Repository path: workshop/experiments/E-20260813d-first-span-again-dakghar/design.md · rendered 2026-09-09

E-20260813d — span A again, and what the declared priming cost

ARM-dakghar step 4, the arm's closing step. Frozen 2026-08-13 before a word of the re-rendering was written, before R05-v1's span-A English was opened, and before Mukherjea 1914 was fetched. Every prediction in §5 was written against the Bengali, the closed register and the collation alone.

1. The question, and why this arm can ask it and the last one could not

«ডাকঘর» is translated whole — 430 units, 5,458 Bengali speech-words, three spans, S160 to S170. The arm owes a craft report (register.md §What the craft report owes), and a craft report can be written as recollection. E-20260806d on «Köyhää kansaa» made one number of it instead, by putting span 1 back through the finished apparatus. This design does the same and adds the thing «Köyhää kansaa» had no way to ask.

Span A was translated under a declared priming event with a known boundary. Before span A was chosen, the lead read Devabrata Mukherjea's authorised English (Macmillan 1914) — its dramatis personae and its Act I through unit [16]. Span A is therefore contamination: high, and ARM-dakghar's constitution says the exposure boundary is what makes it measurable rather than merely admitted. Spans B and C are blind.

Nothing has ever measured what the exposure did. Span A's own study limb (RS-20260811f-dakghar-address) measured overlap on the exposed and unexposed portions of one rendering, which cannot separate this translator was primed here from this passage is easy to converge on. The missing arm of the comparison is a rendering of the same units by the same translator without the exposure. This session's lead has not opened Mukherjea. So:

Re-render span A whole under the closed register, blind to R05-v1 and blind to the comparator; freeze it; and only then open both. Two questions come out of the one freeze:

• Q-learn. What moves when the first span goes back through the finished apparatus — the closed register, the whole-play censuses, the ending?
• Q-prime. Does the exposed stretch [1]–[16] carry overlap with Mukherjea that the unexposed stretch [17]–[86] does not, in v1 but not in v2?

2. The design in one paragraph

Re-render units [1]–[86] whole from the gated copy-text under register.md as closed at span C, blind to v1's English and to Mukherjea. Freeze as T-dakghar-R05-v2-spanA with its own log. Then (i) run tools/dependence_check.py over the three pairs {v1, v2, Mukherjea} × two portions {EXPOSED [1]–[16], UNEXPOSED [17]–[86]}; (ii) diff v1 against v2 unit by unit and code every divergence by cause; (iii) run an independent pre-run critic over the frozen design and the completed coding; (iv) put a stratified sample of unit pairs to three blind seats, who see the Bengali unit and two English renderings in randomised slots and answer one question — do these differ in what they say, or only in how they say it?

3. Materials, and what the translator was and was not shown

The blindness is partial, by construction, and in a declared direction. register.md quotes v1's own lexis — Uncle, Auntie, Grandad, kabiraj, chhatu, nagra, Amal-babu, Mr Datta, Curds — curds — good curds!, the old witch with the matted hair — because that is what a register is for. So v2 is blind to v1's sentences and bound to v1's rules, which biases this experiment towards agreement exactly where §5 predicts agreement. A prediction of no divergence at a rule-governed site is cheap and is labelled cheap.

One leak is narrower than the register and must be named. register.md's V3 row says "and see D9 — the lead had seen a comparator's one-word kin-name". Reading the register therefore told this session's lead that Mukherjea renders ঠাকুর্দ্দা in one word, and V3 binds v2 to Grandad regardless. That single site is excluded from Q-prime by prior declaration, and it is in the UNEXPOSED portion ([20]–[21] onward), so the exclusion works against P1 rather than for it.

4. Why (A) COPY-TEXT is vacuous here, said in advance

E-20260806d reported "(A) COPY-TEXT is zero" as a finding about its collation gate: seven first-edition readings were applied between v1 and v2 and not one changed an English word. This design cannot produce that finding, because span A's four corrections were applied before v1 was written and the file has not changed since. The class stays in the scheme so the count is auditable, and a count of 0 here is a structural fact, not a result.

5. Predictions, registered before translating

5.1 Q-prime — the priming measurement

Reported measures, all from tools/dependence_check.py (shared 7-grams, 12-grams, 15-grams, longest common contiguous run in tokens, each also name-excluded), on six cells: three pairs × two portions.

The power limit, declared now. EXPOSED is 16 units and 217 Bengali words — perhaps 320 English. Independent hands routinely share zero 12-grams over stretches ten times that long. If all three EXPOSED longest runs come back ≤ 5 tokens, P1 is NOT READ and this design records a null with no power, because runs at that length are ordinary function-word coincidence (tools/dependence_check.py header: a single shared 7-gram means nothing; 75 of 78 independent pairs have one).

The confound, declared now. v2 is a reviser's pass by the same hand. If v2 reproduces v1's sentences, it inherits v1's Mukherjea-overlap and P1 fails for a reason that is not "the exposure left no residue". The discriminating datum is therefore reported as its own list: the n-grams shared by v1 and Mukherjea that are absent from v2, and the converse. A P1 failure with LCR(v1,v2) high on EXPOSED is uninterpretable and will be reported as uninterpretable.

5.2 Q-learn — three named sites and one count

Written against the Bengali and the closed register with no v1 English in view. One predicts a null.

The count. Every divergence site is coded to exactly one cause, precedence (E) > (F) > (A) > (C) > (B) > (D):

• (A) COPY-TEXT — a corrected reading changes the English. Vacuous here; see §4.
• (B) LATER RULE — a register rule or names row created after span A determines the reading (V8, V9, V5 as settled, the §Settled retraction).
• (C) WHOLE-WORK — knowledge of the rest of the play determines the reading (the censuses, the ending, the ক্ষেপ- chain).
• (D) FREE VARIATION — the residual. Two renderings of the same Bengali by the same hand.
• (E) CORRECTION — v2 fixes something wrong in v1: a mistranslation, a dropped element, a collision.
• (F) REGRESSION — v2 is worse than v1. New class, added before the run. RS-20260806d §7 recorded that its frozen scheme "has no class for 'v2 is worse than v1'" and called that a defect in the design; two such sites existed and had to be named outside the coding. This design carries the class from the start.

P7 is coder-dependent and is registered as weak. The pre-run critic's first finding against E-20260806d was that a count whose residual class is chosen by the person who registered the threshold is close to unfalsifiable, and it is right. P7 is reported, is not the headline, and §5.1's mechanical measures are the headline instead.

5.3 Register agreement, declared cheap

6. The blind seat run, and its gates

Seats. P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, P3 x-ai/grok-4.5 — the three-seat practical jury. P5 deepseek/deepseek-v4-pro is excluded by NEXT.md's standing finding (five dead bodies of forty-two at S174 on a structured-output task); this run uses plain text and a generous cap regardless, and does not pin reasoning effort.

Item. The Bengali unit, then two English renderings labelled A and B in a slot randomised by a fixed seed, then one question: do these two English renderings differ in what they say (S), or only in how they say it (F)? Plus one line of reason. Seats are told nothing about provenance, authorship, dates, or that an experiment about priming exists. No seat translated anything. The lead does not arbitrate.

Strata, sampled by fixed seed from the coded diff:

Gates. Every gate is a withholding gate: if it fails, the seat primary is not read, and §5.1's and §5.2's numbers stand on their own.

• G1 (floor). REPEAT majority-S ≤ 1 of 4.
• G2 (ceiling). WRONG majority-S ≥ 3 of 4.
• G3 (completeness). Dead or unparseable bodies < 10% of all dispatched cells, and < 10% within each stratum that carries a reported number. Stated against both denominators on purpose: note (bna) fired twice at S174 because one gate was written over usable bodies and another over the wrong denominator.
• G4 (non-confirmatory by prior declaration). If EF has fewer than 3 items, the seat result is reported as descriptive only and is not read as a check on the causal coding — the precedent's FC4.

What the seat run can and cannot do. It is the remedy RS-20260806d §8.1 named and could not run: an external check on a translator's coding of his own two renderings. It is a better remedy here than there, because these seats read Bengali (RS-20260811f G1 passed at 0.778 on the source side) and are shown the source unit. It remains three models sharing a training distribution, and it rates nothing.

7. Order of operations, each step committed before the next

1. This design, frozen — before a word is translated.
2. T-dakghar-R05-v2-spanA and its log, frozen — before v1 or Mukherjea is opened.
3. Mukherjea fetched; §5.1 run; the diff coded. Raw tool output preserved.
4. Pre-run critic over the frozen design and the completed coding — before any seat call.
5. Seats. Raw bodies preserved.
6. Post-run verification recomputing every reported number from the raw outputs.
7. Result page and craft report.

8. Failure criteria for the unit as a whole

• If v2 cannot be produced blind — if v1's English or Mukherjea is opened before the freeze — the design is void and says so.
• If §5.1's EXPOSED runs are all ≤ 5 tokens, P1 is NOT READ (§5.1).
• If a gate in §6 fails, the seat primary is WITHHELD and the mechanical result is reported.
• The arm closes either way. ARM-dakghar step 4 is the craft report; the craft report is owed whatever these numbers are, and an honest null is one of the things it can say.

9. Cost

Translation is the lead's and is free (charter §3, A4). Spend is the critic and the seats.

Today's headroom (UTC 2026-08-13): $1.800491851 of the $5.00 cap, after three sessions. The ceiling is built from max_tokens and not from an expected length — note (abc).

10. Amendment 1 — a leak in the blindness, found before translating

Made 2026-08-13, after §1–§9 were frozen at b38bb6d, before a word of v2 was written and before any measurement existed. Recorded as an amendment rather than folded into §3, because the frozen text must stay readable as what was believed when it was frozen.

§3's claim that Mukherjea was not fetched is true and is not sufficient. The pages this design had to read in order to write P5 and P6 quote Mukherjea's English directly. RS-20260811f-dakghar-address §6 is a table of his renderings at six deference sites, and its last paragraph quotes two more of his lines. The exposure is:

The rule, registered now. The Q-prime cells of §5.1 are computed on the residual portions, with the seven word-quoted Mukherjea units removed:

• EXPOSED-R = [1]–[16] less [5], [8], [9], [12], [14], [15] → 10 units
• UNEXPOSED-R = [17]–[86] less [22], [43], [71], [72] → 66 units

The full-portion figures are computed and reported alongside, so the effect of the exclusion is visible rather than assumed. P1 and P2 are read on the residual portions only. The power limit of §5.1 applies to the residual and is therefore tighter than when it was written: EXPOSED-R is 10 units and roughly 150 Bengali words, and a null on it is close to certain. That is said before the number exists.

P1 is downgraded to a directional observation and is no longer this design's headline. The headline is Q-learn (§5.2) and the three-way overlap picture of §5.1 as a whole, including LCR(v1,v2), for which the leak is irrelevant — the leak is Mukherjea's English, and v1's span-A English is still unopened.

And v2's contamination: declaration follows from this table, not from a recollection: suspected, with the eight quoted units named on the artifact.

11. Amendment 2 — the pre-run critic's dispositions, applied before any seat call

x-ai/grok-4.5 (P3), one call, NEEDS-AMENDMENT, ten findings, six BLOCKING — run against the frozen design, Amendment 1, the completed coding and the completed mechanical measurement, before a single seat call, per §7 step 4. run/critic.json and run/critic.parsed.json. All ten accepted, none overruled. $0.071492400.

What the seat run is now for, stated before it is dispatched. One question only: do the 273 sites this translator called wording-only hide differences in what is said that three independent readers can see? That is the precedent's RT1 and it is worth its money. It is not a check on why anything changed.

And the critic's single_worst_problem is accepted as written: the same interested party registered the causal thresholds and assigned every causal label, so Q-learn's held/failed story is decided by the experimenter. The consequence is that this experiment's reportable findings are the mechanical and documentary ones — the overlap counts, the diff's structure, the root-chain's textual facts, the unlogged rule violation, and the leak absorption — and the causal tallies are printed as tallies.