Repository path: workshop/experiments/E-20260820-rhyme-bearer/predictions-frozen.md · rendered 2026-09-09

E-20260820-rhyme-bearer — predictions, frozen before either comparator was opened

This file exists to be time-stamped. The step-2 pre-run critic of this arm returned BLOCKING 1 because that step's loci and predictions had been written after the published English had been read, which cost the run its confirmatory status. This file is committed before Lane 1839 or Burton 1885 is opened for span F, and the full design is written after. The loci themselves were frozen earlier still, in the span-F translator's log at commit c2dd9bc, before any English for them existed at all.

The claim the translating produced

Span F's fifteen rhyme loci split almost perfectly by where the rhyme sits: in quoted verse six of seven crossed into English; in rhymed prose one of seven did. The lead's explanation, written into D70's section of the log, is not verse is easier but something narrower:

An Arabic sound figure crosses into English when the words carrying the rhyme are not the words carrying the denotation. المداري and الصواري — winnowing-forks and ships' masts — are in the same sentence because they rhyme; to rhyme them in English the translator must rename the things, and to name them he must lose the rhyme. Where the rhyme rides on a verb ending, a participle, an abstract or a grammatical suffix, nothing is named and the translator is free.

Verse-versus-prose is confounded with this in span F but not identical to it, and the two come apart at three of the fifteen loci: F3 is a verse locus whose rhyme falls on a named object (my palm) and was lost; F7 is a verse locus whose rhyme falls on a proper name and was carried; F11 is a prose locus whose rhyme falls on rewordable material and was carried. The bearer of the rhyme, not the genre of the passage, is the registered predictor.

Predictions, registered now

Each is stated so that it can fail.

P1 (primary). Pooled over the published hands, carriage of an audible English sound figure is higher at loci whose Arabic rhyme falls on rewordable material than at loci whose Arabic rhyme falls on named objects. One-sided Fisher exact, α = 0.05.

P2. The same direction holds within each hand taken separately, Lane and Burton alike. If it holds in one hand and reverses in the other, P1 is reported as unsupported whatever the pooled test says.

P3. Where a hand does produce a sound figure at a named-object locus, blind readers will code its English as substituting different things for the ones the Arabic names, more often than where it produces one at a rewordable locus. This is the mechanism, and it is the prediction the lead would most like to be true, so it is stated separately and is not permitted to rescue P1.

P4. Lane carries at fewer loci than Burton. Span A measured Lane at 0 of 7 and Burton at 2 of 7 with 1 partial (RS-20260814-saj-carriage), and span C's confirmatory run ran the other way at the loci where the Arabic repeats a sound (Lane 3 of 4, Burton 1 of 4). The record is therefore contradictory and P4 is the weakest of the four; it is registered so that a third measurement is not chosen after the fact.

P5 (the lead's own text is not evidence). The lead's R05 rendering is censused alongside the two published hands, blind and with authorship stripped, and is excluded from every test above. It was written under S1, which is the hypothesis's ancestor, so its split cannot confirm the hypothesis. It is reported for description only, and any agreement between it and the published hands is reported as such and never as support.

Failure criteria, registered now

1. The coding gate. The predictor is a classification of what each Arabic rhyme falls on. Two panel seats classify all fifteen loci from the Arabic alone, with no English shown and no hypothesis stated. If the two seats agree on fewer than ten of fifteen, the primary is not run and the result is reported as a coding failure. RS-20260816-era experience (S201) is the reason: the lead's own list of an Arabic passage's sound figures agreed with three expert readers on only half of it, and here the coding is the predictor.
2. Cell size. If either predictor cell holds fewer than four loci after the gate, the primary is not run.
3. Absence. A locus with no locatable counterpart in a hand is coded ABSENT from the text by the lead and is dropped for that hand, not scored as a failure to carry. Lane abridges; counting his abridgement as a refusal to rhyme would manufacture P4.
4. One re-dispatch. A body that returns unparsable or truncated is re-dispatched once at double the token cap and, if it fails again, written dead and never re-bought. Dead bodies are reported with the result.

What would make the lead wrong

If carriage is flat across the two predictor cells, the split the translating produced is about verse licence — the freedom to reorder a bayt — and not about what the rhyme is riding on. That is a different and duller finding, and it is the one the design is built to be able to return.

Frozen 2026-08-20, before Lane 1839 (PG #34206) or Burton 1885 (PG #3435) was opened for this span. The full design, materials and procedure follow in design.md.