Repository path: workshop/experiments/E-20260903b-line-end-order/design.md · rendered 2026-09-09

Design — the withheld line-end primaries, re-opened under a gate that is not the one that stopped them

ARM-line-end-order step 1 (T3), 2026-09-03. This design does not replace E-20260901-inversion-habit v2. v2's §4 cells, §5 coding instruction, §7 predictions PR1, PR2, PR2b, PR4, §8 decision matrix and §10 limits are imported verbatim and are not re-opened. What changes is §6 (the gates), which is what withheld the run, plus two additions: a fresh validity set, and the lead's own line ends put through the same instrument.

1. What is being repaired, and the part of it that is post hoc

RS-20260901-inversion-habit bought 523 items × 3 seats and withheld every word-order figure because one seat of a required two cleared both registered gates. The gate that did the excluding was the 24-item synthetic floor, and three things are on the record about it:

1. It excluded the seat with the highest real-item agreement in the run. P2 agreed with the lead's blind gold at 36 of 40 against the voting seat's 32 of 40, and failed the floor at 20 of 24 against a bar of 21.
2. One of the 24 floor items is defective and all three seats caught it. CAL11m's hand-made "canonical" counterpart postposes an adjective and leaves an adverbial before the subject; it is INVERTED, the key says CANONICAL, and every seat lost a bit to it.
3. Both pre-run critics said in advance that the artificial task would establish little (C1-A7, C1-A8, C2-A4, C2-A11). They were right about passing it. What nobody predicted is that failing it establishes little either.

The repair: the synthetic floor is demoted to a reported diagnostic, and agreement with real, hand-coded items becomes the gate.

The part of this that is post hoc, stated plainly and not buried. The decision to demote the floor was taken knowing that P1 passed it and P2 did not, and knowing both seats' gold scores. A gate rewritten after seeing which seats it excluded cannot license anything on its own. Two things are done about that, and they are the reason this design exists rather than a re-analysis:

• A fresh validity set (§3). The gate is evaluated on 60 study items the lead codes blind now, disjoint from the 40 already used, frozen and committed before any seat is scored on them. The threshold is unchanged at ≥ 0.75 — it is v2's, not a new number chosen to fit.
• A seat whose codes do not yet exist (§2). One new non-Anthropic seat codes the whole item set under a gate registered before it was dispatched. If it qualifies, at least one voting seat is qualified by a criterion that could have excluded it.

What remains post hoc even so, and is declared as a headline limit on the result page: P1 and P2's codes were bought before this gate was written. Their re-qualification on fresh items is an unbiased estimate, but the choice to keep looking for a gate they could pass is not something a fresh sample can undo.

2. Seats

Non-Anthropic throughout (charter §5). P4 and P5 are out under notes (bps) and (bne); z-ai/glm-5.2 is out on long prompts, note (brt).

P6 is priced from the API before dispatch — config/models.md carries $1.475 / $4.425 read 2026-08-04, and the pricing gate that table's S182 entry describes applies to any reserve seat.

3. The fresh validity set — the gate

• 60 items, drawn by random.Random(20260903) from the 499 study items, disjoint from the 40 gold_uids of S237.
• Printed by gold2_blind.py with hand, ode, cell and every seat's code stripped, in shuffled order, exactly as S237's sheet was.
• Coded by the lead on ORDER alone, using v2 §5's three-step instruction verbatim, and frozen and committed to gold2.json before any seat is scored against it and before P6 is dispatched.
• Gate: a seat votes only at ≥ 0.75 agreement with gold2 on ORDER. Wilson 95% intervals are reported beside every agreement figure, so a threshold decision is never presented as a point.
• The circularity is declared and is unchanged from v2: the lead wrote the coding definition, so this measures whether a seat applies this definition, not whether the definition is right.

The 24-item synthetic floor is computed and reported for every seat and gates nothing. The defective item CAL11m is reported both included and excluded.

4. The lead's own line ends — the translation limb, wired

T-hafez-shahed-R60-v1 renders غزل ۱۵ whole — the no-radif counterpart of S237's sh77, selected by the frozen mechanical rule in workshop/translations/hafez-shahed/source.md, in the same declared plain modern register by the same hand. 36 line ends go into the coding batches, indistinguishable from Payne's and Leaf's: 20 from sh15 and 16 from T-hafez-daasht-R58-v1 (sh77), one couplet at a time, no hand, no poem, no rule.

The wire, in one sentence. The study limb asks whether a published hand who inverts where a radif obliges him also inverts where nothing obliges him; the translation limb puts the same question to the hand that wrote the log, by rendering a ghazal that obliges nothing and sending both poems' line ends through the study's own blind instrument.

Declared before any of it is coded (R60 §5, and the log's own §5): the hand knew the question while writing sh15, and T-hafez-daasht-R58-v1's log did not. The unbiased side of the contrast is the obligated one; any bias runs to widen the contrast, so the contrast is an upper bound.

5. Predictions

PR1, PR2, PR2b, PR4 are imported verbatim from E-20260901-inversion-habit v2 §7 and are not restated or altered here. PR3 was reported at RS-20260901-inversion-habit §3 and is closed.

New, and registered before the lead's line ends are coded by anybody:

• PR6 (the hand's own spill, direction only). Among the unobligated line ends of the lead's two renderings — the first hemistich of each non-maṭlaʿ bayt — the INVERTED count is higher in T-hafez-daasht-R58-v1, where a radif is carried at every rhyming position, than in T-hafez-shahed-R60-v1, where nothing is carried. n = 7 and 9, so failure criterion 6 fires by construction: counts, not rates, and no interval is claimed. This is the translation limb's own version of PR2 and it is registered as an existence observation.
• PR7 (the obligated slot). In T-hafez-daasht-R58-v1 the rhyming line ends are INVERTED at a higher count than in T-hafez-shahed-R60-v1. This is close to definitional — the radif is a finite transitive verb and English cannot end a clause on one without fronting its object — and is registered as the positive control on the lead's own material: if the seats do not see it, they are not reading the lead's lines the way they read Payne's.

6. Failure criteria

Imported from v2 §9 with one substitution and one addition:

1. CTRL+ below 0.70 INVERTED on the voting seats ⇒ instrument broken, no primary reported, whatever the gates say. Unchanged, and it is the guard that does not depend on the gold set.
2. Any seat below 0.75 agreement with gold2 is dropped and reported. Fewer than two voting seats ⇒ no ORDER figure at all — the same halt analyse.py enforced in S237, and the same code path enforces it here.
3. Mean pairwise agreement on ORDER below 0.70 among voting seats ⇒ the construct is reported as not reliably codeable and the primaries are withheld.
4. Drop rate above 0.20 in any cell ⇒ that cell is reported as a range over the dropped items.
5. Final-token damage rates differing by more than 0.10 between the hands ⇒ PR1 is reported as a range. (Measured at S237: Payne 0.0934, Leaf 0.1329, difference 0.0395 — does not fire.)
6. Denominator below 25 ⇒ counts, not rates. Fires by construction on PR6 and PR7.
7. Sensitivity analysis, always run: every primary recomputed excluding UNCLEAR items and damaged-final-token items.
8. New — the qualification-provenance flag. If the voting seats are only seats whose codes predate this gate (i.e. P6 fails), every reported primary carries that fact in the sentence that reports it, and the result page's headline says the figures rest on seats re-qualified after their codes existed.

7. Statistics

As v2 §7: Wilson 95% on every rate, Newcombe 95% on every difference, and an ode-clustered bootstrap (5,000 resamples of odes with replacement) on both primaries. Outcomes are three-way — supported / contradicted / indeterminate. Non-unanimous items among voting seats are resolved by majority; with two voting seats a disagreement is a drop, and the drop rate is printed per cell.

8. Verification

analyse.py --verify recomputes every reported number from the raw stage files and asserts body integrity; analyse.py --mutate corrupts inputs and requires each mutation to change a figure this page actually reports — the repair RS-20260901-inversion-habit §8 made after a mutation passed harmlessly because the gate had withheld everything.

9. Pre-flight cost

UTC day 2026-09-03 has one prior session, S241, which spent $0.00. The whole $5.00 is available. Declared ceiling $2.60.

Built from max_tokens and the batch size actually dispatched, not from an assumed output length — note (abc). The translation limb, the selection, the fresh gold coding, all sampling and every interval are the lead's own and are not ledgered (charter §3, A4).