the bytes back, then reported six failures on the second run — because score_pair.py and score_controls.py write pair_results.json / control_results.json, and those were left holding the last mutation's output. The inputs were clean; the artifact was not. Restore is not restore unless it covers every file the mutated run can write, not just the file that was mutated — snapshot them before the subprocess and assert them back after, exactly as (bgu) already requires for the mutation itself. Cheap general form: run mutations against a copied tree, or treat any file the scoring script opens for writing as part of the restore set. | S084 | S084 | | (bhl) | A cadence arm's alarm is invisible to check_balance.py, and its own self-test says so. ARM-atelier-cycle declares one visit at least every three sessions, and under-visiting is the alarm — the cadence analogue of a step budget. It was last worked at S085 and the tool reported no violation at S086, S087, S088, S089 and S090, because a budget: cadence arm is excluded from the budget check by construction: tools/check_balance.py's own test asserts violations(... budget="cadence", used="9") == []. So the arm ran two sessions past its declared alarm with a clean tool run every time, and the arm-staleness list — which did print ARM-atelier-cycle 4 — is a count, not an alarm, and names no threshold. Rule: at session start, read every budget: cadence arm's declared cadence off its own page and compare it to N since worked by hand; the tool does not do it. The tool fix was considered and declined as method work under the subject rule — each arm states its cadence in prose, nothing published is false, and no deliverable is blocked. | S090 | S090 | | (bhm) | A declared loss that names no attempt is an assertion, and one counterexample is enough to retire it. register.md V14 said the child-to-parent honorific plural had "no compensation available that would not be a fabrication". E-20260802d put the passage cold to three independent translators in two framings and 6 of 6 cells marked nothing, which supports the claim — and Hertzberg's authorized 1886 Swedish marks it with a kin term («Blir mamma länge borta?»), in a target that also had the pronominal option and set it aside. A device used by the only human translator of the work on record is not a fabrication, so the rule was false while every machine result agreed with it. Rule: a translator's log may declare a loss, but the declaration must state what was tried and rejected; a bare 'nothing is available' is a claim about the language and needs a witness, not an intuition. Corollary, measured here: agreement among cold machine translations is not that witness — six of them missed what one 1886 translator found. | S090 | S090 | | (bhn) | A rule written for a site outside the material in hand was wrong the one time it was tested, and its stated ground was false about the text. Span 1's D11 decided the rendering of «Jumal' antakoon» at ¶79 — a site in span 2 — on the explicit ground that "the greeting recurs and the register must bind". It does not recur: the formula occurs exactly once in the 573-paragraph novella, and the one occurrence was the site being pre-decided. The cost was not merely a wrong rendering (the comparator in fact took D11's side, RS-20260802d §3.4) but a binding register rule carried for five sessions on a false premise about the source. Rule, now register.md V17 and general to any serial work: no rule is written for material not yet read in place; a decision that reaches forward must be recorded as a question for the span that owns the site, never as a binding. | S090 | S090 | | (bhp) | A grammatical loss declared without a forced second attempt is not a finding about the language pair, and across this project's whole record the declaration was wrong at five sites of six. E-20260802e took every site in the filed logs where a marking was said to be carried by a grammatical category the target lacks and declared wholly lost — 8 sites, of which 2 had already been refuted by published comparators (Dole 1896, Hertzberg 1886) — and re-rendered the remaining 6 under a brief requiring the marking to appear in any category. Three independent seats, blind to provenance, recovered the relation from the forced rendering at 5 of 6, from a length-matched unmarked decoy at 0 of 6. The one failure (Spanish grammatical gender) was predicted before the run and is metalinguistic: the relation was female and automatically so, and no English category encodes automaticity. The same census found 8 Class B sites where these same logs record the move being made — pronoun → benefactive verb, suffix → possessive, pronoun → address noun — so the rule was being applied and denied by the same hand, eight times each, with nothing connecting them. Rule: a log may not record a grammatical marking as lost until a rendering has actually been attempted that requires it to appear somewhere else; the absent category is a fact, the absent marking is a hypothesis. This is framework/v0.1/'s R1 and the first prescriptive recommendation this project has admitted. Corollary from T-wang-liulang-R04-v1 D13, which named two English renderings and rejected them as "costume" while asserting "no English slot": impossibility and ugliness are different claims and a log must not make the second while writing the first. | S091 | S091 | | (bhq) | A seat's hidden reasoning is a capacity constraint, not only a cost one, and max_tokens sized from the answer produces zero-content bodies. Note (bhf) has fired three times on price — P2 gemini-3.6-flash at 4–6× P1 for the same passage. E-20260802e fired it on capacity: at max_tokens 4,000, sized generously for 24 one-line answers, P2 returned finish_reason: length on four of four attempts (4,177–5,691 reasoning characters) and P4 kimi-k3 returned length with 16,803 reasoning characters and ZERO content characters for eighteen one-line ratings. Both recovered at 16,000–20,000 with byte-identical prompts. The failed bodies billed $0.196 and are ledgered. Fired again S106: P4 at max_tokens 12,000 returned length with 12,000 reasoning tokens and zero content, billed $0.2128698 — 39% of that session's spend — and the seat was replaced rather than re-sized, which is what (bhf)(i) prescribes once a slug has failed the shape twice. Rule: size max_tokens from a seat's measured reasoning appetite plus the answer, never from the answer alone; and when a body returns length, raise capacity and re-dispatch as a declared amendment rather than accepting a truncated body (note (b) already forbids the latter). The worst-case estimate must then be built from the raised cap — note (abc). | S091 | S091 · S092 — fired as a precaution, not as a failure: caps of 16,000 and 24,000 were set from this note and six of seven bodies then spent over 95% of their completion tokens on reasoning for a 143-character answer; 7 of 7 accepted first attempt, 0 length failures. | | (bhs) | A failure criterion built as a ratio is not a test when its denominator can be measured as zero, and this project wrote one. E-20260803-a4-set registered the paraphrase floor must not exceed 2 × R(s, j), the same juror's retest floor on identical text. P1 reproduces itself exactly on accuracy, voice and affect — retest floor 0.000 — so 2 × R is zero and any movement whatever fires the criterion. It fired on four of P1's six senses and three of P2's, on differences that are every one of them exactly 0.500: one juror, one pass, one point of a seven-point scale. The pooled comparison the criterion replaced says the opposite and is right — 0.139 against 0.233. The amendment that introduced the ratio form was an accepted BLOCKING critic finding and was a genuine improvement in one respect (it paired the two quantities per juror) while introducing this defect in another. Rule: a criterion stated as a ratio, a fold-change or a multiple must state what it does when the denominator is zero or near it, before the run; the cheap general form is an additive floor (P ≤ R + c) or an explicit clamp. Corollary, and the reason this is worth a note rather than a line in a limits section: an accepted critic amendment is not thereby a correct amendment, and the acceptance record must not be read as one. | S094 | S094 | | (bht) | A predictor test needs a no-information benchmark registered beside its null, and this one reversed the reading. E-20260803-a4-set asked whether a translator's frozen log predicts the sense a blind jury scores lowest. It registered a criterion (r̄ ≤ 2.20) and an exact null (each item's rank multiset, 46,656 outcomes). The test fired: r̄ = 2.000, p = 0.0098. It then took nine lines of arithmetic, computed after the run and registered nowhere, to establish that a predictor which never opens the log and names naturalness every single time scores 1.500 — better. The null asks does it beat chance; the benchmark asks does it beat the cheapest thing that is not the predictor, and only the second question was load-bearing. The same shape would have caught a majority-class classifier reported as accurate. Rule: any design whose claim is that some source of information predicts an outcome must register, before the run, the best predictor that ignores that information — usually a constant, a marginal, or the outcome's own base rate — and report the two side by side. A p-value against chance is not evidence that the information was used. | S094 | S094 | | (bhu) | contamination: is a property of the SPAN, not of the work, and a measurement on one span does not transfer to another. T-mare-au-diable-R04-v1 is filed contamination: none on a ch. II measurement against Sedgwick and Ives. At S094 the same translator, the same work, the same two comparators, on ch. XI, returned 16 tokens and 13 shared 12-grams against Ives and 13 / 2 against Sedgwick, with same-book nulls at 2–4 — a discard, twice over. The ch. II figure was then recomputed on the filed text and reproduced exactly (9 and 10 tokens, 0 twelve-grams), so nothing published is false; what is false is the reading of the front-matter field as a fact about the work. Daudet fired the same session at 14 / 3. Three of the project's last five gate runs have fired, and the two firings here are on works whose other spans are clean. Rule: the gate is a per-span measurement and must be re-run for each new span of an already-gated work; a contamination: value states the span it was measured on, or it states nothing. The standing rule in CLAUDE.md already says one call per candidate unit — this note says the unit is the span. | S094 | S094 | | (bhv) | A registered criterion that reverts a decision must be ORDERED against the criteria that vindicate it, and the ordering must be written before the run. E-20260803b measured one string in one sentence: a compensation for a Finnish honorific English cannot carry. It worked — +1.167 on the target rating, exact permutation p = 0.0368 over all 924 splits, 3 of 3 seats moving the same way, 5 of 6 quoting the device unprompted, the specificity control unmoved at 0.167. It was reverted, because a naturalness tripwire registered at ≤ 0.50 came back at 0.833 and the design's decision table made that row unconditional and first. On the raw scale the device gained more than it cost (1.167 against 0.833) and went anyway. The pricing — a defect weighted at half a benefit — was set before any call, on the ground that a translation answers for its faults before its bonuses; the point of the note is not that pricing but that the pricing existed and was ordered in advance. Without the ordering, the same six numbers support keeping the sentence, and the argument for keeping it would have been written after they arrived. Rule: a design with more than one gate states which gate wins, in what order, before dispatch; a design with only gates that can vindicate has no gates at all. Corollary from the same run: the pre-run critic's first BLOCKING finding was that no outcome of the original design could have changed anything — that is a checkable property of a design, and it is worth checking on every design that tests the lead's own work. | S095 | S095 | | (bhw) | A no-information arm can leave a passing result intact and still change what it means. Same run. (bht) required a benchmark that ignores the information; here it was a seat shown no text at all, only a novel set in a Finnish town, 1886; a nine-year-old wakes her sleeping mother because two visitors have come in. It rated the girl's deference at 3.667, five of six cells at 4 — above both English versions. So P − N = −2.167 and M − N = −1.000: both translations under-deliver against what a reader brings to the scene, and the compensation closes just over half the gap. The recovery result (+1.167, p = 0.0368) is unaffected and still true. What changed is the sentence it licenses: not the device works, keep it but the loss is worth about 2.2 points on this instrument and the best available compensation recovers half of it — a price on a grammatical loss where the project had only ever recorded a declaration. Rule: a no-information arm is not only a veto; register it wherever a design will otherwise report an effect without reporting the scale the effect sits on. | S095 | S095** |

| (bhx) | Two accepted critic amendments broke each other, and the collision was invisible to the critic that produced both. E-20260803c carried a POSITIVE arm — the target relation stated outright in the English — as its reading check: a seat that misses it is not reading. The same critic pass then required amendment A12, a header reading "Default to NO … answer YES only if the relation is POSITIVELY MARKED in this English", to stop the instrument reading relations into unmarked prose. A12 worked — the registered floor check passed at 0.095. It also made POSITIVE fail by being read correctly: P2 answered NO to ten of twelve POSITIVE items with reasons of the form "English does not positively mark …", and YES only at the two sites where the marking sits in the speech rather than in the gloss. The registered bar was 11 of 12 per seat, so the run is void, and the void is not the seat's fault. Rule: when a critic pass changes the grading instruction, every registered gate that depends on that instruction must be re-derived under the new wording before dispatch — an amendment is a change to the instrument, not only to the prose. Corollary, and this is what makes it a note rather than a limits line: a reading check that a correct reading can fail is not a reading check, and this one had passed three earlier designs unexamined because no earlier design told its seats to default to NO. | S096 | S096 | | (bhy) | A blind coding payload can hand the seat the answer in the GLOSS, with no target text shown at all. E-20260803d asked three seats, from the Russian alone and with every English rendering withheld, whether a transparent English calque is available for each of fourteen items. The design's own §5 certified that no English was shown. The glosses then read "the spirit of the forest" for леший, "the spirit of the water" for водяной, and quoted the formula outright for крестная сила — i.e. each item's gloss WAS the calque whose availability was the question. The pre-run critic called it BLOCKING and was right; every gloss was rewritten to describe the referent by what it is believed to do and where it is met, with no paraphrase of the word's parts. Rule: when a design withholds one representation of an item, audit every other field for a paraphrase of it — the gloss, the context sentence, the item's own id. Corollary the same pass forced: the repair is reduction, not removal (among trees still carries half of wood-), and the record must say which it achieved. | S097 | S097 | | (bhz) | A probe run to SELECT material is priming, and it primes on exactly the material the study is about. Before translating a word of «Бежин луг», S097 ran a string-frequency count over both published comparators — domovoy, nymph, sprite, water-, wood-, rusalk and twelve others — to decide whether the class was worth a session at all. It printed counts, no sentences, and it still told the lead which comparator transliterates and which substitutes, on the exact fourteen items the run would then measure. The project's standing rule (never prime a workshop translation with a stored published rendering of its own source) was not broken by the letter and was gutted in effect. The design absorbed it — the lead is excluded by construction from the run's primary and the priming is declared in full on the artifact — but that was a repair, not a plan. Rule: a selection probe over a comparator is part of the translation's provenance and must be run in a form that cannot leak, or run AFTER the translation is frozen. A count is not automatically safe; what makes a probe safe is that its output cannot be read as a handling. | S097 | S097 |

| (bih) | A write-once guard that holds across a script's retries does not hold across the script being run twice, and the second run silently overwrites a BILLED body. Note (bgz) made every attempt inside one invocation uniquely labelled. At S101 run_critic.py was launched, appeared not to have started, and was launched again; both dispatches were billed, and the second wrote to the same label and destroyed the first's stored .raw. Its cost survived only in the run console, and the per-request re-sum was short by $0.0586302 until it was added by hand — the key-usage cross-check is what would otherwise have carried the discrepancy. Rule: a runner refuses to overwrite an existing .raw and labels a repeat dispatch _re2, _re3; write-once is a property of the FILE, not of the loop. Second half of the note, and the more interesting one: the two dispatches were byte-identical prompts to one slug at temperature 0, routed to two providers, and returned different verdicts (NEEDS-AMENDMENT, four findings / NEEDS-REDESIGN, six findings) with only two findings in common. How much a single pre-run critic pass warrants is now an open question — the accident was worth more than the call it duplicated, and the union of the two was accepted. | S101 | S101 |

| (bik) | An answer-line guard must accept every format the instruction actually permits, or it converts good bodies into seat failures and spends the reserve on nothing. E-20260804d told seats to output ITEM=<id> ERRORS=<n> and set line_re=r"^ITEM=". P2 wrote ITEM I01 ERRORS=1 — the separator, not the format — and both of its stage-1 bodies were complete, terminated, correctly populated, billed, and rejected as seat failures, triggering two reserve dispatches at $0.0575388 that bought nothing. The bodies were recovered from disk afterwards and scored. This is note (bgw)'s failure mode from the other side: (bgw) says a guard must not carry a stale pattern forward; (bik) says a guard must not be stricter than the task it guards. Rule: write the acceptance pattern to match on the item identifier, tolerating any separator (^ITEM[= ]), and — because a rejection is expensive and silent — print the first 200 characters of any body a guard rejects, so a format mismatch is visible at the moment it costs money rather than at analysis. | S104 | S104 | | (bil) | The key-usage delta LAGS the per-request sum by minutes, so a closing snapshot taken immediately after the last dispatch under-reports — and that lag, not a discrepancy, is what has been reported as "not closing". S105 snapshotted at the last dispatch and got a delta of $0.078312 against a per-request sum of $0.132372; the same key re-read a few minutes later returned a delta of $0.132371900, equal to the sum to 1e-9. The intermediate snapshot showed the reason exactly: it had caught P1's $0.0085345 and neither P2's nor P3's. S104's non-closing cross-check ($0.482830081 against $0.534977746) is the same shape and the same direction, and there is no evidence it was anything else. Rule: take the closing snapshot when the ledger is written, not when the last call returns, and re-read once before reporting a residual. Note (bfv)'s void-on-concurrency rule is unaffected and still governs; this note only removes the false positive that a prompt snapshot manufactures. [fired S107, second time, and the same magnitude: an immediate read gave a residual of −$0.0786, a twenty-second-later read gave 0.000000000. Twice now the lag has been large enough to look like a real billing anomaly, so the rule is restated as an instruction rather than a caution: never report a residual from a snapshot taken at the last response — re-read, and report the second figure.] | S105 | S107 |

| (bit) | A pre-run seat probe is worth what note (bhf) has been claiming since S087 — and a token ceiling learned from it must be applied to EVERY stage, not to the one the author was worried about. E-20260805 ran the probe for the first time. It fired immediately: P2 returned finish_reason: length with empty content twice at max_tokens 64 — S108's exact failure, which cost that session 60% of its spend — for $0.000188. Re-probed at the run's own ceiling P2 answered normally, at $0.001188 against P1's $0.000094 for the identical 24-character reply: 12×, all hidden reasoning. So the probe's yield is not "is the slug alive" but a per-seat minimum token budget, and that is what the probe should return. The lesson then cost money anyway: the amendment raised stage 2's ceiling and left stage 1 at 4,000, and stage 1 burned $0.147 on six length bodies from the two heavy-reasoning seats before every stage was raised. Rule: probe at the ceiling the run will use, record the per-seat minimum, and apply it wherever the probe's condition holds. Seat-failure waste fell from S108's 60% to 23%. | S109 | S109 |

| (bjc) | A difference statistic can shrink from either end, so a design that changes a rating item's wording and compares a DROP must report the two scores that make the drop, separately, or it will report an improvement that is a demotion. E-20260805e re-ran Tier D's failed drop(naturalness) under the revised naturalness string and the number fell from 1.083 to 0.750, across the ≤ 0.75 bar. The whole of that fall is the reference losing points: reference 6.125 → 5.417 (−0.708) against the damaged copy's 5.042 → 4.667 (−0.375), and the reference's ceiling ratings went from 6 of 24 at 7 to 0 of 24. Read as a drop alone it looks like accuracy damage no longer bleeding into fluency; read as two numbers it is a criterion string that marks the undamaged text down. And the same run shows the mirror image: on two new items the identical 0.333 shrinkage came entirely from the damaged copy scoring higher while the reference did not move at all. Same effect size to four decimals, opposite mechanism. Rule: report the reference mean and the comparison mean beside every drop, in every cell, and never let a ceiling criterion stand in for this — the run's FC4 watched only for reference scores at the ceiling and would have passed a cell whose reference had simply been marked down. | S113 | S113 |

| (bjd) | Resolvability must be defined WITHIN a run. Distance from a previously published point estimate is a random throttle, not a noise floor. E-20260805e v1 made a wording effect reportable only if it exceeded |this run's retest − the published 1.12|. The pre-run critic showed the failure mode in both directions: a lucky retest landing near 1.12 licenses almost any effect, an unlucky one suppresses a large one, and the comparator is itself a high-variance random variable on 12 correlated units. The replacement — an exact paired permutation test on the per-unit differences, plus a frozen minimum effect size — is what withheld the run's primary: the effect cleared δ_min at 0.333 and the permutation test returned P = 0.138. The two properties are genuinely different and this run measured both: the cell mean reproduced across three separate runs to 0.037, while the per-unit drops swung by up to 1.5 scale points. Rule: a gate on "is this effect real" reads the run's own paired structure; a published prior figure is a retest STATISTIC to report, never a threshold to gate on. | S113 | S113 |

| (bje) | Navigating a comparator file is priming, and the remedy is to move the locus, not to flag the artifact. Note (bhz) covers a selection probe over a comparator; this is the plainer case that had never been named. At S114 the lead located Quijote I.20 by searching the Gutenberg files for chapter boundaries and printed Ormsby's and Smollett's English of the passage it was about to translate — not a count, the prose itself. R04 §1 was breached before a word was rendered. The temptation is to translate anyway and declare the priming, which the regime permits; that produces an artifact that is evidence about nothing, because a primed rendering cannot speak to independence, to process, or to what the source forces. R04 §1 supplies the real remedy in its own text — "the passage selected must be one whose rendering has not already been seen" — and it was applied: the locus moved forward within the same chapter, to prose whose English had not been opened. Rule: locate the locus in the SOURCE file and freeze its boundaries before any comparator file is opened; if a comparator's prose has been read, move the locus rather than declaring the priming. The declaration is the fallback when nothing can be moved, not the first response. | S114 | S114 |

| (bjf) | Two works by one author can be the same prose, and a corpus arm that does not check for reuse weights one text twice while reporting two. E-20260805f built Machen's "original English" arm from The Great God Pan, The Three Impostors and The House of Souls — and The House of Souls (1906) reprints The Great God Pan whole, along with The Inmost Light. A 500-token probe from the Pan file appears verbatim in the Souls file. The arm was counting the same prose twice, and weighting it by whichever collection happened to be longer. Nothing in the design would have caught this; it surfaced only while checking the premise of an unrelated critic finding about control-cell text reuse. Anthologies, collected editions and "complete works" files are the common carriers, and a title that names a collection rather than a work is the signal. Rule: before blocking a multi-work corpus arm, cross-check every pair of member texts for a shared n-gram run — one contiguous run of a few hundred tokens is a reprint, not a coincidence — and record the check. The cost of missing it is silent and invisible in every downstream statistic. | S114 | S114 |

| (bjg) | A multilingual panel cannot proxy a monolingual reader, and the questions where that matters are exactly the ones about what a reader does NOT know. E-20260805g put a passage containing one untranslated Swedish line to four seats and asked, uncued, for a retelling. All four reported the content of the line — one naming the language with no warrant on its page and translating it, one translating it silently — where the artifact's own declared reader is "a reader with no Finnish … no facing text, no notes". The translator's log had declared a cost that depends entirely on the reader being unable to read that line, and no prompt can make a model not know Swedish; instructing it to pretend would measure compliance, not comprehension. The same trap waits wherever a design turns on ignorance rather than judgment: an unglossed foreign phrase, a dialect a target reader would not have, an allusion the reader is meant to miss, a name whose connotation is supposed to be opaque. Rule: before designing any comprehension measure, ask what the seat would have to NOT know for the measure to work, and check whether the panel knows it. If it does, the panel cannot answer, and the honest output is that the cost is unpriced rather than a number. Corollary: this is the first thing in this project that a bigger, better panel makes worse. | S115 | S115 |

| (bjh) | A guard that records THAT a dispatch failed but not WHY is half a guard, and the half it keeps is the useless one. call.py wrote repr(e) to a .err file on any transport failure. An HTTPError's repr is <HTTPError 400: 'Bad Request'> — the provider's actual message is in the response body, which urllib makes readable exactly once, from the exception object, and which was being discarded. Six mistral-medium-3-5 dispatches died undiagnosably at S115 until one call was reproduced by hand outside the runner, whereupon the provider said "top_p must be 1 when using greedy sampling" — a constraint that fires only when the reasoning field is present, and a one-line fix. Rule: store the error BODY beside the repr, and print the first 200 characters at the moment of failure. This is note (bik)'s rule — print the first 200 characters of any body a guard rejects — extended from bodies that arrive to bodies that do not, which is where it was missing. Repaired in call.py the same session. [FIRED S116, second session running, and the repair paid: mistral-medium-3-5 returned the same 400 twice as a screen seat and the stored body named the cause at the first failure, so the declared reserve took the seat under (bfc) with no hand-made reproduction and no lost time.] | S115 | S116 |

| (bjp) | A serial arm's cadence must be declared in units of the ROTATION, not of sessions, or it is unmeetable the day it is written and every visit reads as an overrun. ARM-atelier-cycle declared one visit at least every three sessions and was visited at S085, S090, S095, S100, S105, S110, S115 — every fifth session, seven times, without one exception. That is not drift: the selection rule gives one principal unit per session to the most neglected of five eligible tracks, so a track returns on a period of about five, and no discipline available to a session could have produced three. Two consequences followed and both were costly. check_balance.py structurally cannot see it — its own self-test asserts that a budget: cadence arm never fires a violation — so seven overruns produced zero warnings and the alarm was raised by hand, first at S090. And the unmeetable cadence was twice used as an override reason to take T1 over the tool's named track, so an alarm nobody could clear became an argument for taking the arm. Rule: a serial arm declares its cadence as one visit every N rotations, where a rotation is the number of eligible tracks; an arm that wants to be visited faster than the rotation is asking for an override, and must say so at birth rather than accrue an overrun that reads as neglect. ES-20260806-craft-report-koyhaa-kansaa §2. | S120 | S120 | | (bjq) | A device you exclude from your own arm by argument is a device a published translator may have used, and the census will tell you — so measure published practice BEFORE ruling a device out, not after. E-20260806-published-loss §5 excluded the archaic second-person pronoun from its FORCE arm, before translating and before any published English was read, on the argument that thou is one-sided: it can mark the down-address and English has no pronoun for the up-address. Oxenford 1844 used it, at four of nine sites, and it works — because once thou is in play, plain you becomes the marked contrast term. The exclusion's provenance is verifiable (75cd449) and it stands as registered; what it cost is that the run's one constructive arm used a weaker device family than the only translator who carried the relation. Rule: when a design must choose among the devices a recommendation licenses, extract what the published comparators actually do FIRST — a device count is textual, needs no panel and no jury, and is available before any translation is written — and let the exclusion be informed rather than argued. The general form: an argument about what a target language can do is a hypothesis, and the published record is the cheapest test of it that exists. RS-20260806e-published-loss §4.3. | S121 | S121 | | (bjt) | A fix applied to a run inside the session that needed it, and not committed to the runner, does not exist for the session that resumes the run. S121 hit repeated provider hangs, pinned deepseek/deepseek-v4-pro's routing to the providers that had returned cleanly, wrote that it had done so in RS-20260806e §5 and in the ledger — and the pin was never in run.py. S122 inherited a runner identical to the one that hung, and the published sentence "routing for that slug was pinned" was false of everything that survived the session. Sessions here are ephemeral and only what is merged exists, which the charter says about work and is exactly as true of configuration: a mid-run mitigation is written into the committed runner, or it is written into nothing. Costed nothing this time — the fix took one edit and the three calls returned in seconds — and the false sentence is the real damage. [FIRED S124 and honoured: the deepseek-v4-pro provider pin was written into the committed call.py, not into the result page alone, before the run was restarted.] | S122 | S124 | | (bju) | [FIRED AGAIN S125, predicted in advance: the same selection rule was reused deliberately so the instrument would be unchanged, intrusion was declared inert-by-construction before the run, and it came back at 4 of 16 — chance.] A selection rule that fixes a property destroys that property as a discriminator, and the rule usually looks like good design when it is written. E-20260806f chose all four loci by one source-side rule — a span in which the narrator presents himself and almost nothing happens — because that maximises persona signal and minimises plot. It also guarantees that every narrator scores high on intrusion: the four came in at 4.50–6.00, and the single-axis matcher on intrusion reached 3 of 16, worse than chance, while irony reached 13. The axis was not weak; the sampling frame had no variance left in it. Same shape as RS-20260806e §7, where an admission condition defined on whether the English conveys the relation selected for sites whose content already conveyed it. Rule: before freezing a selection rule, write down which of the measured variables it holds constant, and either vary it or declare it unmeasurable in this run. RS-20260806f §6. | S123 | S123 | | (bjv) | A request killed before its response reaches the process can still be billed, so killing a hung call is a cost decision and not a free one. Note (bhf)'s S116 entry records the opposite — "the key-usage delta across that killed request is exactly 0.000000000: a request killed before its response returns bills nothing" — and S123 measured $0.0011020 and $0.0165294 across two killed pre-run critic attempts on moonshotai/kimi-k3, the whole of that session's unexplained key-usage residual. The S116 sentence is true of the request S116 killed and is not a general fact, and it is the sentence a future session would rely on before deciding to kill rather than wait. Both entries stand; neither generalises. Practical form: the cross-check does not close after a kill, and the residual is the kill, not a mystery. RS-20260806f §7 limit 11; config/budget.md S123. [FIRED AGAIN S124: killing run_coding.py mid-flight to apply a provider pin left a $0.016075170 residual — the whole of that session's unexplained key-usage gap, identified as the killed call rather than reported as a mystery.] | S123 | S124 | | (bjr) | Provider latency is a design risk, not just a cost risk, and a long run needs resumability before it needs anything else. config/models.md carries a standing caution that OpenRouter routing can move the price of a slug by ~4×. S121 found it moves wall time by more than 20×: deepseek/deepseek-v4-pro routed to eight distinct providers in nine calls, and several hung past a 600-second timeout, twice on the same block, stopping a 36-call grading stage at 33 and forcing the session to land an incomplete run. Two things saved it and one was added mid-run. Every raw body was written to runs/ before anything was computed from it (standing practice), and a resume clause — reuse a stored body instead of re-billing the call — was added while the run was stalled, which made three restarts free and leaves the remainder at three calls. Rule: any stage of more than ~10 calls is written resumable from its stored bodies from the start, and a stage that stalls is restarted rather than waited on. Corollary, learned the same session: pinning provider.order for a slug is an infrastructure fix, not a design change — the slug is what config/models.md specifies and provider is logged as provenance either way — but it must be declared on the result page. | S121 | S121 | | (bjw) | A scale's apparent one-sidedness is a property of the CORPUS put in front of it, not of the scale — and the cheapest test is to make the missing arm. E-20260806b withheld two of three primaries when its gate F1′ found the negative half of the explicitness and popular-speech scales used in 2.1% and 4.8% of 186 cells. Two explanations were open — a scale that cannot go negative, or five translations none of which went down — and no amount of re-reading the run could separate them, because both predict the same table. E-20260806g separated them by writing one deliberately downward translation under a frozen regime and putting it in the same line-up: the rates went to 17.0% and 15.7%, register to 32.7%, and the new arm read −1.500 with all three seats agreeing. The instrument was never the problem. Corollary, and the sharper half: the other arms' negative rates moved by a factor of three on byte-identical texts once the line-up changed, so negative-half usage is a contrast effect and no figure from a comparative coding run is an absolute magnitude. Rule: when a gate fires on scale one-sidedness, the first hypothesis is an empty region of the corpus, not a broken instrument; constructing one instance is a translation, costs $0, and settles it — and any design quoting a magnitude from such a scale states what else was in the line-up. RS-20260806g §10. | S124 | S124 | | (bjx) | A free-text justification tied to "your most extreme code" cannot support a gate defined on ONE named scale, and the mismatch is invisible until someone counts. E-20260806g's F4 — the gate separating the coders read it as low from the coders read it as bad — classified WHY lines attached to register codes, while the prompt (inherited verbatim, deliberately) asks each seat what drove its most extreme of three. Any version whose top code was explicitness or popular-speech therefore yields no register evidence at all, and the gate would have "passed" on whatever happened to be left. The pre-run critic found it; the repair kept the instrument frozen: restrict the gate to cases where the target scale IS the (co-)most-extreme, compute that coverage as a number, and declare a floor below which the gate is UNDECIDABLE rather than passed. Coverage came in at 22 of 45 against a floor of 15. Rule: a gate that reads free text names which text it needs and demonstrates that the prompt produces it — a justification field answers the question it was asked, not the question the gate wants. E-20260806g A1. | S124 | S124 | | (bjy) | A fixed-vocabulary lexical detector reports a PRESERVED pattern as absent, and it can only err in that direction. machine.py's signifying-network check looked for adorn|deck|furnish. Arm V preserved Verga's adornare network — "does the place up for you" → "that's for the place you did up for me" — and the detector returned empty, which reads on the page as network destroyed. Nothing in the run would have caught it: a zero from a keyword detector looks exactly like a true negative. The pattern was extended and the network found, and RS-20260806b's "one network survives and one does not" is a lower bound for the same reason. Rule: every count from a keyword detector is a LOWER bound, a zero from one is never evidence of absence, and an arm that scores zero is read by hand before the zero is recorded. RS-20260806g §4. | S124 | S124 | | (bjz) | A recognition confound is broken by rebuilding the corpus, not by arguing about it — and the gate must be probed on BOTH sides, because a prior can only anchor a match if the seat identifies both texts. RS-20260806f measured a source→target persona crossing at 16 of 16 and had to report in its own headline that both recognition seats named all four works and all four authors from the English, leaving carriage and shared-prior equally consistent with the table. E-20260807 rebuilt the corpus out of authors with no reachable English translation record, held the instrument byte-identical so fame was the only intended variable, and promoted recognition from a reported number to a gate: 0 of 8 from the English, 1 of 8 from the source, crossing 16 of 16 anyway. Three parts of the rule. (i) The English-side probe alone is not the gate the confound needs; the pre-run critic's BLOCKING 1 is why the source-side probe exists, and it is where the run's only true hit came from. (ii) The scoring rule must be the correct surname, not the recognised flag: when these seats guess they guess a famous author of the right language and period — de Maistre for Karr, Saltykov-Shchedrin for Levitov, Shiga Naoya for Miyaji — so four confident wrong names read as four recognitions under the flag and as zero under the rule. (iii) The gate is a tolerance, not an α: naming a correct nineteenth-century author by chance has no estimable rate, and a design that promises to calibrate one is promising the assumption it is calibrating. What the gate buys is bounded and must be said in those words: it excludes identification-mediated recognition, never a diffuse shared prior. RS-20260807-narrator-unknown §§1–2. | S125 | S125 | | (bka) | Matching your TARGET lengths does not defeat a size-surrogate matcher, because the surrogate compares target size against SOURCE size — match the sources. RS-20260806f §6 found a one-dimensional nearest-neighbour on word count getting 3 of 4 works right and told the next design to match on length. E-20260807 matched the four English renderings to 448–466 words, ±2.6%, and wrote into its own design that the matcher was "expected to be at chance". It came back 2 of 4 — because the four source spans were 352 / 431 / 466 / 391 in the design's units and the surrogate reads across the pair, not within a side. One of its two hits is free by construction: the frozen rule fits the CJK character-to-word divisor to the item so the surrogate is as strong as it can be, which makes that item match itself exactly. Rule: a length-confound control is defeated by matching the sizes the surrogate actually compares — for a cross-language matcher that is the SOURCE sizes, in the surrogate's own units, including any fitted conversion. Corollary: a design's stated expectation about its own control is a prediction like any other and is reported when it fails. RS-20260807 §6. | S125 | S125 | | (bkb) | A mutation test that cannot change the decision it tests is not a test, and a uniform perturbation cannot change a nearest-neighbour argmin. verify.py's inherited mutation set added +3.0 to every axis of one profile vector and asserted the recomputed match count changed. At S125 it was caught 2 of 3 — and the miss is a property of the mutation: a uniform shift moves a point equally away from all four consensus points, so the argmin need not move at all. The verifier was not weak; the perturbation was orthogonal to the decision rule. Five mutations chosen to be decision-relevant — four label swaps within a seat, one flattening of a seat's whole target side — were added and caught 5 of 5, and both sets are reported. Rule: pick each mutation by asking what the analysis's decision rule is a function of, and perturb THAT; a mutation set that has never been checked against its own decision rule can report a clean verification of nothing. Corollary: report the mutations that did not bite, with the reason, rather than only the count. RS-20260807 §8. | S125 | S125 | | (bkt) | A positive control must manipulate the SAME COMPONENT OF THE GRAMMAR as the phenomenon whose absence is being claimed. E-20260808g set out to show that a Hungarian tag's tense morphology is not carried into English, and gated that null on a control that changed the tag's lexeme (mondá → kiáltá, said → shouted). The pre-run critic's BLOCKING 2: models carry a lexeme change trivially, so passing it would have licensed nothing about morphology, and the null would have been unreadable. The control was replaced before dispatch with a same-lemma, same-slot tense change (mondá → mondja, said → says). Both controls then returned exactly 1.000 — but only the second one is the reason the 0.000 means anything. Rule: write the control by asking what kind of thing am I claiming is not carried, and manipulate a thing of that kind; a control one component away is a demonstration that the instrument works on something else. Companion to (bkb): pick the perturbation from what the decision rule is a function of. RS-20260808g §3. | S138 | S138 | | (bku) | Before claiming a translation loses a source's marking, manipulate the marking and see whether independent hands move. This project has retired one prediction after five attempts in five language pairs (framework/v0.2 S1), and every one of them began by assuming the source's marked form meant something and going to look for its loss. E-20260808g inverted the order on a new device: it took eleven paragraphs, changed one token in each — the tag's tense — gave the two versions to four independent hands who never saw both, and asked only for a translation. Cross-arm disagreement 0.0000 against a same-arm floor of 0.0000, with the component-matched control at 1.0000. The cost was $0.169 and four API calls, against the several thousand dollars and dozens of sessions the loss-hunting shape has consumed. Rule: where a design is about to assert that English cannot carry a source form, the cheap first move is a minimal-pair generation probe — the loss claim is only worth pursuing if independent hands' English actually moves with the source form. Corollary, and the substantive one: a form bound to a construction may be marking something the target marks by other means — Hungarian marks the attribution slot morphologically, English marks it with a bleached said, and neither loses anything to the other. | S138 | S138 | | (bkv) | A gate must be measured at least as precisely as the quantity it gates. E-20260808h's primary was read on 6 ratings per site (3 rater seats × 2 calls); its source-side gate — the precondition that the manipulation was legible as a change of manner at all — was read on 2 ratings per site, one per arm, on the same five-point scale, whose observed between-seat spread is 0.696 scale points. The gate's bar was +1.00. It returned +0.667, F1 fired, and a primary computed from three times as much data was withheld by a coin. Worse, the same gate returned the wrong sign on the positive-control block (−0.75 on 8 ratings) while the English side returned +0.786 on 24 — the two sides disagreeing about a manipulation both were reading. Nothing in the design was wrong except the allocation: the manipulation was grammatical (0 residual forms, 0 grammaticality flags in 32 site-arms) and every other check passed. Rule: when a design has a gate, count the ratings behind the gate before dispatch and give it at least the precision of the thing it can withhold — a gate is not a cheap formality bolted to an expensive measurement, it is the measurement's licence, and an under-powered gate does not fail safe, it fails loud. Corollary, and the one that costs money: a gate is the wrong place to economise seats. RS-20260808h-trajectory §1, §4. | S139 | S140 — FIRED AND DEMONSTRATED, which is rare for a note of this kind: the repair was carried out and the number moved. E-20260809a stage R re-measured S139's source-side gate on the byte-identical texts, the same prompt string (asserted against the stored request before dispatch — the builder refuses to run otherwise) and the same +1.00 bar, changing nothing but the allocation: 2 ratings per site from 1 seat became 6 from 3. It returned +1.389 where it had returned +0.667, the gate passed, and a primary withheld for a whole session was released. S139's other under-measured figure resolved with it: the positive-control block's wrong-sign −0.75 on 8 ratings came back +0.500 at parity. So the note's cost is now priced in both directions — an under-powered gate does not fail safe, and repairing one is worth a session's principal unit. RS-20260809a-trajectory-ja §3. | | (bkw) | A seat that returns an empty body after spending its whole cap on hidden reasoning can often be repaired by turning reasoning off, not by re-seating. Note (bhf)'s remedy — zero content means change the seat, not raise the cap — is right about the cap and incomplete about the seat. In E-20260808h four bodies came back with zero content at cost: two critic seats and two of the three hands, $0.273475, 42.5% of the session. Both hands were recovered on the same seat by sending "reasoning": {"enabled": false} (deepseek-v4-pro, which then cost $0.0024 against $0.0040 for the dead call) or "reasoning": {"effort": "low"} where the first form 400s (grok-4.5). Rule: on a zero-content body, try the reasoning switch on the same seat before re-seating — re-seating costs a whole new call and loses the seat the design chose; two provider-specific parameter forms exist and one of them will be accepted. Caveat measured in the same run: google/gemini-3.6-flash rejects reasoning.enabled with a 400, so the switch is attempted, not assumed, and the fallback to re-seating stands. Companion to (bhf), which keeps the prohibition on raising the cap. RS-20260808h-trajectory §9. | S139 | S140 — extended: APPLY IT BEFORE THE FAILURE, NOT AFTER. E-20260809a dispatched every seat with a dead-body record (nemotron-3-ultra, kimi-k3) with "reasoning": {"enabled": false} from the start rather than waiting for the empty body. nemotron, which spent $0.046932 on nothing as S139's registered critic, returned a ten-finding critique for $0.011126; kimi-k3, dead in four of its last six load-bearing roles, returned two clean gate calls. Session waste fell from S139's 42.5% to 7.2%. A second measurement sharpens the note against (bhq): on stage R, three of six calls died at the cap that had returned clean bodies from the same seats on the byte-identical prompt one session earlier; the cap was raised to 12,000 per (bhf)'s seat/cap distinction and two recovered, and the third recovered only on the reasoning switch, at $0.0010549. A cap raise that failed and a reasoning switch that worked, on one seat, in one stage. RS-20260809a-trajectory-ja §9. S142 — the note's own dead end is closed. Its caveat says google/gemini-3.6-flash REJECTS reasoning.enabled with a 400 and leaves re-seating as the fallback; E-20260809c lost two stage-1 bodies at that seat (one an echo of the passage, one truncated at 8 items of 14, $0.0516, 17.9% of the session) and recovered BOTH on the same seat with the other form, {"reasoning": {"effort": "low"}}, at a raised cap. Applied prophylactically to every later stage, the seat then returned 9 of 9 clean bodies. So the caveat is two-sided: that seat rejects one form and accepts the other, and re-seating was never needed. RS-20260809c-generic-voice §9. | | (bkx) | A control is not frozen until its constituent has been shown to exist, as the control needs it, in the actual candidate material. E-20260809a froze a positive control that swapped an address title for a second-person pronoun in the tail, on the strength of a whole-novel census showing 先生 75 times and 奥さん 17 times inside dialogue. Applied to the selected sites it collapsed: most of those tokens are third-person reference, not address — the narrator talking to Sensei's wife about Sensei — and where the tail's address token was a pronoun it was the senior speaking to the junior, for whom no title exists. Only 2 of 6 sites could carry the manipulation at all, and only in one direction. The control had to be rebuilt after the design was frozen and after two critics had reviewed a version that could not be built. The census that justified it counted the right strings and answered the wrong question: presence of a token is not availability of a manipulation. Rule: before a control is frozen, run the mechanical check that the constituent it manipulates is present in the role the control assigns it, at every site the design will use, and commit the per-site count. One grep over the tails, keyed to the addressee, was the whole cost. Companion to (bkt), which governs what kind of thing a control manipulates; this one governs whether the thing is there at all. RS-20260809a-trajectory-ja §7 limit 5; E-20260809a amendment A7. | S140 | S140 | | (bky) | A reading probe whose raters cannot quote the manipulated constituent is not measuring the manipulation, and the required quote field says so for free. Two consecutive runs on two language pairs have now produced the same shape. E-20260808h: on the pronoun block 24 of 70 quotes contain an address term and every one of them is a term the manipulation never touched, because a Russian pronoun switch leaves nothing in the English to point at. E-20260809a, on Japanese: 2 of 36 quotes on the primary contain a contraction — while the mechanical census over the same post-split English shows contractions moving +2.639 per 100 words in the manipulation's direction — and 2 of 36 quotes on the positive control name the inserted formula, the control returning −0.125 against a +0.75 bar. In both runs the raters quote events instead: "the door of my house is closed to you", "The wife's eyes filled with tears", "he burst out laughing". The hands encode the change; the probe does not see it. Rule: where a design asks readers whether something changed, make the required quote a first-class measure and check it before the primary — if the quotes do not name the manipulated constituent on the block where the manipulation is largest, the probe has not measured the manipulation, whatever its means say. The corollary is cheap and load-bearing: validate a positive control by its quote rate, not only by its effect size. RS-20260809a-trajectory-ja §5; RS-20260808h-trajectory §5. | S140 | S142 — third consecutive firing, on a third language pair, and the first time the note WITHHELD a primary rather than explaining one after the fact. E-20260809c promoted the quote to a declared gate G6 with a chance baseline computed and committed before dispatch (0.2887, the share of the manipulated arm's tokens lying inside a declared edit span) — the pre-run critic's one wholly correct finding, which caught that the original bar of 0.33 was below chance. Observed: 0.2857, at chance, so F5 fired and P1 is descriptive only. The baseline is the addition the note now carries: a quote-rate gate without one cannot be read at all. RS-20260809c-generic-voice §7. | | (bkz) | A ladder of regimes executed by one hand cannot tell the programme costs accuracy from this translator's rule-following costs accuracy, and the arm that separates them is one extra call. RS-20260807e §5 and RS-20260808c §5 both published following either declared programme cost about 1.1 points of propositional accuracy — in two language pairs, from opposite programmes, with a contamination control and a compliance audit attached. Neither run contained an unruled independent arm: both had IND-R07 and IND-R08 and no IND-R06, so the quantity the claim is about had no lead-free measurement anywhere in the design. E-20260809b added that one arm for $0.0092 and it inverts the generality of the claim: the unbriefed hand's tax is +0.024 (P = 1.000, CI [−0.166, +0.250]) on the very passage where the lead's is +0.714, and +0.167 (ns) on the new one. The tax is real — it survives on a source no model can quote an English translation of, P1 +0.354, P = 0.039 — and it is a property of the lead translating under rules, not of rules. Rule: whenever a finding is about what a REGIME does, the design needs at least one arm of that regime executed by a hand that is not the lead, including the baseline regime — a ladder missing its own control condition in the independent half will produce a claim about programmes from evidence about a translator, and no amount of contamination control or compliance auditing inside the lead's half can catch it. Corollary measured in the same run: the published magnitude did not survive re-measurement either — the identical statistic on byte-identical texts under a byte-identical prompt came back +0.619 against +1.095, so cite such a figure for its direction and not for its size until it has been re-measured. RS-20260809b-programme-tax §3, §4.1, §4.3. | S141 | S146 — AMENDED by note (blk): the remedy stated here is sufficient for a positive finding and NOT for a null. Six hands returned +0.98 where this note's one hand returned +0.024 | | (bla) | A distributed carrier can move without any edited site naming it, so a site-level check cannot certify that an operator left it alone. E-20260809c's AWKWARD arm was built to roughen English without touching any carrier of voice, and arms.py asserted exactly that, mechanically, against ten named CLOSE strings — all ten survived byte-identical. The arm nevertheless came back +0.476 of diction formality above CLOSE, P = 0.03125, and register is the first of the five carriers voice's own entry names: twenty-one nominalisations, agentive passives and with the exception of periphrases are individually neutral and collectively bureaucratic. The design gated the other arm's register (G2 = 0.143, passed) and never gated this one's, so a registered reverse dissociation failed with part of the failure built into the manipulation. This is the mirror of RS-20260806-same-man's finding that register and rhythm are distributed properties a site edit cannot move: it cannot move them at a site, and it can move them in aggregate. Rule: where a design asserts that an operator holds a property, the property is measured on every arm the assertion covers, not only on the arm the primary is read from — a site-list check certifies sites, and a distributed property needs a distributed measurement, which here was one more blind rating stage the run had already built for the other arm. RS-20260809c-generic-voice §5, limit 1. | S142 | S142 | | (blb) | A control's items must be matched on every property its scale can read, not only on the property it is named for. E-20260809d's G2 asked whether raters were scoring the speaker rather than the words, and selected its four items on one criterion: turns with no address content. The two turns that fell to speaker A are his long heraldic answers — "A huge human foot d'or, in a field azure…" — and the two that fell to speaker B are his short questions. Both pairs are address-free and the pairs differ in length and diction formality, which is exactly what a ceremony scale reads. G2 came back at 0.7917 against a registered bar of 0.75 and withheld a primary of +3.500 at P = 1.08 × 10⁻⁵, and the run cannot say whether the gate caught a speaker halo or its own item selection. The pre-run critic named this defect (SERIOUS 4) and it was overruled on the ground that address-freeness was the property the gate needed — the wrong ground, because a control has to be matched on the scale's inputs, not on the hypothesis's variable. Rule: before freezing a control, list what the rating scale can respond to and check the arms are balanced on each of them; a control matched on one dimension and unmatched on three is an uninterpretable gate, and an uninterpretable gate still fires. Companion to (bkx), which governs whether a control's constituent exists at all; this one governs whether its items are comparable. RS-20260809d-forced-choice §4, §5a. | S143 | S143 | | (blc) | Substituting the names does not blind a canonical text, and the recognition probe that says so costs one line of prompt. E-20260809d replaced every proper name and the title object in Poe's "The Cask of Amontillado" — Fortunato→Guarnieri, Amontillado→Madeira, palazzo→house — and asked each seat, after its ratings, whether it recognised the story. 11 of 12 bodies named it, with author. What survives name substitution is the situation: a masonic exchange, a chaining, a walling-up. The canonical reading of that story is the run's own hypothesis, so every rating was taken by a seat that already held it. Rule: where a design needs raters blind to a work's received reading, name substitution is not a blind; ask for recognition in the same call, report the rate whatever it is, and prefer material the panel has no reading of — for a canonical author of a canonical story there may be no such passage, and that is a materials constraint to settle before dispatch rather than a limit to write afterwards. The probe is nearly free and is the only thing that turned an assumption into a number here. RS-20260809d-forced-choice §5b. | S143 | S143 | | (bld) | A permutation that cannot move most of its tokens is not a test, and the immovable count is what says so. E-20260809e's registered null permuted MARKED/UNMARKED labels within part × lemma — the right stratification, forced by the pre-run critic, because it controls for a lemma simply being frequent. On this corpus it was nearly vacuous: 40 of the 55 marked tokens, 72.7%, sit in strata containing no unmarked token at all, because most of Mikszáth's archaic verb forms are hapax lemmas with no ordinary-past counterpart in the same part. Immovable tokens contribute the same floor to the observed statistic and to every simulated one, so the null's spread comes from a quarter of the data and the p-value (0.1253) is substantially a power failure rather than evidence of absence. Rule: before reporting a p-value from a within-stratum permutation, report how many units the permutation can actually relocate; if it is not most of them, the test's failure is reported as underpowered and the design says so in the same sentence. A conditional null bought correctness at the price of power, and both halves have to be published. RS-20260809e-legend-layer §4a. | S144 | S144 | | (ble) | A test run on the corpus that contains the passage which generated the hypothesis must report the same test with that passage removed, and that sensitivity is often the result. E-20260809e tested whether Mikszáth's narration-internal archaic past clusters, on a hypothesis formed from two paragraphs of Part I. On the whole novel the statistic sat at the 97.5th percentile (p = 0.1253); with Part I excluded it collapsed from 7 to 1, p = 0.8050 — six of the seven contributing paragraphs were in the part that produced the rule. The pre-run critic named exactly this (SERIOUS 9: randomising across the whole novel partly re-discovers the passage that motivated the inventory) and the sensitivity exists only because the finding was accepted; had it been overruled, a p just under the bar would have been published as a device. Rule: name the passage the hypothesis came from before freezing, and register the leave-it-out recomputation as a required output, not as a robustness check to run if time allows. Companion to (blb): that note is about accepting critic findings on controls, this one about the specific finding a hypothesis-generated-from-the-data design will always attract. RS-20260809e-legend-layer §4b. | S144 | S144 | | (blf) | A billed API request survives the client that dispatched it, so a call started in a foreground shell with a shorter timeout than the call is money spent on nothing. S145's first loc dispatch ran in a shell the harness killed at two minutes while the request was open. The request completed server-side and was charged $0.0597891; no body was ever received, nothing was written to runs/, and the loss showed up only as a gap between the key-usage delta and the per-response sum — which is the one place it can show up, and the reason that reconciliation is worth doing every session. Rule: dispatch every API call in the background, or with a client timeout longer than the call can possibly take; never in a foreground shell that something else can kill. Corollary for the ledger: an unexplained positive gap between the key delta and the body sum is a lost body, not rounding, and it is named rather than absorbed. RS-20260809g-device-cross §10. FIRED AGAIN S171, at $0.062188150: E-20260812i's first run_hands.py was launched in the foreground, killed by the harness at 120 s, and had already billed at least one translation whose response was never written. The remedy is unchanged and was not applied. RS-20260812i-footing-channel §7.| S145 | S171 | | (blg) | A permission is not an application, and a design that prices two permitted devices is measuring their REACH unless it says otherwise. E-20260809g crossed two register devices as may use switches and defined its primary as the difference between what each bought. The pre-run critic's first three findings were one finding: each revision arm changes the English only where its own device has something to offer, so the two quantities are measured on different units, and a device looks weaker whenever its opportunities are rarer. Measured in that run: eleven of twenty-eight ⟨hand, site⟩ device opportunities produced no change at all, and the two devices' applied sets overlapped at fewer than half the sites. The primary was split before dispatch into a policy contrast (all sites, zeros included — what a translator gets across a book) and a conditional contrast (applied sites only, reported with its own n and not as a paired test), with neither permitted to be called the other. Rule: whenever arms differ by a permission rather than by an instruction that must be executed, register reach as its own estimand and say which of the two any headline number is. RS-20260809g-device-cross §2, §4.1, §6. | S145 | S145 — and S150, where the reach the note insists on registering became the whole result. E-20260810c gave two hands the located-idiom permission at 30 Japanese narration sentences: they used it at 4 of 60 hand-sites (0.067), against 32 of 60 (0.533) for the respelling permission, with 116 of 180 revision cells byte-identical to their own baseline. The manipulation check on the located arm returned −0.0167 and 0.0000 against +0.20 — an arm that is the baseline at 93% of hand-sites cannot be more located than it — and the REG primary was withheld. Because reach had been registered as its own estimand before dispatch, the run has a licensed finding (the permission is not taken) where it would otherwise have had only a withheld one. Extension, and it is cheap: compute reach in the gap between generation and rating. Here it was computable from the frozen artifacts alone, cost $0, and named the session's result before a single rating call went out. RS-20260810c-register-reach §5. | | (blh) | Derive the power gate from the exact test you will actually run, and state the smallest attainable P before dispatch. RS-20260808f reached P = 0.125 as the smallest value its test could return — four of eight sites were exact ties and significance was unreachable whatever the data did — and nobody knew until afterwards. E-20260809g registered the arithmetic in advance: a sign-flip permutation on k non-tied blocks bottoms out at 2/2^k, so k = 6 → 0.03125 (readable at α 0.05) and k = 5 → 0.0625 (not), and a cell with fewer than six non-tied sites is reported UNDERPOWERED rather than as a null. Both of its cells then failed that gate and the primary was withheld — which is the note working, not the note failing. Rule: a power gate stated as a round number is decoration; state it as the k at which the planned test can first reach your α, and register what a cell below it may and may not be said to show. RS-20260809g-device-cross §3, §4. | S145 | S145 — and S150, where the gate PASSED and the run was withheld anyway. E-20260810c re-derived k for a 30-site cell and registered 12 non-tied sites; it got 20, at a smallest attainable P of 1.9e-06. The power problem the note exists to catch was genuinely removed, and the comparison was still withheld — on the manipulation check. That is the note's real limit, worth carrying: a power gate answers could this design have said something, and never is the thing it measures the thing you meant. RS-20260810c-register-reach §4. | | (blp) | A symmetric agreement rule over free-form annotator spans must fix its unit of agreement in advance; clustering the spans themselves chains. E-20260810c replaced a spine-based site vote with a symmetric one on a critic's finding, and implemented it as single-link clustering under an overlap relation. Three annotators returned 312 stretches at three different grains — one sub-clausal, two whole-sentence — and transitivity collapsed them into 26 blobs, several spanning half a paragraph, in a criterion whose own exclusion 3 forbids a site longer than one sentence. The repair, registered before any generation call and with no English in existence: compute each annotator's character coverage, split the passage into its sentences, and admit a sentence at ≥ 2 of 3. Rule: when annotators are free to choose their own spans, agreement must be scored on a unit the design fixes — a sentence, a clause, a character — never on the spans themselves, because overlaps is not transitive and single-link clustering will make it so. E-20260810c/critic.md A21. | S150 | S150 | | (blq) | Before reusing a selection criterion on a new work, measure its admission rate on that work; a criterion built to find islands says nothing where the whole text is the island. E-20260808e's site criterion finds stretches below the neutral written register of their own language, and was built on Verga, Maupassant and Turgenev, where the low register is an island inside standard prose. Re-applied unchanged to Sōseki's 俺-narration in 「坊っちゃん」, it admitted 97 of 111 sentences — 0.874 — and 80% of the characters in the span. Nothing was wrong with the criterion; the text has no islands. The consequence had to be registered mid-run: the phrase marked narration sites was struck and the estimand narrowed to criterion-positive narration sentences of this span, because at 0.874 a "site" is not a place, it is the book. Rule: an admission rate is a property of the criterion crossed with the material, and it is one cheap count — take it before designing anything on top of the criterion, and treat a rate near 1 as a finding about the work rather than a supply of sites. RS-20260810c-register-reach §3. | S150 | S150 | | (blu) | Two panel slugs cannot have reasoning disabled at all, so note (bkw)'s remedy has an exception and the fallback is effort: low. x-ai/grok-4.5 and google/gemini-3.6-flash both return HTTP 400, "Reasoning is mandatory for this endpoint and cannot be disabled", to a body carrying "reasoning": {"enabled": false} — measured on both slugs at S153 against openai/gpt-5.6-terra, deepseek/deepseek-v4-pro and qwen/qwen3.7-max, which all accept it. The failure is at dispatch, not a dead body, so it costs nothing and is caught immediately; but a runner that applies note (bkw) uniformly will not run at all on those two seats. Rule: apply {"enabled": false} by default and carry a named exception set dispatched at {"effort": "low"}, which is what S151 recovered on. E-20260810w-recurrence-census/run.py. | S153 | — | | (blv) | A substring gate over an OCR witness fires on the scan, not on the seat: de-hyphenate before gating. E-20260810w's F3 required every quoted expression to be present in the passage it came from, and fired twice on two different seats quoting Garnett correctly — the Internet Archive scan carries end-of-line soft hyphens inside words ("others of still greater conse- quence") and the seats had de-hyphenated. Had the gate been applied as written, two of three cells on one hand would have been voided as fabrication, on the strongest single piece of evidence in the run. Rule: any substring or exact-match gate over an OCR witness normalises word-internal hyphen-plus-whitespace on both sides first; the repair is a materials repair, applied identically to every arm and independent of any answer, and is declared as one. RS-20260810w §6. | S153 | — | | (blt) | A retry-on-exception path re-dispatches a request that may already have been billed, so a key-usage reconciliation should PREDICT its residual from the retry log rather than expect zero. E-20260810t reconciled to a residual of $0.0072064, not to zero, and the dispatch log explains it exactly: of seven retries, six were 429 (no body, no charge) and one was IncompleteRead(319 bytes read) — a response that began arriving, was billed upstream, never reached the client, and was re-dispatched and billed again. The figure is one source-present call at the top of the observed range. A run that reports "delta exact to zero" as a hygiene claim will read a truncated-read retry as a defect, and a run that reports the residual without attributing it has not reconciled. Fires at: every cost reconciliation on a runner that retries on exception. Remedy: count the non-429 retries and name them beside the residual. | S152 | — | | (bmt) | A structural signal in a file is a candidate list, not a count — confirm it against the source before publishing a figure from it. Note (bmm) counted the empty speech-heads in ARM-dakghar's copy-text file and published "section ৩ has sixteen" missing speeches; it was carried forward by the arm page, the register and NEXT.md for five sessions. The whole-span refetch at S170 found fourteen. The other two are empty in the print as well — a speech-head above a centred stage direction, which the file correctly carries as its own unit — so the signal was detecting layout, not loss. The direction of the error is the benign one here, but nothing about the method guarantees that: the same test would have under-counted had the extraction dropped a speech whose head it also dropped. Fires at: any figure derived from a structural proxy in a file — empty fields, unmatched delimiters, count mismatches — that is quoted before the source has been consulted. Remedy: state such a figure as at most N candidates until the source has been read at each one, and let the confirmed count be the first thing the next session's gate produces. dakghar/collation.md §7A, ARM-dakghar S170 log. | S170 | S170 | | (bmu) | An assertion inherited from a state page is not evidence, and a frozen artifact must not repeat one without checking it against the source. register.md §Unresolved said at span B that nobody addresses Amal as তুই in sections ১–২ of «ডাকঘর». It is false — the Headman does, at [190], in a speech this project had already translated and already argued about (note (bml) turned on the same speech). The false premise then propagated into the span-C translator's log, D40, which was written and frozen before the census that caught it, and it took an erratum (D51) to remove. The register is a hand-off convenience; it had acquired the standing of a measurement simply by being written down. Fires at: every frozen log, design or result that repeats a fact from an arm page, a register, NEXT.md or a prior result rather than from the artifact. Remedy: any inherited claim that a frozen artifact restates must be re-derived from the source in the same session, and cited to the source rather than to the state page; where the re-derivation is cheap — here it was one census over 430 units — run it before the freeze, not after. RS-20260812h-dakghar-grade §2, T-dakghar-R05-v1 D51. | S170 | S170 | | (bmr) | A "licensed" or correspondence rate for ONE published hand must be reported per copy-text, or the page must say it was not — on this hand the same statistic is 0.8723 and 0.1034 depending only on which printed volume the transcription came from. E-20260812g censused Balmont's Poe across seven tales whose Wikisource transcriptions name four different printed volumes (Скорпион 1901 vol. I, Собр. соч. 1906 vol. VII, Скорпион 1906, Скорпион 1911 vol. III). Pooled licensed(RU) = 0.4414 — a perfectly quotable number that is an average over two populations differing by a factor of 8.4. And it is not simply lost markup: STRESS retention is 1.000 in both volumes, and an independent re-keying of one tale on az.lib.ru carries 27 of the 28 body italics the Wikisource text has, so the unlicensed marks are in both electronic witnesses. Either the translator's practice changed between volumes or their compositors differed, and a pooled figure hides the question. Fires at: every census, count or correspondence rate computed over a single translator whose texts were gathered from more than one edition or more than one electronic route — which is most of what a free corpus gives you. Remedy: record the declared copy-text per text at build time, report the statistic per copy-text as well as pooled, and where the split is large treat the pooled figure as uninterpretable rather than as the headline. Applies retrospectively as a caution, not a correction: prior single-hand rates on this shelf were not computed this way. RS-20260812g-foreign-italic §5, framework/v0.2 §7.8a. | S169 | S169 | | (bms) | Put a same-hand-different-text reference cell in every corpus build and read it before the corpus is used — it catches page furniture that inspection does not. E-20260812g built seven Russian texts from Wikisource and looked right: sane paragraph counts, balanced delimiters, plausible span totals. The defect surfaced only when a reference cell compared the same translator on two DIFFERENT tales and returned a 64-token identical run, which no two translations of different texts can have. It was a category footer. Repairing that exposed three more classes of furniture, all of which had been counted as the translator's marks: {{right}} epigraph blocks, source-edition lines, and — worst — the editor's notes section, which glosses the source's foreign phrases IN ITALICS, i.e. exactly the spans the run was counting. One page headed that section === КОММЕНТАРИИ === in capitals and a case-sensitive rule let five of its glosses through, including one of Poe's own French sentences. Span totals moved 128 → 118 → 112 across the repairs, all before any correspondence call. Fires at: every corpus built from a wiki, an e-text library or any page that carries apparatus around the text. Remedy: build the reference cell into the corpus script, not into the analysis — one call comparing the same author or translator across two of the corpus's own texts — and require a near-zero long-run before the corpus is used; then assert that the admitted span count equals the body span count per text, which is the check that stayed green afterwards. Kin to (bmm), which says the same about silently SHORT extractions; this is the silently LONG case. RS-20260812g-foreign-italic §6.5. | S169 | S169 | | (bmq) | Before a design classifies sites by the source's effect depends on something the target language lacks, measure whether two readers of the source agree on that classification at all — on this material they agreed at ZERO sites of fifteen. E-20260812f cut a Bulgarian comic chapter into 15 segments and asked two source-side seats, holistically and with the feature checklist struck by its pre-run critic, whether each segment's effect on a Bulgarian reader depends on something English does not have. 0 both-yes, 12 both-no, 3 split — and the three splits go in different directions, each seat naming a different segment as the untranslatable one. The segment the translator's own frozen log had spent two entries on — «Санким честна ли е? … изведнъж гут моргин», a Turkish sneer and a mangled German phrase — came back carriable from both. This is the third time this project has turned a judgment about untranslatability into a yes/no coding and the third time it has failed to converge (RS-20260804b §5 at five of seven loci; RS-20260804h's yardstick stage; here) — and the first time the category was the translator's own rather than a critic's. Fires at: any design that strata-fies sites by difficulty, untranslatability, "where the translator faced a real choice", or any near relative, from the source side. Remedy: measure the classification's inter-reader agreement in a cheap first stage and size the design to what survives; if nothing survives, the design's primary is dead before dispatch and the classification is the finding. Sibling of (bmo), which says the same about assumed floors. RS-20260812f-affect-unprompted §4. | S168 | S168 | | (bml) | Before buying a measurement, ask what the frozen artifact already answers by inspection; if the primary's outcome is legible in the text, cancel that part of the run and report the inspection. E-20260812g reached its second frozen design with twelve translation calls and two blind-coder calls in the pre-flight. Its critic's finding 16: "the supplied frozen item text already visibly contains 'boy' at [190]; absent a very unusual coding decision, the outcome is apparent before dispatch." True — the primary was a sealed self-report about three speeches, and one of the three visibly contradicted it. The calls would have bought a majority vote on a fact anyone can read, and would have dressed a count of three as a measurement. Fires at: any design whose primary is a property of text this project already holds frozen. Remedy: before the pre-flight is written, state which predictions could be settled by reading the artifact, settle those by reading it, and let the run buy only what reading cannot give. RS-20260812c-grade-shift §6. | S165 | S165 | | (bmm) | A text extracted into this repository can be silently SHORT, and neither the image gate nor the index lint can see it: diff the extraction against the upstream source before translating. ARM-dakghar's copy-text file was missing ten whole speeches of span B — 101 Bengali words, 4.7% of the span — one at the foot of each of ten pages, present in every case in the upstream Wikisource page text and lost by this project's own S160 extraction. The only trace was an empty speech-head, which no tool reads; the arm's declared page-image gate would have compared images against a text that had already dropped the lines. Section ৩ was forecast to have sixteen more; it had fourteen — see (bmt). Fires at: every span of every serial work whose source was extracted from a paginated upstream. Remedy: the gate runs in two stages and in this order — extraction against the upstream page text first, then the reconciled text against the page images — and the word count of the restored span is compared with what the arm page declared. RS-20260812c-grade-shift §3, dakghar/collation.md §5, §7A. | S165 | S170 | | (bmn) | A copy-text set from a collected edition decades after first publication is a signal to go looking for the earlier witness, not a copy-text to accept — and the looking happens BEFORE translating. T-shukuhai-R06-v1 was frozen declaring one witness and asserting that no second freely reachable text existed. False: 『道理』(春陽堂, 1921), the story's first appearance in book form, is on NDL Digital Collections under a Public Domain Mark as page images, and the Aozora card names its own 底本 — a 1960 collected edition — on its face. The witness was found by a contamination search run after the translation and the whole experiment had been frozen and run. The late collation found five divergences, one of them a change of sense ([11] 興味 → 意味) landing on the exact sentence the translator's log D12 had spent an entry adjudicating between two readings of a word the author may not have written; none of the five fell on a measured site, which was luck and is not a method. Fires at: every translation limb whose copy-text is an e-text set from a modern collected edition. Remedy: read the e-text's own 底本 statement as step 1 of the copy-text gate, and search for the earliest reachable witness before the draft is begun — the same gate S160/S165 ran on 「ডাকঘর」 and S161 on 「Kaşağı」, run for the right reason and at the right time. RS-20260812d-slot-or-carrier §10, shukuhai/collation.md. | S166 | S166 | | (bmo) | temperature 0.0 is not determinism, and a floor that assumes it is will fail — measure what one hand's own re-rendering moves before setting any detection bar. E-20260812d registered G3 at ≤ 0.60: the same hand, the same passage, two calls at temperature zero, rated 0–3 for how much the two English versions differ. It came back at 0.8333 and withheld every primary in the run. None of the six pairs was byte-identical; their similarity ran 0.0229 to 0.8511 — one pair was effectively a different translation. The same arbiters on byte-identical English scored 0 at 12 of 12 (post-hoc, unregistered), so none of the floor is arbiter false-alarm and all of it is the hand. Fires at: any design whose statistic is a rated difference between two generated texts, which is most of this project's arrival, carriage and recovery measurements. Remedy: measure the hand-variation floor in a cheap first stage, as E-20260812d already measured source-side visibility before selecting windows, and set the detection bar above the measured figure rather than at an assumed one; report every cell as a difference from that floor, not as a raw mean. Corollary: prior published arrival numbers whose floor was assumed are not thereby wrong, but their smallest reportable effect is unknown. RS-20260812d-slot-or-carrier §3. | S166 | S166 | | (bmp) | One dispatch process per output directory, and a runner you mean to stop must be PROVEN dead before the replacement starts — a pkill that reports nothing is not a proof. E-20260812e switched a seat's reasoning configuration mid-run, killed the runner, and started a replacement. The first runner did not die. Both dispatched against the same runs/ directory for the rest of the run, racing on the same (pair, seat) files: 90 of 108 bodies on that seat came back at the OLD configuration and 18 at the new one, and the defect was invisible in the costs and in the answers — it was caught only by auditing the request bodies file by file for the parameter that had been changed. Where both processes wrote the same slot, both were billed and only the later writer's record survived: $0.407307700 with no surviving record, plus $0.074160000 of bodies deleted in the repair — 30.6% of the run's spend, buying nothing. Fires at: any session that restarts, resumes, or reconfigures a runner that writes into a directory keyed by item. Remedy, three parts: (i) confirm the process is gone (ps returning empty, not pkill returning quietly) before starting anything that writes the same directory; (ii) make the runner record its configuration in a lock or manifest file and refuse to start if a different one is live; (iii) when a configuration changes mid-run, move the superseded bodies aside — never delete them — because the raw output is the record and the accidental duplicate is a free robustness check. The 18 deleted bodies here destroyed exactly that check. RS-20260812e-dose §7.1. [FIRED S168, used correctly, and remedy (iii) needs one more clause. The runner wrote runs/_config.json on its first call and refused to start on a changed configuration; three deliberate reconfigurations went through a --reconfigure flag that preserved the previous configuration in a _history list; every kill was verified by ps returning empty rather than by pkill returning quietly. The part that failed was the move. Superseded bodies were moved into ONE FLAT directory on two separate occasions, and the second move overwrote same-named bodies from the first — 63 moved, 51 survive — destroying exactly the cost records the residual reconciliation needed. Clause: move superseded bodies into a directory named for the superseded configuration, never into one shared bin.] | S167 | S168 | | (bmg) | When a critic amendment narrows the question a control is scored against, re-check every existing control against the narrowed question before dispatch — two amendments accepted separately in one sitting can be jointly incoherent. E-20260811f accepted the critic's SERIOUS 3 (add a control on the right axis) and its ADVISORY 6 (narrow the arbiter question from how they stand to each other to the social relationship) in the same adjudication. The narrowing made the two pre-existing controls — a change of referent and a change of tense — inapplicable, because neither changes a social relationship; G2a duly failed at 0.2500 and withheld a primary that the new control (G2b, 0.8333) had already licensed. Nothing in the disposition process compares amendments with each other. RS-20260811f-dakghar-address §4. | S160 | | (bng) | An open-ended enumeration is the one output shape whose cap cannot be guessed: if the answer is a LIST whose length is set by the input, either bound the output form or shard the call — do not pick a number. E-20260813g's LAT call asked one seat to classify 758 word types by etymological origin and return the R-list and the O-list, on the reasoning that the lists would be short. Both attempts returned finish_reason: length at a 6,000 cap with effort pinned low, the second returning zero characters, and $0.076043850 — 34.7% of the session's spend — bought nothing. F6 struck the measure, F3 reduced the axis to one measure, and F1 then withheld half the run's registered predictions: one mis-sized cap cost an entire axis. Note (bmb)'s remedy (pin the effort and size the cap to the pinned effort's observed output) could not be applied here, because on a new task shape there is no observed output — which is exactly when this note fires. The remedies, in order of preference: (a) ask for a closed output form whose size is fixed by the input rather than by the answer (a code per item, a fixed-width table, counts), (b) shard the list across several calls with a bounded slice each, (c) if neither is possible, price a pilot on a tenth of the list before dispatching the whole. The sibling calls in the same run asked for closed-form output over the same six texts and returned 618 and 550 characters at a 4,000 cap. | S178, 2026-08-13 | fires at dispatch design, on any call whose output length is a function of the input's length. S183 — fired and NOT repaired: P4 truncated mid-finding-10 at a 6,000 cap and the verdict line never arrived; NEXT.md recorded that a second $0.11 would have bought a verdict word and it was not spent. S184 — fired, and repaired, and the repair is not the one this note prescribes for a critic: raising the cap to 8,000 did not work — the seat spent 6,085 of the 8,000 on hidden reasoning and truncated in the same place. See note (bnr), which carves the pre-run critic out of remedy (a)/(b)/(c) and prescribes a continuation call instead. | | (bnr) | A pre-run critic is the one open-ended enumeration whose cap must NOT be raised — budget a CONTINUATION CALL instead. Note (bng)'s three remedies all assume the output can be bounded or sharded; a critic's cannot, because bounding the form is exactly what stops it finding the thing nobody anticipated. Raising the cap does not work either: S183 truncated P4 at 6,000, S184 raised it to 8,000 and truncated in the same place, the seat having spent 6,085 of the 8,000 on hidden reasoning — the visible budget was 1,915 tokens at both caps and the cap is not a bound on a reasoning seat. What worked, at $0.035: a second call carrying the design and the truncated critique, asking only for the unfinished finding, any further findings, and the verdict line. It returned finding F-G complete, five further findings and the verdict — and two of the five were accepted amendments, so the continuation was not a formality. Fires at: every pre-run critic dispatch. Remedy: size the first call at whatever is affordable, expect truncation, and hold ~20% of the critic's budget for a continuation; never let a run proceed on a critic that did not return a verdict line. RS-20260814f-carriage-elevation-2 §9. | S184 | S184 | | (bny) | When a reasoning seat returns ZERO characters, reasoning.max_tokens is the remedy and reasoning.effort is not — and the two may not be sent together. E-20260815's pre-run critic (P4 moonshotai/kimi-k3) was dispatched at max_tokens 6,000 with reasoning: {"effort": "low"} — the pin note (bnk) prescribes — and returned finish_reason: length with content of zero characters, the whole cap spent on hidden reasoning. $0.100636, 31.3% of the run, for nothing. Note (bnr)'s remedy is a continuation call, and it was inapplicable: a continuation needs a truncated critique to continue from and there was none. The retry that worked, at $0.075334, sent reasoning: {"max_tokens": 2000} with max_tokens 9,000 and returned nine findings and the verdict line complete. Sending both keys is HTTP 400 ("Only one of reasoning.effort and reasoning.max_tokens can be specified") — free, but it costs a round trip if discovered at dispatch. Fires at: every first call to a seat that reasons, and hardest on the critic, whose output cannot be bounded by shape. Remedy, which supersedes (bnk)'s pin for this case and closes the gap (bnr) leaves: on any call whose value is the written answer, send an explicit reasoning cap — not an effort pin — sized to a fraction of a total that leaves the answer room the reasoning cannot take; keep (bnr)'s continuation budget for the case where the answer starts and stops, and use this note where the answer never starts at all. RS-20260815-register-room §10. | S188 | S188 | | (bns) | A non-zero reconciliation residual is not evidence of an unrecorded call until a SECOND key read, with no calls in between, has been taken. S184 closed with per-request sum $0.263575750 against a key delta of $0.381083810 — an unattributed $0.117508060, 45% of the session's own spend, which on every prior session's convention reads as a defect to hunt. A second GET /api/v1/key immediately afterwards, with nothing dispatched between the two reads, returned a figure $0.001734720 higher: the key was still settling. NEXT.md had already observed the behaviour once in prose (S181 $1.109046 plus $0.265127 settling late) without a way to test for it. Fires at: every session close that reconciles. Remedy: when the residual is non-zero, take the second read before writing anything about it, and report the movement-with-no-calls figure beside the residual; CLAUDE.md's rule that per-request cost is primary and the delta is the sanity check is what is ledgered either way. The check costs one curl and distinguishes settlement lag from a call nobody recorded — which are different defects with different repairs. RS-20260814f-carriage-elevation-2 §9. | S184 | S184 |

| (bnz) | A call in flight when you kill a runner is billed and lost — note (bnx)'s append-and-resume protects everything that RETURNED, not the one in the air. E-20260815b's judging runner was stopped four times: twice by a harness wall clock and twice deliberately, to hold back a gate and protect the write-up. (bnx) worked exactly as written — 110 of 110 cells survived every interruption and not one had to be re-bought. The residual was $0.004922400, 0.7% of the spend, and note (bns)'s second key read returned a figure identical to the first, so it is not settlement lag: it is the sixth QD request, in flight at the moment of the kill, billed by the provider and never written to disk. Fires at: every session that stops a dispatch loop by signal rather than by letting it finish — which, on a long serial judging run, is most of them. Remedy: none is worth building; the figure is one call's cost and the repair (a shutdown flag the loop checks between calls) would trade a bounded, known loss for a runner that cannot be stopped promptly. What this note is for is the reconciliation: a residual of roughly one call's cost, after a kill, needs no hunt — record it as the kill's price, name the stage it belongs to, and do not spend a session looking for a defect that is a signal arriving mid-request. RS-20260815b-fluent-carriage §9, config/budget.md S189. | S189 | S189 | | (boa) | A verifier proves the arms differ where they should and nowhere else; it cannot prove they say the same thing — put the content-parity control BEFORE the judging calls, not after them. E-20260815b built three arms whose differences were mechanically exact: 342 pre-run checks, 0 failures, 8 of 8 mutations caught, asserting that the carriage manipulation and the oddity manipulation touched disjoint paragraph sets and nothing else. Every one of those checks passed and the run's carriage half is still unreadable, because an independent seat — after catching 3 of 3 planted errors — returned SAME on 0 of 7 segments. Three of the seven declared form-only device classes turned out not to be: a suspension cannot be flattened without inventing its completion; a reduplication reads as a different claim; varying a held keyword changes the degree asserted. Two of thirty-one oddity edits were also propositional defects the author did not catch — one reversed an agency — and no mechanical check could have seen either, because no check tests meaning. The parity control was dispatched last, so it cost the whole run instead of one call. Fires at: every design whose primary contrasts two texts asserted to differ in form only — which is this project's commonest shape. Remedy: dispatch the parity control first, on the frozen arms, as a gate; a design that cannot pass it has no primary and should learn that for one call's price. Corollary that travels further than this run: R14's operator table has never been put to a parity control of this kind, and RS-20260808d's sixteen-device flattening is the figure most exposed to it. RS-20260815b-fluent-carriage §6. | S189 | S189 |

| (bob) | A "first N by rule" locus selection is only as good as the enumeration behind it, and an enumeration is not an enumeration until its exclusions are published. E-20260815's translator's log declared eight PATTERNED loci as the first eight in copy-text order under a two-part criterion, and named a ninth as the last. Both claims were false: at least two further criterion-(b) loci exist in the same 1,036 words, one of them earlier in the text than the second locus taken, so "the first eight" never described the set. Nothing was wrong with the rule; what was missing was the rejection list. The pre-run critic did not find this directly — it found that the domain pooled sound with grammar (BLOCKING 1), and acting on that finding forced a re-enumeration, which is what exposed it. The repair that works is cheap and was applied the same session: the small domain (SOUND, 4 loci) is claimed exhaustive and prints every candidate it rejected, with the reason; the large one (COGNATE, 5) drops the completeness claim and says so. A reader can now test the first and cannot be misled by the second. Fires at: every design that selects loci by "all X, take the first N" — which is this project's standard shape for census work, used at E-20260814 and E-20260814g before this. Remedy: a selection rule that claims exhaustiveness must publish its exclusion list in the frozen design; where the list would be long, drop the exhaustiveness claim instead of asserting it untested. Corollary: an author enumerating figures in his own translation is the worst-placed reader for the ones he did not notice while translating — the misses here are all in sentences the log discusses for other reasons. RS-20260815-supplied-sound-confirm §9.5, erratum E3 on T-alf-layla-R05-v1. | S190 | S190 | | (boc) | A confirmatory primary registered as a strict inequality can PASS while the effect it stands for collapses — register a magnitude, or the replication cannot fail. E-20260815 was built to confirm RS-20260814g's Burton supplies English sound where the Arabic has none, at 4 of 8 plain loci. Its pre-run critic rightly demoted the rate threshold (0.375) as circular — it was mined from the run being confirmed — and the primary became the between-hand contrast Burton > Lane. That primary HOLDS, at 1 of 8 against 0 of 8, and the effect fell from 0.500 to 0.125. One cell satisfies a strict inequality; on n = 8 with a comparator at floor, the primary was close to unfalsifiable in the direction that mattered, and the honest reading had to be assembled afterwards from the secondary figures and from where the surviving cell sat (in the confounded speech stratum, with the unconfounded slice at 0–0). Fires at: every replication whose comparator is at or near zero, which is common here because the null hand is usually the flat one. Remedy: when a primary is a contrast against a floor, register both a direction and a magnitude — e.g. > Lane and within 0.25 of the original point estimate — and pre-commit to reporting the replication as failed if the magnitude misses even where the direction holds. The circularity the critic named is real and the answer is not to drop the magnitude but to source it from outside the run being confirmed, or to declare the study a direction-only replication in its title. RS-20260815-supplied-sound-confirm §5. | S190 | S190 | | (boe) | A re-dispatch ladder that overwrites the failed response's usage under-reports the run's cost by exactly the attempts it replaced — accumulate every attempt, not the one that worked. E-20260816 closed with per-request costs summing to $1.022045450 against a key delta of $1.047809700, the key reading $0.025764250 high, and note (bns)'s second read returned an identical figure, so it was not settlement lag. The cause is one line of the runner: on an unusable reply it re-calls and rebinds r, so only the second body's usage is written to run.jsonl. Ten bodies were re-dispatched and every one of them had failed by truncation, which means the discarded attempt had burned its whole max_tokens and was the most expensive body of its stage, not an average one: 9 P1 at cap + 1 P2 at cap ≈ $0.029 against a measured residual of $0.0258. Fires at: every runner in this project — they all share this dispatch shape, and the defect has been invisible because a residual of two or three cents reads as settlement. Remedy, one line: keep a per-body list of attempt costs and write their sum, with the attempt count, so the ledger is the money spent rather than the money spent usefully. Distinguish from (bnz), which is about a call in flight when a runner is killed — that one returned nothing and is unrecoverable; this one returned, was billed, and was discarded on purpose by code that meant to keep the good answer. RS-20260816-answering-figure §5; config/budget.md S191. | S191 | S191 | | (bof) | The key-usage delta is not a cross-check on this project's spend, and must not be reported as one. E-20260815c closed with per-request costs of $0.542027300 against a key delta of $1.008081890. This is not settlement lag (note (bns)) and it is not note (boe)'s runner leak, which is fixed in that runner and ledgered inside the total: two key reads twenty seconds apart, with no request from the session in between, differ by $0.005285, and over a 60.9-second idle window the figure moves $0.013748 — about $0.81 an hour while nothing is being dispatched. Something outside these sessions is drawing on the same key continuously. That retro-explains the movements S190 ($0.327797), S191 ($0.253922) and S192 ($0.306086) each opened on and each recorded as unexplained, and it means every past reconciliation that "closed exactly" did so because the drift happened to be small over that window, not because the delta is an instrument. Fires at: every §Spend section and every config/budget.md reconciliation. Remedy: per-request billed cost from "usage": {"include": true} is the ledger and is exact; a key read may be recorded as provenance but no residual against it is a defect to be explained, and no session should spend its §Spend section explaining one. Supersedes the reconciliation practice of notes (bns) and (boe) without touching either note's own finding. RS-20260815c-footing-direction §9. | S192 | S221 | | (bog) | An instrument control for "what does the WORDING mark" must hold the narrated EVENTS flat, or it is not a control. E-20260815c's negative control removed every rank word from a scene and kept the scene: a household standing back, a younger woman curtsying, the hostess coming forward to thank the visitor. All three seats named the same person as socially highest in the unmarked version as in the marked one, and all three quoted the behaviour, not a word — "the household stood back and the hostess came forward and thanked her". The gate fired and every registered primary of the run was withheld. A second pair, same three women, no deferential behaviour in either member, separated at the ends of the scale: REL 1, 1, 1 with WHO E, E, E unmarked, against REL 7, 6, 6 with WHO B, B, B marked. Fires at: any design asking a judge what the wording does while showing them a story. The trap is that the natural way to build a deference control is to write a deference scene. Remedy: build the control on an incident in which nobody defers to anybody — people sitting and talking — and vary only the words; then, separately, run the deferential-events version to measure how much the events swamp. The two together are the measurement, and in this case they gave the run its only unwithheld finding: with events flat, four rank words are worth 5.33 scale points and a named person; with events deferential, the same four words are worth 2.33 points and change the named person not at all. RS-20260815c-footing-direction §2–§3. | S192 | S192 | | (bph) | A content check that a response "looks right" does not catch a response that was cut off — assert finish_reason explicitly, and set the reasoning budget well below the content cap. E-20260815d set reasoning: {"max_tokens": 400–600} inside a content max_tokens of 400–500 for its short stages, so the reasoning budget consumed the whole allowance and 21 of 120 bodies came back with finish_reason: "length", cut off mid-sentence. Every one passed the run's content check, because a truncated Japanese sentence still contains four Japanese characters, and one of them (「「いや、決して間違いなどあってはならない。それは必ずやそこにある」) was missing the polite ending that would have changed its coded value. Note (bny)'s working shape is a ratio — 2,000 inside 9,000 — and this session copied the numbers without the ratio. Fires at: every dispatch with a max_tokens under about 2,000, which is every per-item stage. Remedy, both halves: (i) keep reasoning.max_tokens at roughly a fifth of the content cap and never at or above it; (ii) make finish_reason != "length" a kept-ness condition in the runner and an assertion in the verifier, so a truncated body is re-bought automatically and can never reach the analysis. Re-purchase is selected on finish_reason, never on what the body says. RS-20260815d-supplied-footing §8. | S193 | S198 — fired four more times in one run, on two seats and two stages; see (bps), which is the advance check that makes it preventable. | | (bpv) | max_tokens must be strictly GREATER than the reasoning cap, and two seats fail in opposite directions when it is not — one with an HTTP 400, one with a silent empty body. E-20260816f sent max_tokens: 900 with reasoning: {max_tokens: 1500}. qwen/qwen3.7-max refused the request outright — HTTP 400, "max_completion_tokens [200] must be greater than thinking_budget [400]" — which is the benign failure, because it costs nothing and is impossible to miss. openai/gpt-5.6-terra accepted it, billed it, and returned finish_reason: length with an EMPTY content string, twice, at 600/1500 and again at 1400/700 on a short prompt. Fires at: every dispatch that sets a reasoning budget, which by (bph) and (bps) is every dispatch. Remedy: make the two caps a ratio and assert it in the runner before the first call — this run settled at 1,500/900 for the long bodies and 1,400/700 for the short ones, and all four seats were clean. Kin to (bph), which is the truncation seen from the content end, and (bps), which is the per-seat probe; this note is the one line of arithmetic that makes both cheap. RS-20260816f-night-seam §4, config/budget.md S200. | S200 | S200 | | (bpx) | temperature: 0 does not make a second call a second reader, and a design that counts replicates as independent bodies inflates its n by the replicate factor. E-20260816g's v1 planned three temperature-0 replicates per seat and treated the bodies as independent observations; the pre-run critic's BLOCKING 3 named it — "repeated deterministic calls are pseudoreplicates, while model families are fixed, heterogeneous instruments rather than a random sample of readers". Served models are not bit-deterministic, so the bodies do differ; the point is that the variation is not a sample from anything the design has defined, and the four or five seats are the instruments. Fires at: every design that buys more than one body from the same seat on the same stimulus. Remedy, two lines: set temperature above 0 so replicates are genuine samples, and make the seat, not the body, the unit of the estimator — average within a seat first, then across seats, so a seat contributes once however many bodies it supplied. E-20260816g did both; its Δ is the mean over three seats of each seat's own FULL-minus-FLAT, and its two permutation nulls are over labels and loci, never over bodies. Flagged, not adjudicated: earlier runs whose exact tests permute bodies — RS-20260816f's placement P over 3,432 relabelings is the nearest — bought temperature-0 replicates, and whether that inflates their nulls is a question for whoever next needs those figures, not a claim here that they are wrong. | S201 | S201 | | (bpi) | Two frontier models translating the same passage independently are not two observations, and the dependence check will say so if it is pointed at them. E-20260815d measured openai/gpt-5.6-terra against google/gemini-3.6-flash on 501 words rendered EN→JA from identical prompts, no shared context: 69 shared 7-grams, 35 twelve-grams, 26 fifteen-grams, longest contiguous run 33 characters — more long n-grams and a longer run than the run's known dependent pair, a 1930 published translation against its own 2021 revision (80 / 23 / 11, run 18). Both models were clean against the published hand (runs of 8 and 11). Fires at: any design whose criterion is "k of n independent model hands", which this project has written many times. Remedy: run dependence_check_cjk.py / dependence_check.py between the model hands, not only between each hand and the human comparator, and report a "2 of 2 hands" criterion as what the measurement says it is. This does not invalidate model panels used as judges of distinct objects; it bites when the models are being counted as independent producers. RS-20260815d-supplied-footing §7.1, A-mikami-silver-blaze §Contamination. | S193 | S193 |

| (bpj) | A form-only device set is a hypothesis about meaning, and only an independent reader can test it — this project has now had five of seven classes fail that test on one story. E-20260815b declared seven Class A "form-only" device classes on Makino and its parity control convicted three: a suspension cannot be flattened without inventing its completion, a reduplication asserts more than a single token, and varying a held keyword moves the degree asserted. E-20260815e dropped those three and the same control convicted a fourth — folding a standalone attribution into a dialogue tag forced the flattener to invent an addressee at one site and a speech verb at another, because the source attributes by juxtaposition. A blind carriage audit then refused a fifth: the reduplicated mimetic fails as carriage at 5 of 6 sites, against the translating log's claim that English sound-symbolic verbs do the work and "cost nothing". Only utterance-first, attribution-after transferred entire. Fires at: any design whose arms are asserted to differ in form only — this project's commonest shape — and at any R14 table. Remedy: declare the class set, then have an independent reader price it before the figures are read; report a translator's carriage census as a claim (this one was priced at 25 of 33) and never as a measurement. Note (boa) says when to run the control; this one says what it keeps finding. RS-20260815e-fluent-carriage-2 §3–§4. | S194 | S194 | | (bpl) | When the lead is primed on a specific published rendering of the very phrase it must translate, write the primed solution down BY NAME and exclude it — an invisible contamination becomes a stated constraint, and the excluded thing stays checkable. D43 had to render the Nights' night-closing formula, a rhymed sentence repeated about a thousand times, and the lead had seen Burton's rhyme (day / say) before span A was written. Every rhyme the desk then produced had to be checked against a solution already in the room. What settled it was naming the excluded pair in the log, listing the three further candidates refused and why, and recording the Arabic end-rhyme as lost with a compensating figure at the same seam rather than reaching for the known answer. Fires at: any R05-style source-only rendering of a phrase from a heavily translated work — the formulae, the proverbs, the titles — and at any artifact carrying contamination: high. Remedy: name the primed rendering in the translator's log before choosing, exclude it explicitly, and publish the rejected candidates; a contamination: line that says only high leaves a reader unable to tell exclusion from imitation. Kin to the standing contamination rule in CLAUDE.md, which governs measurement; this governs the decision when a measurement is not what is at stake. T-alf-layla-R05-v1 D43, register.md V12. | S195 | S195 | | (bpm) | Counting a phrase with a literal-space regex under-counts every wrapped text, and the miss is silent and directional. E-20260815f first counted Burton's night formula at 25 by grep on the raw Project Gutenberg file, and at 33 over a whitespace-normalised token stream. The eight missing instances are the ones where the line break falls inside the phrase. Twenty-five was believed long enough to be written into a working note, and it is the retention figure — so the error would have published a translator as dropping a quarter of his boundaries. Fires at: every census over PG, Wikisource, Aozora or any hard-wrapped source, which is all of them. Remedy: normalise whitespace or tokenise before counting, and make the independent verifier use a \s+-joined pattern; where a count is load-bearing, have the two paths agree. RS-20260815f-night-formula §7.5, verify.py. | S195 | S195 | | (bpk) | A source-blind "which followed the original?" score is not evidence about carriage, and the same seats answer differently when they can see the original. E-20260815e is the first carriage manipulation here whose arms passed an independent content-parity control (SAME on 7 of 7, 5 of 5 planted errors caught). On it: an arm carrying nothing and merely badly written was chosen as the source-follower over a clean flattening in 20 of 21 cells; the cell where carriage and non-fluency point at opposite arms came back indeterminate for the second time in two builds (13/21 then 11/21, seats disagreeing both times); and the same pair with the original printed above it went to the carrying arm in 20 of 21, both reversing seats moving the same way (2/7 → 7/7, 3/7 → 7/7). Fires at: any use of perceived-source-carriage, and at any evaluation that asks a source-blind reader an attribution question. Remedy: score this sense with the source available, or report the figure as a fluency judgment; a source-blind carriage score may not be cited as evidence that anything was carried. Corollary for oddity arms: on a source whose content names its language, no English-side awkwardness could be built that an independent seat did not attribute to that language — 25 of 35 edits, and retiring the operator a critic named moved it not at all. Stilted collocation was the one operator that escaped (4 of 12 against nominalisation's 11 of 12, circumlocution's 7 of 8 and cleft's 3 of 3): if a design needs unlicensed markedness that does not leak the source language, make it lexical, not syntactic. RS-20260815e-fluent-carriage-2 §7–§8. | S194 | S194 | | (bpn) | If the primary is "does X do anything", DERIVE the second arm from the first — an independently written contrast arm cannot carry the attribution, and a critic shown both will say so before you spend a dollar finding out. E-20260816b first wrote its two arms separately: same declared device classes, same locus set, word cost matched to ±2. The pre-run critic returned NEEDS REDESIGN with the whole point as BLOCKING 1 — two independently written arms differ in propositions, member count and rhetorical form as well as in the manipulated property, so their difference is not attributable. The rebuild made the second arm a de-sounded twin of the first: same added material, same member count, same word count at every locus, token-identical except at 15 positions across thirteen loci, one at eleven of them, and a builder check that asserts it. The primary then fired at +0.3846 and could be read, because the seats' own free-text reasons name the changed word. Fires at: every design contrasting two versions of one text — this project's commonest shape, and the shape of R14, R31/R32, R34/R35 and every arm table since S133. Remedy: build arm B by editing arm A at declared slots, never by writing it; make the token-level diff a builder assertion with a hard maximum; and record per locus the material both arms add, so a residual rule-1 violation is common to both and cannot generate the effect. Corollary the same run establishes: exact synonymy does not exist, so a one-word substitution always changes sense a little — say so as a limit rather than claiming equivalence, and let the seats' reasons show what they answered on. RS-20260816b-invented-figure §3, §10.2; E-20260816b §11. | S196 | S196 | | (bpo) | A contrast arm denied the resource the real move uses measures the resource, not the move — and this project published such a reading and had to withdraw it one session later. RS-20260816 built its invention arm under a substitution only, no added words rule, measured it inaudible at 0 of 13, and wrote into framework/v0.2 §7.12 that ornament cannot be invented, with the length restriction named as a caveat. E-20260816b gave the same hand, the same loci and the same device classes the words an ornamentalist actually takes and the same instrument heard the source-inference at 9 of 13 — indistinguishable from genuine compensation at 8 of 13. The starved arm had measured the starvation. Fires at: every decoy, control or contrast arm built under a constraint the genuine move does not obey — length caps, substitution-only rules, single-site budgets. Remedy, in the order that works: measure the move's profile in the wild first, on continuous prose, and build the arm to that profile; here the translation limb rendered a whole fresh chapter under the permissive regime and counted 22 ornamented sites, +6.1% in length, +1.55 words per site before a single arm locus was written, and the arm was then funded above that rate so that a null would be conservative. Where the constraint cannot be lifted, the finding is about the constrained condition and the title must say so. RS-20260816b-invented-figure §7; T-kalila-nasik-R34-v1 §4. | S196 | S196 | | (bpp) | A control carried over from an earlier design must be re-validated against the NEW prompt, not merely against the old result. E-20260816c imported the four third-party Q1 instrument controls of E-20260815 verbatim, as RS-20260816b had, and they had been unanimous for three sittings. But Q1 had been rewritten for this design to say "Look only at the span marked ⟪ ⟫", and the control passages carry no marked span. CTL.POS came back N from P1 — reason: "N; no marked span is present" — the abort gate fired, and the run stopped after 12 bodies and $0.042105 spent on an ill-formed stimulus. The control was right, the prompt was right, and the pairing was wrong. Fires at: every design that reuses an instrument while changing the prompt around it, which is what reusing an instrument means. Remedy: the prompt's presuppositions are part of the cell contract and belong in the same assertion block as the cell counts — here one line, exactly one ⟪…⟫ pair in every dispatched string, now in cells.py, would have caught it before dispatch. RS-20260816c-checked-ornament §8. | S197 | S198 — FIRED AGAIN AND GENERALISES BEYOND CONTROLS TO MECHANICAL PARSE RULES. E-20260816d carried E-20260815d's accept-rule for a Japanese body verbatim — at least 4 CJK characters — into a census whose shortest turns are \"Uncle!\" and \"Nothing!\", whose correct Japanese is three characters. It rejected a perfectly good 「おじ様！」. Anything carried from an earlier design — a control, a prompt, a parser, a threshold, an accept-rule — is re-validated against the NEW material's extremes before dispatch, not just against its typical case. | | (bpq) | A stop-loss set EQUAL to the worst-case estimate is not a margin, it is the estimate again — and it destroys whatever the dispatch order put last. E-20260816c §11 built a worst case of $2.40 from max_tokens, correctly per note (abc), and then set the runner's ceiling to $2.40. Those being the same number, the run halts exactly when the estimate is right rather than when it is wrong, and the arms dispatched last — which carried a registered primary — are what a correct-but-tight estimate would have cost. Fires at: every run with a stop-loss. Remedy: put the stop-loss above the worst case, at whatever the day's headroom allows, because its job is to catch an estimate that is wrong — a routing surprise, a runaway cap, a re-dispatch storm. Where headroom will not allow that, dispatch the registered primary first and the descriptive arms last; and where the order is instead a gate order, as here, say in the design that the two criteria conflict and which one won. E-20260816c §10, §11. | S197 | S198 — FIRED AGAIN, FROM THE OTHER SIDE, AND THE RULE IS NOW TWO-SIDED. (bpq) said a stop-loss EQUAL to the worst case is not a margin. E-20260816d v1 over-corrected and set it BELOW the worst case ($1.25 against $1.45), and the pre-run critic's F13 caught what that means: the design licensed its own truncation — a healthy run could be halted mid-census and the halt would look like a defect in the data rather than in the plan. The correct structure is worst case < stop-loss < ceiling: the stop-loss fires only when something is actually wrong, and is still a margin. Where the true worst case will not fit under an affordable stop-loss, the design is reduced until it does — tighter caps, fewer cells — and the reduction is written down. v2 landed at $1.97 < $2.05 < $2.20. | | (bpr) | A gate that withholds a whole family of predictions should withhold only the ones it actually threatens — check, before freezing, which of them are between-group comparisons and which are within-item contrasts. E-20260816c gated P2, P3 and P4 on a manipulation check that the panel's labelling agreed with the design's. It failed, and all three were withheld — but only P2 is a between-stratum discrimination that needs the labels to be right. P3 and P4 are within-site contrasts (the same source locus, the same marked span, with and without the English device) and are well-posed under any labelling; the gate cost them for nothing, and the rule could not be relaxed afterwards without relaxing a rule after seeing it bind. Fires at: every design with a manipulation check in front of more than one prediction. Remedy: in the design, list under each prediction whether it needs the gate and why; gate only those. RS-20260816c-checked-ornament §10.3. | S197 | S197 | | (brt) | A max_tokens cap verified on one task shape does not transfer to another, and the failure is silent — an empty body billed at the cap. Note (brr) fixed qwen/qwen3.7-max (QR) at 3000 and recorded that at 3000 it ran 56 calls without a single failure. E-20260826c sent the same seat the same cap on a different task shape — read a 130-word passage, return a short JSON list — and 29 of its 68 calls came back with an empty content string and finish_reason: "length", 44% of the seat, while P1 and P2 at 2500 lost nothing. The seat spends the whole allowance on hidden reasoning when the visible answer is tiny, which is the opposite of the intuition that a short answer needs a small cap. Fires at: every reuse of a per-seat cap across designs, which is what a note like (brr) invites. Remedy: treat a cap as verified for the task shape it was measured on, and on any new shape either probe one call per seat before the fan-out or budget at twice the previous cap; and make finish_reason != "length" a re-dispatch condition in the runner, which (bph) already requires and which here converted a silent loss into a counted one. Kin to (bpp) — anything carried from an earlier design is re-validated against the new material. RS-20260826c-register-cost §3. | S225 | S229 | | (bru) | A quote-what-you-would-query instrument scores asymmetrically if the stored target is longer than the phrase a reader would naturally quote — score by LOCATION, not by string. E-20260826c froze its hit rule as the returned query contains the target string. On the seeded control SC4 all six seats quoted the planted defect — "three fish", "three" — and all six scored zero, because the stored target was the longer there were three fish. A one-word target (screwdriver) is hit by any quotation containing it; a four-word target is missed by every quotation shorter than itself, so the same rule is strict or lax depending on how the analyst happened to type the target. The defect inverted the run's instrument gate: 15 of 22 blatant seeds under the frozen rule, 22 of 22 under location scoring. Fires at: every design whose outcome is a free-text quotation matched against an expected string — this project's whole family of name the place where… instruments. Remedy: locate each returned quotation in the passage it came from and score overlap with the manipulated span (all 308 quotations in this run were found verbatim, so no judgment was needed); or, where that is not possible, store the minimal distinctive string. Report both scorings when the correction is made after the run, as RS-20260826c-register-cost §5 does. | S225 | S225 || (brv) | A locus pool built by hand will violate its own admission rule, and two independent critics will both find it — build the pool with a script and publish a disposition for EVERY candidate. E-20260827-declared-play froze 23 تجنیس loci hand-picked from the Persian under a written phonemic rule. Both pre-run seats returned BLOCKING on it independently: P2 showed two admitted loci that fail the stated distance test while a refused one passes it; P1 named a block (p041) that satisfies the rule and carries no disposition at all, which falsified the page's claim that the rule was applied once, block by block. The rebuild — enumerate_loci.py emits candidates under a rule fixed before it is run, adjudicate.py records a coded disposition for every candidate and fails if one is missing or names a candidate that does not exist — cost the lead four of its own favourite loci and added six it had missed or wrongly demoted, including one the hand-built page had excluded by name. Fires at: every design whose primary is what happens at the places where the source does X — this project's commonest shape. Remedy: generate mechanically, adjudicate in the open, publish the refusals beside the admissions, and state where lead judgment enters and what it cannot see. The blind spot is then a declared property of the rule rather than an undisclosed property of the analyst — and §6 of the result page is what one blind spot cost: a real locus sat in the control pool, and the declarer's one kept promise was scored against him. RS-20260827-declared-play §2, §6; critic-response.md §2. | S226 | S226 | | (brw) | A runner that retries a failed call must accumulate the cost of EVERY attempt, or the ledger under-counts by exactly the discarded ones. E-20260827-declared-play's run.py recorded only the last attempt's usage.cost per cell. 17 cells needed a second attempt and then a third at a repaired cap; the per-request sum came to $1.541667931 against a key-usage delta of $1.747847221, and the $0.206179290 gap is the discarded attempts, not drift — which is why this is the first session able to explain a divergence rather than record it (contrast note (bof) and the S225 row). Fires at: every runner with a retry loop, which is all of them since (bph). Remedy: sum costs across attempts inside the cell record, and keep a separate attempts count; where the defect is found after the fact, ledger the key delta, which is the larger and therefore the conservative figure. RS-20260827-declared-play §9. | S226 | S229 | | (brx) | At temperature 0 a malformed reply is REPRODUCIBLE, so the mandated single re-dispatch buys nothing — a formatting failure needs a tolerant parser, not a retry. E-20260827b's eight void cells are one defect: openai/gpt-5.6-terra returns {"choice":"A","why":"… “quoted phrase.} — an unterminated final string, because it opens a curly quote inside why and never closes either — and every one of the eight was re-dispatched once at double the cap and came back byte-identical, which is what temperature 0 is for. The choices were all plainly recoverable from the raw text. Fires at: every design whose F4 missingness rule is re-dispatch once, which is all of them since (bph); the retry rule was written for truncation at finish_reason: length, where the cap change does something, and it does nothing for a parse failure at finish_reason: stop. Remedy: branch the re-dispatch on finish_reason — retry a length, and for a stop that will not parse either repair the string (close an unterminated final value) or re-ask with a changed prompt; and where the repair is made after the run, report the registered figure and the repaired one side by side, as RS-20260827b-shown-or-told §7 does. | S227 | S232 |