Repository path: workshop/experiments/E-20260903b-line-end-order/design-v2.md · rendered 2026-09-09

Design v2 — the operative design: a prospective re-run, both gates binding, no reused code in any primary

This supersedes design.md, which stays frozen as the record of what was proposed and refused. Eight BLOCKING findings across two critic seats landed on one point and it is granted: a gate rewritten after seeing which seats it excluded cannot license the figures it lets through, and disclosing that does not repair it. critic-response.md carries the dispositions.

E-20260901-inversion-habit v2 §5's coding instruction and §7's PR1, PR2, PR4 are imported verbatim. PR2b is dropped with its cells (§2 below); PR3 was reported at RS-20260901-inversion-habit §3 and is closed.

1. The rule that governs everything else

No ORDER code bought at S237 enters any primary, any interval, or any sensitivity analysis. S237's codes are read once, at the end, as a between-run diagnostic, and the verifier asserts that no primary function ever opens stage_b_P*.json from that run.

Every voting seat is dispatched after gold2 is coded, frozen and committed, and after both gates are registered in this file.

2. Item set — trimmed so that three seats fit the day

The cells are E-20260901-inversion-habit v2 §4's, unchanged in definition. L-RHYME, O-P and N-P are dropped, which drops PR2b with them: PR2b was secondary and v2 §7 already said its definitional component made it weak. What remains is exactly what PR1, PR2 and PR4 need, plus the instrument control.

Payne 306, Leaf 81. 37 of the 387 carry a damaged final token and are flagged, as at S237.

3. Seats — three, all dispatched after the gates are committed

New tags Q1–Q3 are used deliberately so that no file, figure or sentence can confuse a prospective code with an S237 one. Non-Anthropic throughout (charter §5). P4, P5 and z-ai/glm-5.2 are out under notes (bps), (bne), (brt). x-ai/grok-4.5 at reasoning: {"effort": "low"} is out by registered rule, note (bsp), and at default effort it prices at roughly thirteen times its S237 cost, which does not fit the day.

qwen/qwen3.7-max is priced from GET /api/v1/models before dispatch, as the gate in config/models.md §S182 requires for any reserve seat, and the read is recorded in the cost table.

4. Two gates, both binding, both registered before any code exists

Gate 1 — the synthetic floor, RESTORED. 24 pairs, each one real printed line whose printed order is canonical plus a counterpart made by moving that line's finite verb to the end and changing no word. Bar: ≥ 21 of 24.

The one key correction, under a rule fixed before any new code exists. A calibration item's key is corrected if and only if (a) the lead's independent reading finds the key wrong and (b) all three S237 seats coded against it. Applied to all 24 items, exactly one qualifies: CAL11m, whose hand-made "canonical" counterpart "To me the East wind yesternight hath brought the tidings rare" postposes its adjective and leaves the adverbial before the subject. Its key changes CANONICAL → INVERTED. The floor is scored on the corrected key, and both the corrected and uncorrected scores are reported for every seat. The audit over all 24 items is committed at runs/RS-20260903b-line-end-order/calib_audit.json.

A consequence of the correction, recorded because it weakens the floor slightly. Pair 11 now has both members keyed INVERTED, so the construction property "exactly one member of each pair is INVERTED" no longer holds for that pair. The floor is therefore 24 keyed items, not 12 certain-by-construction pairs, and CAL11m's key rests on the lead's reading plus three seats' agreement rather than on construction. Pair members still never share a batch.

Gate 2 — real-item validity. 60 study items, drawn by random.Random(20260903) from the 387, disjoint from S237's 40 gold items, printed with hand, ode, cell and every seat's code stripped and the order shuffled, coded by the lead twice in two passes separated by a reshuffle, with disagreements adjudicated and the adjudicated key frozen and committed before any seat is dispatched. Self-agreement between the two passes is reported. Bar: ≥ 0.75 agreement on ORDER.

A seat votes only if it clears BOTH gates. This is stricter than S237, which is the point: the charge the critics made is that the author loosened a rule after it fired, and the answer is to tighten it and buy fresh codes rather than to argue.

What decides is the observed point — floor ≥ 21 of 24 and gold2 agreement ≥ 0.75. Wilson 95% intervals are printed beside both and decide nothing (C1-A16, C2-A12).

5. Concordance among qualified seats

• Three qualified seats ⇒ item code is the majority; no majority ⇒ dropped.
• Two qualified ⇒ a disagreement is a drop.
• Fewer than two ⇒ no ORDER figure at all (failure criterion 2).
• Per-seat reporting trigger (C2-A4): if any qualified seat's INVERTED rate in any of the four primary cells differs from the pooled rate by more than 0.15, every primary is reported per seat as well as pooled, and the result page says which seat is the outlier. A qualified seat is never silently outvoted.

6. Enjambment — a separate field, and mandatory stratified sensitivity

C1-A11, C1-A12 and C2-A8 are granted: the three-step instruction asks whether the marked line's last word "would still be last" in prose, and for a line whose clause continues overleaf the honest answer is not a word-order fact about that hand. Two changes, neither of which touches ORDER's three values or the halt logic:

• A fourth returned field, CLAUSE: COMPLETE — the clause containing the marked line's last word finishes inside the printed couplet; CONTINUES — it runs past it.
• Both primaries are recomputed on CLAUSE == COMPLETE items only, and that recomputation is mandatory, not optional. If the two versions of a primary fall on opposite sides of the registered bar, the primary is reported as indeterminate and the enjambment difference is the finding.
• The per-hand CONTINUES rate is reported as a fact about the two books.

7. Failure criteria

1. CTRL+ below 0.70 INVERTED on the qualified seats ⇒ instrument broken, no primary reported, whatever the gates say.
2. Fewer than two seats clearing both gates ⇒ no ORDER figure at all. Enforced by analyse.py returning before any cell rate is computed, and asserted by the verifier.
3. Mean pairwise agreement on ORDER below 0.70 among qualified seats ⇒ primaries withheld.
4. Drop rate above 0.20 in any primary cell ⇒ that cell is reported as the interval between assigning every dropped item CANONICAL and assigning every dropped item INVERTED, and the prediction outcome is read off the worse end (C1-A13).
5. Final-token damage rates differing by more than 0.10 between the hands ⇒ PR1 reported as a range. (S237 measured 0.0934 and 0.1329, difference 0.0395; recomputed here.)
6. Denominator below 25 ⇒ counts, not rates.
7. Sensitivity, always run: every primary recomputed (a) excluding UNCLEAR, (b) excluding damaged final tokens, (c) on CLAUSE == COMPLETE only.

8. The translation limb — exploratory, and labelled so everywhere

T-hafez-shahed-R60-v1 renders غزل ۱۵ whole, the no-radif counterpart of S237's sh77, selected by the frozen rule in workshop/translations/hafez-shahed/source.md, in the same declared plain modern register by the same hand. Its 20 line ends and the 16 of T-hafez-daasht-R58-v1 go into the coding batches indistinguishable from Payne's and Leaf's.

PR6 is withdrawn as a registered prediction (C1-A9, C2-A6): at n = 7 against n = 9 a count comparison has no decision rule worth the name. What is reported is the counts, per seat and pooled, as a descriptive case observation.

PR7 is relabelled a manipulation check (C1-A10, C2-A7), not a control: that the rhyming line ends of a poem carrying a finite transitive radif are coded INVERTED more often than those of a poem carrying nothing is close to definitional, and it is reported to show the seats read the lead's lines at all, never as evidence for anything.

The "upper bound" claim of v1 §4 is WITHDRAWN (C1-A8). The hand knew the question while writing sh15; that awareness could have acted in either direction, and nothing here bounds it. No confirmatory claim rests on the translation limb.

9. Verification

analyse.py --verify recomputes every reported number from the raw stage files, asserts body integrity, and asserts that no S237 stage file is opened by any primary path. analyse.py --mutate requires each mutation to change a figure this page actually reports.

10. Limits carried into the result page

1. The gold key is one coder — the same coder who wrote the definition, double-coded but not independent. An independent human coder is unbuilt and has been on the record as unbuilt for twenty-three sessions.
2. A null on PR1 is not fully separable from an instrument that cannot read natural verse (C1-A14, C2-A10). The restored floor shows a seat can detect a constructed contrast; nothing available to this project shows it reads subtle natural displacement.
3. Two hands, and register is confounded with hand. PR2 is the within-Payne primary and is where register is constant.
4. Eleven matched poems for PR1: Payne's volume 1 stops at Brockhaus CC.
5. No page images. All couplets are printed in runs/RS-20260901-inversion-habit/line-ends.md.
6. The translation limb is exploratory (§8), by one hand that knew the question.
7. Tier D is NOT PASSED. No sentence may say a reader hears, prefers or wants anything.

11. Pre-flight cost

UTC day 2026-09-03 has one prior session, S241, which spent $0.00. The whole $5.00 is available. $0.089021400 is already spent on the critic pass. Declared ceiling $4.50.

Built from max_tokens and the batch size actually dispatched, not from an assumed output length — note (abc). If the Q3 probe prices the seat above $0.09 a batch, Q3 is dropped and the run goes with two seats, which §5 permits and which the result page then says. The translation limb, the selection, both gold passes, all sampling and every interval are the lead's own and are not ledgered (charter §3, A4).