Repository path: workshop/experiments/E-20260829-radif-hands/critic-response.md · rendered 2026-09-09

Pre-run critic pass — E-20260829-radif-hands

Two independent seats on two labs, on the design frozen at commit 31e6026b, before any other call was dispatched. P1 = openai/gpt-5.6-terra (OpenAI, $0.047860), P3 = x-ai/grok-4.5 (xAI, $0.044958). Raw bodies in runs/RS-20260829-radif-hands/raw/C1.json, C2.json. Note (brr) honoured: P2 was not used, the cap was 12000, and neither seat truncated.

Both returned NEEDS-REDESIGN. 25 findings, 9 of them BLOCKING. Twenty are accepted, three are accepted in part, and two remedies are overruled in writing. The design is amended to v2 below and the amendments were made before the classification stage was bought.

Accepted, and what changed

A1 — C1-1 and C2-1, both BLOCKING: the coder never checks that the repeated English tail renders the Persian radif. This is the strongest finding of the pass and it is right. As frozen, any repeated line-ending counted — an English monorhyme, a stock refrain, a pronoun. Amendment: a CORRESPONDS check is added and is now part of the coding. For every cell the coder returns carried, the result page prints the Persian radif, a literal gloss and the English tail side by side, and a cell whose tail is not a rendering of the radif is recoded not carried. The check is declared as what it is — a judgment, made in public, on nine printed pairs — not as machinery.

A2 — C2-4, BLOCKING: G2 made refutation conditional on agreement with the party who wrote the rule. Exactly backwards, and the sharpest finding in the pass. Amendment: lead-agreement is struck as a gate. The trust gate is now inter-seat agreement — at least 2 of 3 seats identical on every checklist bit — and lead-versus-seat disagreements are printed as a diagnostic side table with no power to withhold anything.

A3 — C2-3, BLOCKING: the seat schema did not collect every predicate the published cells turn on. Amendment: the seats are given a closed checklist with one bit for each predicate in §7.36.1, §7.40 and the S230 clause — part of speech; finite or not; preterite or not; transitive or not; bare copula or not; case particle or not; imperative with fronted complement or not; standing in ezāfe/genitive to the preceding word or not — plus an explicit UNSURE. Application is then a lookup.

A4 — C2-5 and C1-10, BLOCKING: the item set was not named. It could have been, because the census and the match were run before any money was spent. Amendment: the nine items are named in v2 §2a below and every gate is expressed over that list.

A5 — C1-2, C1-3 and C2-10, BLOCKING: coverage, and vacuous passes. Amendment: a coverage gate G6 is added — a cell with no matched item is reported NO-TEST, never as "not refuted". Coverage on the frozen list is: case particle 1 item, transitive finite verb 4, oblique pronoun in genitive 1, copula 2, adverbial particle 1. Every cell P1 and P3 speak to has support, and the support is small and is reported as small.

A6 — C1-4 and C2-6, BLOCKING/MAJOR: a denominator in printed lines. Accepted in part. Both of the coded books print two lines to the bayt, so B − 1 is the right bar for them and the hit counts are printed beside B for every cell. Amendment: Bell, who prints seven-line stanzas and is not matched item-by-item, is coded by a fraction rule instead — a tail on at least 45% of the poem's lines — and the rule used is named for each hand rather than assumed common.

A7 — C1-9 and C2-8, MAJOR: the question claimed more than the design can license. Accepted. The question is rewritten. It is not does the predictor sort carriages from failures; it is can a published hand supply a counter-instance to the impossibilities §7.36.1 and §7.40 state, with a descriptive carried-against-cell table beside it and no sorting claim.

A8 — C1-5 and C2-7, MAJOR: P2 compares two different corpora. Accepted. P2 is downgraded to descriptive, loses its gate, and carries its confound in the sentence that reports it: Bell's zero is a fact about her book, not a controlled comparison with Leaf's.

A9 — C1-6 and C2-12, MAJOR: Payne's role was unstated. Accepted. Amendment: Payne's item set is named — Brockhaus 8, 52, 79, 123, 151, 198, being the six of the nine that fall inside his volume 1 — and the other three are ABSENT, because volumes 2 and 3 are not freely reachable.

A10 — C1-11 and C2-9, MAJOR: G4's constructed positive was unfrozen and tunable. Accepted. Amendment: G4 is now two named, held-out sets: Bell's 32 poems must code not carried (a hand who never attempts the device), and Leaf's ode XXI and Payne's ode 52 must code carried; both are printed with their line-ends.

A11 — C1-12 and C2-11, MINOR: pricing, retries and tokenisation. Accepted. The dispatcher retries at most twice per call on transport failure and never re-buys a parsed body (note (brx)); the normalisation functions are the archived census.py and carriage.py, committed before dispatch; the price table is config/models.md as of this session.

A12 — C1-13, MINOR: unanchored premises. Accepted. "Payne is monorhymed" and "Bell promises nothing about form" are now supported by their own printed pages, quoted in the result.

Accepted in part

B1 — C2-2 and C1-8, BLOCKING/MAJOR: the census excludes bound radifs, so whole cells can go empty. The premise is right and the consequence is not what the critic thinks. A radif is by definition a repeated word; a repeated bound morpheme is part of the rhyme, and that is the standard definition, not this project's. What is accepted is the reporting duty: the exclusion is now stated as a limit on the frozen census, and the phenomenon it hides — a hand rendering a Persian bound suffix as a free English word, and so making a radif where the Persian has none — is reported as a finding rather than dropped, because it turned up in the P4 control.

B2 — C1-7, MAJOR: model seats are not a validity anchor. Accepted as a limitation and its remedy overruled below. The seats are a reproducibility and independence diagnostic, and the result page says so.

B3 — C1-1's remedy asks for position-by-position correspondence. Accepted in substance (A1) and narrowed: correspondence is checked at the level of the poem's tail, not at each of its nine positions, because a tail that stands at B − 1 positions is by construction the same tail at each of them.

Overruled, with the reason written

O1 — C1-7's remedy: "use qualified human annotation as the adjudication anchor". Overruled. The project has no human annotators — that has stood on its named-not-built list for twelve sessions — and inventing one would be worse than having none. What is done instead: the nine classifications are printed in full, in Persian, with the gloss, so that anybody who reads Persian can check them; and the claim they support is an existence refutation, which turns on classifications (را is a postposition; کرد is a finite past transitive; تو is a pronoun in ezāfe) that are not in dispute in any grammar of Persian.

O2 — C1-9's stronger remedy: "register a genuine discriminative test with a defined failure outcome". Overruled for this dispatch and routed to step 2 of the arm, which is what step 2 was constituted for. A discriminative test needs the association powered, and powering it needs Payne's two hundred odes matched to the source — a session's work by itself. Registering a test this dispatch cannot power would be the failure mode G5 was written to avoid.