Repository path: workshop/translations/README.md · rendered 2026-09-09

Translations

Translations are first-class artifacts (charter §3). This directory is also where Tom reads the project's prose (charter §2.3, A5): it is catalogued in workshop/translations/INDEX.md (generated by tools/build_index.py), and every session that translates also quotes an excerpt in the journal.

Layout: workshop/translations/<work-slug>/<regime>-v<N>/ containing

• translation.md — front matter type: translation, id: T-<work>-<regime>-v<N>, translated-by: (lead or a panel role) and contamination: (none | suspected | high), plus: source pointer (manifest or anchor), regime id and version, resolved models where any were used (provenance), date, cost ($0 for lead translation), lineage, evaluation pointers; then the translation text itself.
• run/ — full raw request/response JSON for every API call that produced it. Lead translations have no run/; the artifact and its log are the record.

The translator's log (required)

Every translation.md carries a ## Translator's log section, written at translation time and frozen before any evaluation of that translation is designed (charter §3, A4). A log written after seeing scores is worthless, which is what the freeze protects.

It records decisions and alternatives, not quality: what was hard, which options were live, what was chosen, what was given up. Any evaluative sentence in it carries internal-judgment-only. The lead never judges its own translation — the non-Anthropic panel does, blind, with authorship stripped (charter §5).

The log exists because it is evidence the project cannot get any other way: process observation from inside a translator, at no cost. It feeds the typology (which senses actually collided), the theory (do the claimed regularities show up as felt problems?), and eventually the framework, whose recommendations should come from documented decisions rather than plausibility.

Contamination

Declare it per work, with a one-line basis. A lead translation of a heavily-translated famous work is a practice artifact, not an independent test of translating ability — published versions plausibly sit in the lead's training data (the calibration run measured 93% work-recognition on Botchan's opening). The canon manifests already carry a contamination axis (D-20260723-02 criterion 6); use it. Where a comparison carries weight, prefer low-contamination sources. Never prime a workshop translation with a stored published rendering of its own source — translate from the source alone.

Retranslation of a canon work under a new framework version is core activity, not duplication — new version directory, lineage recorded. Evaluations live with the experiment or evaluation record that produced them and are linked from the translation page; every evaluation names its goodness senses and carries anchors or internal-judgment-only, plus provisional: true while jury calibration is unpassed.

Translations of Tom-provided texts do not go here — they stay inside private-texts/ (charter §7).

Binding condition on the translator's log, absorbed from wiki/backlog.md at S074 (row opened S064, review-or-retire fired)

Any tally over a log's rows is computed by a script that parses the log, and the script is committed with the tally. Note (bey): three of five R07 tally lines were wrong on first writing, every one was caught by parsing and none by rereading, and S064's own pre-registration quoted a 21-row denominator for a 44-row table. A translator's-log table is a data structure, and this project kept reading it as prose.

S074's R10 pair did this without being asked — the A / X / N target codes of both logs were extracted mechanically by analysis/checks.py, and the extraction is asserted by an independent verifier that re-reads the frozen log — which is the shape the backlog row wanted and is now the rule.