Repository path: workshop/experiments/E-20260803c-occupied-slot/design.md · rendered 2026-09-09

E-20260803c — does R1 have an unstated precondition: that the slot is free?

ARM-framework-v01 step 2 (T5). Frozen before any seat was addressed, before the published English of this span was opened, and after T-kammacher-R04-v1 was committed with its log.

AMENDED 2026-08-03 after the pre-run critic returned NEEDS-REDESIGN with seven BLOCKING findings, all ten accepted (critic.md). Amendments A1–A8 are operative and are marked in place below. The largest is A2: the author-written repair arm is withdrawn from scoring and replaced by a mechanical one applied identically at every site, because the design as first frozen could not distinguish the slot is spent from the translator patched harder where patching was easier. The frozen translation, the twelve sites and the German are unchanged.

1. The question

framework/v0.1 carries one operational recommendation:

R1 — displaced marking. Where the source marks a relation or attitude by a grammatical form the target lacks, the absence of a same-category counterpart is not the absence of the marking. Render the site again under a brief that requires the marking to appear, and let it fall wherever the target does mark such things — on a pronoun, a verb, an address noun, a courtesy formula, an adjective, the argument structure of the clause.

R1 assumes those places are empty. Nine sessions of evidence went into admitting it and not one of them asked what happens when the source has already used the device the translator would reach for.

S095 hit exactly that, once, by accident (RS-20260803b-honorific-carry). Canth's nine-year-old wakes her mother with a plural imperative — the honorific — and the compensation the project's own register file had licensed, a kin term carrying the deference, was unavailable because Canth had already written the kin term: «Äiti, täällä on rouvia. Nouskaa ylös!» The vocative slot was spent by the source's own vocative. The session recorded this in one sentence and moved on.

One case is an anecdote. This experiment asks whether it is a condition.

What this unit teaches about translating literature, in one sentence — rewritten under amendment A11 (critic pass 2, findings 2 and 11), because the first version claimed far more than this design can reach: whether one of the repairs R1 licenses — adding an address noun to carry a relation the target's grammar cannot carry — gains a reader anything at a site where the source has already written an address noun of its own. That is a claim about one device, in one translator's English, on one scene. It is not a claim that R1 is "structurally impossible" anywhere, and nothing in framework/v0.1 turns on it (A5).

2. Materials

Source. Gottfried Keller, «Die drei gerechten Kammacher» (1856), paragraphs 16–28, 1,981 German words: the picnic scene on the hill. DE→EN is a pair R1 is not evidenced on — framework/v0.1 §2 declares IT→EN, FI→SV, FI→EN, JA→EN, RU→EN, LZH→EN, ES→EN — so the material is fresh ground for R1. It is not, after amendment A8, a scoring of v0.1's registered prediction 1: see §6.

The filed translation. T-kammacher-R04-v1, committed at 13e3e6b with its 19-decision log, before this design existed and before the census below was built. Its log records the losses at D10 (the whole Sie/ihr system), D11 (the one Ihr to a single man) and D12 (Konjunktiv I) without any repair attempted, which is what an R04 close translation does.

Contamination, measured before the span was chosen and not after. A 229-word gate span from outside this span was rendered by the lead from the German alone and compared with Wolf von Schierbrand's 1919 English (PG 34505) by tools/dependence_check.py: 13 shared 7-grams, 2 shared 12-grams, longest common run 13 tokens, verdict DEPENDENT?. Raw at materials/gate/. Consequence, binding on §7: the lead is not independent of Schierbrand on this work, so the HUMAN arm is a descriptive comparator and never an independence claim. Every primary test below is a within-lead comparison and does not turn on that independence.

3. The census, and the occupancy rule — both fixed before any rendering under R1's brief

Class A site, restricted to one family. A stretch of the span in which a grammatical form of the second person marks the social relation between speaker and addressee — T/V choice, number used as register, or the older polite plural Ihr — for which English has no same-category counterpart, because English has one second-person pronoun and no register-marked verb morphology. Konjunktiv I (log D12) and the diminutives (log D9) are Class A by the project's standing definition and are deliberately excluded: mixing marking families would confound family with occupancy.

The licensed device inventory, taken verbatim from R1's own wording and then narrowed to what English actually has for address relation: (i) an address noun — title, name, epithet-plus-name, kin term; (ii) an explicit courtesy formula; (iii) verb-phrase modality; (iv) the thou/you contrast; (v) syntactic elevation or inversion.

Primary device — narrowed by amendment A1 (critic finding 1). As first frozen the rule counted (i) or (ii) as occupying. The critic's objection is decisive: occupancy has to be a device-level reuse test — can the same device the repair uses be added without redundancy or clash — and the repair (A2) inserts an address noun. So occupancy is now defined on (i) the address noun alone; the courtesy-formula detector is removed from occupancy(). (iv) thou/you is left available at every site, occupied or free, which runs against the hypothesis and is left in for that reason.

The occupancy rule, as amended. A site is OCCUPIED if its source excerpt already contains an address noun directed at the hearer. Otherwise FREE. Coded mechanically in materials/sites.py against a token list fixed before coding, and re-asserted by analysis/verify.py.

The arithmetic consequence, stated before the run and not after: A1 moves O7 (occupied only by the courtesy formula «Mit Verlaub») from OCCUPIED to FREE. The split is 7 FREE / 5 OCCUPIED, not 6/6, and the permutation test below is over C(12,5) = 792 splits.

The mechanism the rule is meant to capture, stated so it can be wrong. At an OCCUPIED site the faithful English already contains the device — "immodest Dietrich", "By your leave", "Dear friends" — so a reader can already read the relation off the filed rendering, and a compensation placed there adds nothing a reader can detect. The prediction is therefore not that R1 fails at occupied sites but that it has no headroom there.

Twelve sites, seven FREE and five OCCUPIED after A1, in three strata by who addresses whom:

Stratum C is unbalanced and the reason is a fact about the book, declared here rather than discovered later: the journeymen essentially never address Züs without a vocative or a courtesy formula. Occupancy is partly a property of direction of address in this text. Strata A and B carry the within-stratum comparison; the pooled figure is reported with this stated.

4. The arms, as amended

• FILED — verbatim substring of T-kammacher-R04-v1, asserted as a substring by the verifier.
• FORCED_MECH — amendment A2, and the heart of the redesign. R1's repair applied by a function with no per-site judgment in it (materials/sites.py:forced_mech): insert exactly one address noun, drawn from the closed three-item list {familiar → "lads", ceremonious toward a man → "sir", ceremonious toward a woman → "ma'am"} by the site's frozen footing, set off by a comma, immediately before the first punctuation mark of the excerpt. Change nothing else. The identical repair is applied at every site, free or occupied. This is what makes the question answerable: if the repair lands at free sites and breaks or vanishes at occupied ones, that is a fact about the slot and not about how hard the translator tried.
• FORCED_AUTHOR — the hand-written arm from the first freeze, retained in sites.json and NOT SCORED, so the record shows what was withdrawn and why (critic finding 3: loud markers at free sites, token gestures at occupied ones).
• DECOY — amendment A4. Reworded, change-matched to FORCED_MECH and marking-neutral. Its job is precisely and only to catch a grader who scores any changed rendering as marked; it is not an unmarked baseline, and at occupied sites it necessarily keeps the source's own address noun. Three DECOYs the critic named as accidentally carrying register were rewritten.
• POSITIVE — amendment A3. The relation stated outright. Removed from the scored batch and graded in a separate call, so it cannot teach the key mid-batch.

HUMAN — Schierbrand 1919 at the same excerpt — is added by materials/human.py in a separate commit after this design is frozen, so that neither the design nor any lead rendering can be tuned to it.

5. Procedure

1. Three non-Anthropic seats (P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, P5 deepseek/deepseek-v4-pro). P3 is the pre-run critic and grades nothing.
2. Each seat receives, per site: the relation stated neutrally, with the source's device never named, and the four scored renderings (FILED, FORCED_MECH, DECOY, HUMAN) unlabelled. The German is not shown. This is a declared deviation from the E-20260802e procedure that v0.1's prediction 1 names: there the sources were Japanese, Russian, Literary Chinese and Spanish, and showing them leaked little; German Sie read by a frontier model leaks the answer outright. The deviation is conservative — it can only lower recovery — and prediction 1 is scored under it.
3. Question per rendering: does an English reader with no access to the source get this relation from it — YES / NO, plus one line of reason.
4. Order is fixed by sha256, never chosen by the lead: the order of sites within a call, and the order of the five renderings within a site, are both derived from sha256(site|seat|rotation). Each seat sees two rotations; call.py posts temperature: 0, so a rotation is a different prompt and not a re-draw of the same one (note from S095's critic).
5. Two halves of six sites per rotation, split by sha256 within each occupancy class, to bound output length. 3 seats × 2 rotations × 2 halves = 12 calls, plus 3 attention-check calls (POSITIVE only, all twelve sites, one per seat) = 15 calls.
6. Judgment is not parallelised; calls are strictly sequential.

6. Predictions — registered, six, before any call

Renumbered Q1–Q5 under amendments A5 and A8. Nothing in framework/v0.1 is amended contingent on any of them (A5, critic findings 6 and 10): whatever this returns enters the release as an open question with a measurement attached, in the section for what v0.1 is not evidenced for.

Not scored here (amendment A8, critic finding 9): v0.1's registered prediction 1. It names the E-20260802e procedure, and §5.2's German-hidden deviation is not that procedure. The raw FORCED_MECH recovery figure is reported as a measurement under a more conservative instrument and is not a discharge of prediction 1.

Registered floor check (A12, critic pass 2 findings 3 and 10). If FILED recovery at FREE sites is not near floor — operationally, if the mean FILED YES-rate over the seven FREE sites exceeds 0.50 — the instrument is reading the relation into unmarked English and Q1 is reported but not interpreted. The grading header now reads "Default to NO … answer YES only if the relation is POSITIVELY MARKED" for the same reason.

Registered sensitivity (A13, critic pass 2 finding 7). Q1 is reported (a) pooled, (b) strata A and B only as the primary stratified figure, and (c) under leave-one-out over the five OCCUPIED sites, so no single clash site can drive it.

Statistics. Q1 is tested by an exact permutation over all C(12,5) = 792 ways of splitting twelve sites into a seven and a five, one-sided on the registered direction, on the site-level gain score (mean YES across 6 gradings per arm per site). Reported stratified as well as pooled (A7): stratum C is 1 FREE / 2 OCCUPIED after A1 and carries nothing on its own. Nothing is tested with anything else; no test is chosen after the numbers exist.

7. What this run cannot show

• Nothing about quality. No seat is asked whether any rendering is better. Tier D is NOT PASSED and every arm of this design is availability-grading, evidence class X2/X1b, never X3.
• Nothing about human readers. Three machine seats. No human reader is available to this project (charter §9).
• No independent human corroboration. The 13-token gate measurement forbids it (§2).
• One work, one translator, one span. Twelve sites in one scene of one Swiss novella.
• The FILED arm was written by a translator who already knew this question existed. It was written first and committed before the census, which is the strongest control available and is not the same as blindness. If FILED is inflated by that knowledge, Q1's effect is understated.
• Amendment A6 — the contamination reaches the baseline, not only the HUMAN arm (critic finding 7). Span-level dependence between T-kammacher-R04-v1 and Schierbrand 1919 measures 27 shared 7-grams, 4 twelve-grams, one fifteen-gram, longest common run 15 tokens (materials/span-dependence.out). What this run measures is therefore recoverable address register in this translator's English — not a property of English, and not a property of R1 as a general repair.
• The twelve sites were chosen by the lead and that cannot be repaired inside this project (critic finding 2). There is no second translator here. A2 removes the compensation-writing channel; site selection remains a lead judgment and is declared, not argued away.

8. Budget

Pre-flight, worst case built from max_tokens (note (abc)):

Against $4.086701847 of headroom on UTC 2026-08-03. Lead translation, the census, all scoring and all verification are free.