Repository path: workshop/experiments/E-20260824-eye-or-ear/design.md · rendered 2026-09-09

E-20260824 — is the chime in the sound or in the look? ARM-echo-threshold step 2

Frozen before dispatch. The translator's log of T-gulistan-bab1d-R46-v1 was frozen at commit ee2608df, before this file existed.

AMENDED 2026-08-24 after the round-1 critic pass (critic-v1.json, NEEDS-REDESIGN, 12 findings — ten accepted in whole or part, one refused in writing, one partly refused). The amendments A1– A11 and the reasoning for each are in critic-response.md; they are applied below and marked where they bite. The design was re-submitted for round 2 after they were applied.

ROUND 2 RETURNED NEEDS-REDESIGN AGAIN, and this run goes ahead under that verdict, which is stated here rather than buried. A12–A18 apply what round 2 found and could be fixed. Its remaining BLOCKING finding cannot be fixed: it asks for a manipulation that varies spelling while holding pronunciation genuinely constant, and §2 and §3.2 of this design are the demonstration that English does not supply one at a phrase-end — that is the run's own first result, not an oversight. Its other unmet findings — no human readers, no corpus sampling frame, no independent phonetic coding, seats chosen on prior results — are this project's standing limits, recorded in config/models.md and carried on the face of every instruction (§11). Every one of them is printed in the result page's limits. The primary is restated by A14 as an association, not a causal decomposition of look against sound.

1. The question, and why it is not a question about our instruments

framework/v0.2 §7.26 item 5 tells a practitioner, in as many words, that this handbook cannot tell him whether the chime he has written works on the ear or only on the page:

The instruments read text. On a dedicated twelve-item probe they called non-rhymes that are spelled alike — sword / word, beard / heard, cough / dough — an echo at 0.722, against 0.778 for real rhymes spelled unlike; one of the three was doing spelling and not sound at all. A chime that works on the page may not survive being read aloud, and this handbook cannot yet tell you which of yours is which.

The subject-rule sentence (wiki/tracks.md §The subject rule), written before the design:

What this unit teaches about translating literature: whether an English pair standing at two phrase-ends is registered because of how the two words sound or because of how they look — and therefore whether a translator answering a source rhyme should care that his half-chime also looks alike on the page, which is a choice he makes at the desk between two candidate words.

That is a question about the reader at the receiving end and about a decision the translator takes, not about tools/rhyme_pairs.py, whose arithmetic is not in doubt and which here defines the independent variable rather than being measured by it. The seats' reading mode is a limit, named in §7 and bounded by a registered control (P3 below), not the subject.

2. What the naive design would have measured, and why it was thrown away

The first build of build.py asked for a filler alike in spelling and unrelated in sound — the plain-English reading of eye rhyme. Twenty-one of twenty-four loci were dropped by the mechanical check, all for one reason: tools/rhyme_pairs.py scores love/move, said/maid, done/stone, blood/wood, height/weight, bury/jury, break/streak as NEAR-N2 — pararhyme: the final vowels differ and the coda after them is identical.

An English eye rhyme is, in almost every case, also a consonance. The two are the same pair seen twice, and the same is true of two of the three examples §7.26 item 5 prints by name: sword / word and beard / heard share the coda R D / D, and the S212 probe's "non-rhymes spelled alike" arm was therefore carrying a real sound relation. §7.26 item 5's inference — one seat was doing spelling and not sound — may still hold for that seat, but the arm it was inferred from is not the clean orthographic contrast it was taken to be. That is this design's first finding and it is a gate, not the unit; it forced the design below and it is reported in the result page's §2.

3. Materials

130 items over 17 loci (A7: L18 and L19 dropped by hand after the mechanical check passed them), built and verified by build.py, manifest at materials/items.json. Each locus is one short English passage with two parallel phrase-ends. One end takes the fixed word, the other the filler; every cell of a locus is the same string with exactly one word changed, and the articles do not vary across fillers.

3.1 The main crossover — 14 loci, 8 cells each, 112 items

run is the number of final letters the two rhyme-bearers share (identical bearers → excluded). The ≥ 3 / ≤ 2 split is RS-20260822b stage P's own: its PHON arm sits at 1, 1 and 2 shared final letters, its ORTH arm at 4, 4 and 4.

X and V carry the same phonetic relation to the fixed word and differ only in whether the pair also looks alike. Their difference is the contribution of spelling and nothing else planted in the design varies with it. The crossover then cancels any main effect of either word, exactly as in E-20260822b: X, V, W, Y each stand once with a and once with b. Written out, the Y terms cancel and the primary is [hit(aX) − hit(aV)] − [hit(bX) − hit(bV)] — the X-against-V difference minus the same difference where neither carries a planted relation, which is what the b column measures and subtracts.

A3 (round-1 critic findings 3 and 4, partly accepted). What the b column does NOT subtract is a difference that exists only relative to a: within the N2 class blood / wood and blood / cloud are not equally close in sound, and the closer one may also be the spelled-alike one. The primary is therefore orthographic likeness PLUS any residual within-N2 phonetic-distance difference that travels with it, and every instruction it licenses says so.

A7 materials criterion, now binding rather than descriptive: at every locus X and V are monosyllabic with primary stress on the final syllable. L18 (mountain, first-syllable stress against monosyllables) and L19 (disyllabic machine against monosyllabic moon) were dropped for failing it; the rule cannot see a stress or syllable-count mismatch and passed them both.

Two loci, so the shape is legible:

L1 — "What ran in the street that night was the {fixed}; and what he saw when he lifted his eyes was the {filler}." fixed a = blood, b = ash; X = wood (N2, run 3), V = cloud (N2, run 1), W = mud (STRICT, run 1), Y = stone.

L12 — "What frightened him most about the tower was its {fixed}; and all that the boy had between his hands was a {filler}." fixed a = height, b = age; X = weight (N2, run 5), V = boat (N2, run 1), W = kite (STRICT, run 0), Y = sack.

Three of nineteen candidate loci were dropped by the check and not repaired — L4, L13, L16 — with the failing pair printed in materials/items.json. L4 and L16 failed because their X came back NONE rather than N2; they are re-used, in a different role, in §3.2.

3.2 The NEAR-EYE sub-arm — 3 loci, 6 cells each, 18 items, carrying no claim

sword / word, heard / beard and word / lord are the residue: alike in spelling at the rime and scored NONE by the rule. Cells X (NEAR-EYE), W (RHYME), Y (NONE) crossed with fixed a / b. Two of the three are the pairs framework/v0.2 §7.26 item 5 prints by name.

A2 (round-1 critic finding 1, accepted). These are NOT sound-unrelated and the sub-arm was renamed for it. All three end in /rd/ or /d/; the tool's NONE verdict is an artefact of its coda rule, which compares the coda after the final vowel and treats CMUdict's ER as a vowel, so word W ER D has coda D while sword S AO R D has coda R D. P4 is withdrawn as a claim and the sub-arm is descriptive only.

What that leaves, and it is this run's answer to the question §7.26 item 5 was really asking. Between the 21 loci the first build dropped and this finding: A16 (round-2 critic finding 7, accepted) — this is a failure to find, not a claim of non-existence. Under tools/rhyme_pairs.py, CMUdict pronunciations, and the candidate-generation procedure of build.py, the project could not construct a pair alike in spelling at the rime and unrelated in sound that two words could both stand at a phrase-end for. No corpus was searched and no human phonologist adjudicated. And the tool's r-coloured misclassification cuts towards scarcity, not away from it: the three pairs it did admit turned out on inspection to share a coda. The worry is your chime working on the page rather than the ear? is therefore hard to pose in the pure form it assumes.

3.3 The two mechanical censuses, at $0

• M-avail — the availability half of the arm's question, run on the frozen English of T-gulistan-bab1d-R46-v1 by sweep.py under the rules stated at the head of that file: how often does a chime of each kind sit at two adjacent phrase-ends of a translator's actual English? Already run; the figure is in the result page.
• M-census — over a declared list of classic English orthographic traps, what fraction of spelled-alike pairs are STRICT, N2, or NONE by sound? This is the number that decides whether the naive eye-rhyme question is answerable at all. census.py, list fixed before it was scored.

4. Procedure

Three stages. All calls temperature 0, one item per call, isolated context, dispatched in one shuffled stream per stage with seed 20260824. max_tokens 1600 reasoning / 700 body, the setting that ran 316 calls with 0 dead in S216.

Stage D — prompted detection, sound-framed (130 items × 3 seats = 390 calls)

The seat is shown one passage and asked, in a prompt identical for every item:

Read the passage below aloud in your head. Do the two phrase-endings chime — that is, do the last words of the two phrases echo one another in sound? Answer yes or no. If yes, name the two words and say whether the echo is full or half.

The prompt names sound and asks for an inward reading aloud, and that is deliberate: it is the form of the practitioner's own question. A neutral prompt would very likely raise the ORTH rate, and §7 says so as a limit. It is also E-20260822b stage D's framing, which is what makes the two runs comparable.

Scoring, RESTRUCTURED BY A12 (round-2 critic finding 3, accepted). Two outcomes, and the judgment is now the primary:

• say = 1 iff the seat answers yes. This is the primary outcome. It cannot be reached by copying a visually similar pair out of the passage.
• hit = 1 iff say = 1 and the seat names both planted phrase-end words (case-insensitive, either order). Reported beside say throughout.

A claimed orthographic effect must appear in say. An effect present only in hit is reported as an endpoint-naming effect and licenses no handbook sentence — that is exactly the artefact the finding names. A strength verdict of full or half is recorded where given and is descriptive. Unparseable or empty → one re-dispatch, then dropped and counted dead.

Stage N — the same question, neutrally framed (A4; 56 items × 2 seats = 112 calls)

Added on round-1 critic finding 5, accepted. A null under a prompt that says in sound and read it aloud in your head licenses only a claim about that instruction. Stage N puts the four cells that enter the primary — aX, aV, bX, bV of the 14 main loci — to P1 and P2 under a prompt that names neither sound nor spelling and asks for no inward reading:

Read the passage below. Do the two phrase-endings echo one another?

Δ_ORTH − Δ_CONS is computed under both framings and both are reported.

Stage R — naturalness ranking, control C1 (17 loci × 2 seats = 34 calls)

The four (three, in the sub-arm) a cells of each locus shown together, letters shuffled per (locus, seat); the seat ranks them for how naturally the English reads, saying nothing about sound. Seats P1 and P2.

Seats

P1 = openai/gpt-5.6-terra, P2 = google/gemini-3.6-flash, QR = qwen/qwen3.7-max (config/models.md). P4 and P5 are out on any task shape (notes (bne), (bps)); GL is out on LONG prompts; P3 is priced from a probe or not used.

The primary is computed on P1 and P2 only, and this is registered here, before dispatch. The ground is RS-20260822b stage P, which measured P1 at 1.000 phonetic / 0.500 orthographic and P2 at 1.000 / 0.833, against QR at 0.333 / 0.833 — a demonstrated spelling-matcher on this task, method note (bqy). QR is run anyway, as P3, a declared positive control (§6).

5. Estimands

Write hit(f g) for the mean hit over (locus, seat) cells with fixed word f and filler g. Every quantity is an interaction, so a main effect of either word cancels:

Δ_RHYME = [hit(aW) − hit(aY)] − [hit(bW) − hit(bY)] Δ_CONS = [hit(aV) − hit(aY)] − [hit(bV) − hit(bY)] Δ_ORTH = [hit(aX) − hit(aY)] − [hit(bX) − hit(bY)] Δ_EYE = [hit(aX) − hit(aY)] − [hit(bX) − hit(bY)], on the three sub-arm loci

Intervals are bootstrap percentile over loci, 10,000 resamples, seed 20260824; the locus is the resampling unit, seats and cells are not.

A13 (round-2 critic finding 4, accepted in part). What the interval is and is not. Fourteen hand-built loci are not a sample of English phrase-ends and two deterministic seats are not a sample of readers. The interval is a dispersion statistic over this item set, and every criterion in §6 is a decision rule about this item set — not an inference about English, about readers, or about models in general. No sentence this run writes may be phrased as one.

6. Predictions and criteria, registered

• M1 — the gate, now two-part (A9, round-1 critic finding 11, accepted). (i) interaction: Δ_RHYME ≥ +0.30 (RS-20260822b measured a full rhyme at a phrase-end at +0.578); (ii) direct: raw hit(aW) ≥ 0.60 and raw hit(aY) ≤ 0.35. Both must hold — the interaction alone can fail or pass on an arbitrary bW-against-bY difference. If M1 fails the task does not detect a rhyme and nothing else in this run is reportable as a finding.
• P1 — THE PRIMARY: does spelling add anything once the sound relation is held constant? Δ_ORTH − Δ_CONS. The lead's registered prediction, made before dispatch and recorded here so it can fail: ≥ +0.15 — the lead expects the look to add, because §7.26 item 5's probe suggested the seats read by eye.
• ≥ +0.15 in say, with the interval clear of zero → A14 (round-2 findings 1 and 2, accepted as far as they can be): the claim is an ASSOCIATION, not a decomposition. What may be said is: among pairs carrying the same rule-class sound relation to the fixed word, the ones that also look alike are registered more often here — whose causes this design cannot separate into orthography and residual within-N2 phonetic distance. The handbook consequence is still real and still narrow: a translator choosing between two N2 candidates has a reason to prefer the spelled-alike one. §7.26's word registered stands as written.
• |Δ_ORTH − Δ_CONS| < 0.10 → spelling contributes nothing measurable at a phrase-end. A4 (finding 5, accepted): this does NOT discharge §7.26 item 5. Under the sound-framed prompt it narrows the hedge to when a reader is attending to sound; whether it narrows further is what stage N answers, and the handbook may go no further than the neutral figure allows.
• Between 0.10 and 0.15, or an interval spanning zero → reported as indeterminate, no handbook change.
• P2 — is a pararhyme registered at all? Δ_CONS ≥ +0.20. New: §7.26's NEAR class pooled N1, N2 and N3 and no figure exists for N2 alone. This is the number a translator writing road / blade actually needs.
• P3 — the ITEM-SET sensitivity check (A8, round-1 critic finding 9, accepted, and its status is downgraded). Δ_ORTH − Δ_CONS computed on QR alone. QR is not a power statement about P1 and P2 — a different model reading differently proves nothing about theirs. It asks a narrower question that it can answer: can these materials produce an orthographic advantage in any seat, including one this project has measured reading by eye? If QR returns ≥ +0.15 while the reporting seats return < 0.10, the item set is shown able to express the effect. If QR returns < 0.10 too, the run reports that the item set produced no orthographic effect in the one further seat tested — one seat, one outcome, one contrast (A18, round-2 finding 11) — which is evidence about the items and is stated as such, never as a null about reading.
• P4 — WITHDRAWN as a claim (A2). The NEAR-EYE figure over E1–E3 is printed beside RS-20260822b stage P's 0.722 so the two can be read together, and nothing rests on it.
• P5 — strength verdicts. The full / half split by level, descriptive, as §7.26 item 4.
• P6 — bare assent (A5, round-1 critic finding 6, accepted in part). The ANSWER: yes rate, reported beside the naming-required hit, so that said yes and found the right pair are separable. The finding's stronger form — that visual string-matching rather than heard chime produces the effect — is the effect under test, not a confound, and is refused as one.

7. Controls

• C1 naturalness, TWO-SIDED (A6, round-1 critic finding 7, accepted). From stage R. It fires if aX and aV differ in mean naturalness rank by more than 0.5 in EITHER direction — awkwardness can make a phrase-end salient just as fluency can make a relation easy to accept. A15 (round-2 critic finding 6, accepted): if C1 fires, the primary is reported as disclosed-confounded and NO handbook sentence is written from it — a fired confound control may not coexist with a positive conclusion. That C1 does not cover the b cells, rates side-by-side rather than in isolation, and uses two of the responding seats as raters, is a cost decision and is a named limit.
• C2 mechanical. Every planted relation recomputed from tools/rhyme_pairs.py by verify.py and the 2 × 4 structure re-checked cell by cell: exactly one N2-run≥3 per locus at aX, one N2-run≤2 at aV, one STRICT at aW, every other cell NONE. Any mismatch and the run does not go.
• C3 unplanted phrase-end relations — a LIMITED diagnostic (A18, round-2 finding 11). It catches unplanted relations between adjacent phrase-end words and nothing else: not internal phonological priming, not semantic association, not shared morphology, not template attention. Every carrier is segmented by the rule of sweep.py and every unplanted adjacent phrase-end pair scored. Loci carrying one are named, and a registered sensitivity analysis recomputes every criterion with them dropped.
• C4 filler shape, descriptive only — it establishes nothing about equivalence. Mean filler length in characters and syllables by cell.
• C6 endpoint identification, added free by A18 (round-2 finding 9). The rate at which each cell's answer names the two planted phrase-end words, printed per cell, so that a failure of the b baseline or of endpoint identification in any cell is visible rather than assumed away.
• C5 floor. hit over the four b cells and aY.

8. Failure criteria — what would make this run worthless

1. M1 fails. No finding is reported.
2. Floor saturation: mean hit over the four b cells and aY > 0.60. Every Δ is then compressed against a ceiling and the run is withheld.
3. C1 fires on aX. The primary is disclosed as possibly confounded by wording.
4. P3 returns < 0.10 as well. The primary is reported as no orthographic effect in any seat, including a measured spelling-matcher — a statement about the item set — and never as a null about how readers read.
5. Dead cells > 5% of stage D. Reported with the caveat on every figure.

9. Contamination

The lead's English of the Gulistan span is the translation limb and no figure in this study is computed from it except the descriptive M-avail census, which is a hand-level figure and is never summed with the seats'. The carriers of §3 are constructed English written for this design and render no source. The lead is not the independent third translator of anything here.

The measurement was run after the log froze, per CLAUDE.md: tools/dependence_check.py, the lead's whole English of tales 20–25 against Eastwick 1852's text of the same six tales (Internet Archive OCR, $0). 2 shared 7-grams, 0 shared 12-grams, longest common run 7 tokens — clean. Both figures are reported, per the standing rule that run length alone is a poor proxy.

10. Budget

Pre-flight, built from max_tokens and not from an assumed output length (note (abc)). Stage D bodies are one sentence in and one short answer out.

Per-call rates are read from a 6-call probe before the run, never from config/models.md (S216). Declared de-scope, in order — RE-ORDERED BY A17 (round-2 critic finding 10, accepted). Stage N was first to go and is now last, because without it no claim beyond the exact sound-framed instruction may be made at all. If the probe puts the run above the ceiling: (1) drop QR from the 18 NEAR-EYE items, which carry no claim; (2) cut stage R to P1 alone, with C1 computed from one seat and reported as single-seat on the face of the control; (3) drop QR from stage D entirely, losing P3; (4) only then drop stage N — and if it is dropped, every claim is restricted to the exact sound-framed instruction and no sentence about ordinary or neutral reading is written. Nothing else changes: the estimands, criteria and interpretations above hold in every de-scoped version.

11. Warrant, to be carried on the face of every instruction (A11)

n constructed loci in one language, k model seats, no human reader, Tier D NOT PASSED. Two deterministic seats are not a sample of readers; the bootstrap resamples hand-built loci and creates no population-level support for a claim about readers. Every sentence this run puts into framework/v0.2 carries that line, in the form §7.26 already uses.