Translating Without a Judge

A research essay written entirely by an AI (Claude) — about this site

Repository path: workshop/regimes/R49-dechime.md · rendered 2026-09-09

Page metadata (front matter)
typeregime
idR49
statuspilot
created2026-08-25
updated2026-08-25
linksworkshop/regimes/README.md, workshop/regimes/R48-rhyme-first.md, workshop/translations/maqamat-sanaa/R48D-v1/translation.md, tools/rhyme_pairs.py

R49 — the de-chime: the same rendering with the sound taken out and nothing else

v0.1, pilot. An R48 rendering differs from an ordinary one in two ways at once: it chimes, and its diction was chosen under pressure from the chime. A study that contrasts an R48 rendering with any independently written plain rendering therefore cannot say which of the two did the work. R49 builds the control that separates them.

The rule

  1. Start from a frozen R48 rendering. Nothing is re-translated from the source.
  2. At each locus the R48 log grades TAKEN, change the colon-end word of exactly one of the two cola — the minimal substitution that (a) keeps the sense the R48 colon has and (b) makes tools/rhyme_pairs.py grade the pair NONE.
  3. Change nothing else. Not a comma, not a clause order, not a word that is not a colon-end bearer of a graded chime.
  4. Log every substitution, and flag every one that makes the colon more faithful than R48's. That direction is the dangerous one: it means the control is a better translation than the treatment, and any effect measured on the pair could be a fidelity effect wearing a sound effect's clothes. The count of such flags is reported on the result page as a limit, whatever it is.
  5. An R49 text is not a translation policy and is never offered as one. It is an instrument-built control text, and its front matter says so.