Repository path: workshop/translations/maqamat-hulwan/R48D-v1/translation.md · rendered 2026-09-09

«المقامة الحلوانية» — the R49 de-chime of the R48 rendering

This is not a translation policy and is never offered as one (R49 rule 5). It is an instrument-built control text: T-maqamat-hulwan-R48-v1 with the sound taken out and, as far as a rule can enforce it, nothing else.

Why it exists. E-20260827b's pre-run critic returned NEEDS-REDESIGN from both seats, and one of the blocking findings was that a preference shift measured on RHY against ORD cannot be attributed to the chime, because the two arms differ at every translation decision the chime forced as well as at the chime itself. The critic's remedy — "construct controlled stimuli that vary only the targeted English end-rhyme while holding wording as constant as feasible" — is what R49 was built for at S222, and this is its second execution.

Built by the lead, in session, at $0, from a R48 rendering frozen and committed at 86d1136b before this text existed. Not ledgered (charter §3, A4).

The rule, and how rule 3 is enforced rather than promised

../dechime.py reads substitutions.tsv — one row per changed colon — and fails if any colon not in the table differs from R48, if any substitution is a no-op, if any chime the R48 grading found survives, or if the de-chime creates a chime that was not there. It exits 0.

49 colons, not 42. Seven loci needed a second substitution because the machine grading found a chime spilling across a locus boundary or a run deeper than two cola, which R49 rule 2 — written for two-colon loci — does not cover. That is a declared extension of the rule, made before any data call, and every one of the 49 rows is in the table.

Four substitutions were made twice. The first pass left two chimes standing (examination ⁄ consideration, men ⁄ woven) and created two that were not in R48 (meet ⁄ brought, well ⁄ people) — which is the failure mode the checker exists to catch, and it caught all four before anything was dispatched. A hand-checked de-chime would have shipped them.

Twelve of the 49 substitutions move more than two tokens, against R49 rule 2's minimal substitution; the longest moves seven tokens out and eight in (p085). Reported here because the count is a bound on how well "nothing else changed" holds.

R49 rule 4 — every substitution that makes the colon MORE faithful than R48's

This is the dangerous direction: a control that is a better translation than the treatment turns any measured effect into a fidelity effect wearing a sound effect's clothes. Eight of the 49 are flagged, and all eight are places R48's own log had already declared as leaning:

Six of the eight are the four leans R48's own frozen log listed plus two of its chime-driven word choices — that is, the de-chime, applied mechanically to break sound, walked back exactly the places the translator had already written down as the price of sound. Nothing was found that the frozen log had not already declared.

The consequence for E-20260827b, registered on its design page: the placebo pair RHY-D vs ORD is, at eight colons, a more faithful text opposed to ORD than RHY is. If the disclosure conditions move preference toward RHY and not toward RHY-D, that shift is running against a fidelity advantage the control holds, which is the conservative direction.

The text

../R48D-v1/cola.txt, generated by ../dechime.py; the substitutions are substitutions.tsv; the machine grading is grades.json. The verse is identical in every arm and is excluded from every study segment.