Repository path: workshop/experiments/E-20260821c-echo-availability/design.md · rendered 2026-09-09

E-20260821c — does an Arabic rhyme reach English only where the ordinary English words already chime?

Track T1, ARM-alf-layla step 7. Study limb of span G (the copy-text's fourth night, pp. 33–36). This design and its locus table are frozen before the span is translated and before either published comparator is opened, which is the condition RS-20260815-supplied-sound-confirm established after the pre-run critic ruled that step 2's post-hoc predictions could describe and not confirm.

1. Where the question comes from

RS-20260820-rhyme-bearer (span F) closed with an observation it explicitly refused to call evidence:

Unregistered and labelled so: the one Arabic rhyme that does reach English does so in all three hands, and it is the one where the plain English words for the two things already chime.

Three spans of census stand behind it. Arabic rhymed prose (saj') reaches English almost never — Lane at 0 of 7 locatable loci in span A, Burton at 2 with 1 partial, the lead at 1 of 7 in span F's prose — and the exceptions are not distributed as effort would distribute them. The conjecture is that they are distributed as the English lexicon distributes them.

This is the only form in which that observation can become evidence: registered, with its predictor measured independently, before the English of a fresh span exists.

2. The registered hypothesis

H. At a rhymed Arabic locus, an English hand produces an audible sound-echo only where the ordinary English words for the rhyming members already echo. Where the plain lexicon does not supply the echo, no hand — however committed to sound — manufactures one.

H is a claim about translating literature, not about this project's instruments: it says what a translator's commitment to a source's sound can and cannot buy, and it is the kind of statement framework/ exists to carry.

Two hands cannot be influenced by this design. Lane published in 1839 and Burton in 1885. They are the primary. The lead is a third hand and is hypothesis-aware — it will translate span G after reading this page — so the lead's row is reported and is excluded from the primary. What governs the lead's own behaviour is register.md S1, a standing rule from span A: attempt an English rhyme at every locus; take it only where the sense survives; record the loss where it does not.

3. Materials

• Source: workshop/translations/alf-layla/source-hindawi-2022-spanG.txt — Hindawi 2022 vol. 1, pp. 33–36, the copy-text's fourth night, 1,356 Arabic words including one poem of five abyāt.
• Hands: H-LANE Lane 1839 (PG #34206), H-BURTON Burton 1885 (PG #3435), H-LEAD T-alf-layla-R05-v1 span G. All three already stored in this repository.
• Seats (config/models.md; slugs logged as provenance in run.jsonl): stage A P2, QR; stage B P1, P2, P3. P4 and P5 are out (notes (bps), (bne)).

4. The locus table — frozen from the Arabic, before any English for it exists

Inclusion rule, fixed before enumeration: a locus is a run of two or more consecutive syntactic members standing in a coordinated or parallel construction whose final words share a rhyme — identical final consonant with the identical preceding vowel. Assonance without the shared consonant does not qualify (نديمي / حبيبي is excluded on this rule); a shared final short vowel without the shared consonant does not qualify (المضرة / النافعة is excluded).

Fifteen free loci, 39 members. L16 and L17 are listed because they are present and would otherwise look overlooked. Both are excluded by a rule written here, before any datum exists: each is a recurring frame formula whose English is already fixed for the lead by a register decision taken at an earlier night (V12 at span D; V14 at span E), so neither locus is free, and a locus that is not free cannot test a hypothesis about what a translator does with a free one.

Amended 2026-08-21, before translation and before any datum: L17 was in the free set when this table was first written, and the check for V14 recurrence — run against every stored span source — found the same sentence at the span E seam. The exclusion is recorded here rather than made silently.

5. Procedure

Sections 5–7 were rewritten 2026-08-21 after the pre-run adversarial critic returned four BLOCKING findings and no call had been dispatched. Every finding is answered in critic-response.md; nothing was refused. No second round was bought, per note (bqp): every change below narrows the primary.

Stage A — availability, measured blind, one member at a time

For each of the 39 members, a seat is given that member alone, inside the locus's canonical Arabic context with every other member of the same locus replaced by …. The masking is done in code (materials/build_items.py), not by hand, and every masked string is stored in items.json so a reader can confirm that no sibling member survives. The seat is asked for the most ordinary English word or short phrase for the marked item. It is never told the passage is rhymed, never shown another member, and never told what the answers are for. A seat that cannot see the pair cannot rhyme on purpose.

Two residues are declared rather than solved: L7's unmasked أصبح shares a root with the masked الصباح, and L14 is a proverb famous enough that a seat may complete it from memory.

Two seats (P2, QR), temperature 0, one member per call. 78 calls.

Stage B — echo detection

Every item has one shape: a list of two to five English expressions, unlabelled, with no Arabic, no source, no hand identity and no statement of what the list is. The seat answers ECHO: RHYME | NEAR | NO and, when not NO, quotes on a PAIR: line the two expressions it heard.

All 87 items are shuffled together by random.Random(20260821) before dispatch — the seed is fixed in run.py — so controls are interleaved with data in one reproducible order.

P2 sits in both stages. It answered stage A one member at a time, in separate calls, never seeing a list; the risk that it recognises an assembled list as its own is small and is declared rather than assumed away. The primary is computed with P2 in and again with P2 dropped.

What a HAND item contains, and the one discretion left

A HAND item is the hand's own words, verbatim: for each member, the shortest verbatim span of the hand's sentence that renders it — normally its whole colon. Nothing is reduced to a chosen final word. Where a hand renders more cola than the Arabic has members, all of them are kept and the surplus is flagged on the item; H-BURTON's L7 has four cola against the Arabic's three, and four is what he wrote.

This is the answer to the critic's BLOCKING 2, which showed that the earlier "last content word" rule was not being applied and was making choices that changed what a detector would hear. The remaining discretion is the choice of span, and every span is stored in items.json beside the hand's full sentence so a reader can check it.

The change makes the design conservative against its own hypothesis, and that is said before the run: verbatim cola are longer than single words, so a detector has more chances to hear an incidental echo anywhere in the list — including at the availability-0 loci where H predicts none. The HAND block is biased toward falsifying H.

One rule for a member a hand does not render

A locus is a cell for a hand when the hand renders at least two of its members. Members it does not render are recorded as ABSENT, do not appear in the list the detector sees, and are counted in the report. A locus with fewer than two members rendered is not a cell for that hand and is named in the result. One locus is affected: L2 is a Lane cell with three of its five members, because Lane has no counterpart for الأوان or for أرض رومان.

The definitions the analysis runs on, fixed before dispatch

• A stage-B item is an echo when a majority of surviving seats answers RHYME or NEAR.
• A locus's availability is how many of its two AVAIL lists (0, 1 or 2) are echoes. Availability-0 = both majority-NO. Availability-2 = both echoes. Availability-1 is reported on its own row and pooled with 2 only for the ≥ 1 comparison PR2 names.

6. Predictions, registered

• PR1 — availability is rare. At ≤ 5 of the 15 free loci will both AVAIL lists be echoes.
• PR2 — the primary, and it is a tolerance claim (H-tol). Over the published-hand cells (Lane + Burton, less dropped loci): the echo rate at availability-0 loci is ≤ 0.10, and is strictly lower than the rate at availability ≥ 1 loci. A tie fails the directional half.
• H-strict, reported and never folded in. Every availability-0 published-hand cell coded as an echo is a counterexample to the universal form of H. The result reports the exact count and quotes every one of them. One is enough to withdraw the universal claim, and the result will say so in those words.
• PR3 — names ride free, tested at L9 only. L9's two members are proper names in all three hands and in the AVAIL lists. PR3 holds if L9 is coded an echo in every hand and both AVAIL lists and the seats' PAIR: lines name the two names. L2 is not used for PR3: Lane does not render رومان, and a five-member list can be coded an echo on a pair that has nothing to do with names.
• Descriptive, not a prediction. L4, L5, L6 and L13 carry their Arabic rhyme on a bound morpheme English has not got as a rhyme-bearer. Their cells are reported beside the rest with their counts. The earlier PR4 is withdrawn: it fixed neither denominator nor failure condition, and the critic was right that it could not cleanly succeed or fail.

7. Failure criteria, registered

1. Instrument. Each stage-B seat must return the registered verdict on ≥ 10 of 12 CTRL items. ECHO controls pass on RHYME or NEAR; NONE controls pass on NO only; NOT-RHYME controls pass on NEAR or NO. C5 additionally requires the PAIR: line to name thunder and wonder. A seat below the bar is dropped; fewer than two seats surviving and no primary is reported.
2. Withhold. If fewer than three availability-0 published-hand cells exist, PR2 is withheld and the run is reported as descriptive.
3. Withhold. If fewer than 10 of the 15 loci are cells for both published hands, the run is reported as descriptive and says so.
4. A seat's item is dead after two dispatches; > 5% dead in any block and that block is reported with its loss rate on the face of the result.

8. What this cannot establish

• No human reader is involved. The detector is a language model, so "audible" means a model coded it as an echo. Every figure carries that qualification.
• Relocated compensation is invisible. S2 allows an English rhyme to fall away from the Arabic's position; an item that lists the members' own renderings cannot see a compensating figure built elsewhere in the sentence. This design measures echo at the locus, and prior spans' censuses measured the same thing.
• The lead's row is not evidence for H, having been produced after this page was written.
• ABSENT stays ambiguous between the hand dropped it and it was not in his Arabic: Lane worked from Būlāq and Burton from Calcutta II, and the Būlāq/Calcutta II collation is closed as an honest null (ARM-alf-layla §Blocks).

9. Budget

Declared ceiling for the whole experiment: $1.40. The critic pass cost $0.087891, leaving $1.312109; the runner carries its own ceiling of $1.30 and a stop-loss of $1.15.

Pre-flight built as (max_tokens + reasoning.max_tokens) × output rate + prompt × input rate, note (bqs), and built on the assumption that every call needs the doubled-cap re-dispatch, note (bqk). Caps: stage A max_tokens 100 / reasoning 80; stage B 120 / 80 — max_tokens strictly greater than the reasoning cap in both, note (bpv).

$1.108 is under the runner's $1.15 stop-loss, which is what (bqk) requires: the guard sits above a worst case in which nothing goes right. Session worst case including the critic: $1.196, under the declared $1.40.

Realised 2026-08-21: $0.84032408, neither guard approached. The stage-A answer cap did not survive contact: QR renders its deliberation into the message content and truncated before its answer line, so 7 of the first 52 cells died — the cap was raised to 350 and the parser repaired twice, all three changes recorded in run.py's docstring with their arithmetic, none of them touching a prompt, a seat or a cell's content. Note (bqu).

Today's UTC ledger stood at $3.09812465 of $5.00 before this session, with $1.90187535 unspent.