Repository path: workshop/regimes/R04-lead-close.md · rendered 2026-09-09

R04 — lead close translation (source-only)

v1.0 (frozen 2026-07-27, S041). The regime under which the lead agent translates in session as a labeled subject (charter §3, §5, amendment A4). It exists because R01/R02 are API regimes — they specify prompts, temperatures and max_tokens — and none of that applies to a translation the lead writes itself. R04 specifies the discipline instead, which is the only thing that is actually constrained.

Frozen at v1.0 because the condition its own changelog set has arrived: v0.1 said "Not frozen: freeze before the first experiment that compares R04 against another regime," and E-20260727b-arm-identifiability is that experiment. Five translations were filed under v0.1 (genji-yomogiu, black-cat, rayo-de-luna, son-makara, signal, mare-au-diable); v1.0 changes no wording that applies to them retroactively and adds one requirement going forward (§Procedure step 2a).

R03 remains reserved for the multi-agent regime sketched in wiki/program.md Slate D.

Roles

• Translator — the lead agent, working directly, as a labeled subject. No panel call, no API cost.
• Judge — never the lead (charter §5). Evaluation, if any, is a separate design and a separate artifact.

Human involvement

None. (The human-in-the-loop variants of charter §1 are specified separately and are never run live.)

Procedure

1. Source only. The translator reads the source text and whatever non-translational apparatus it needs (dictionaries, commentary on the source language, the source's own critical text and collation notes). It does not read any published rendering of the same passage into any target language before the translation and its log are frozen. Where a parallel modern rendering exists in the same location as the source text, the passage selected must be one whose rendering has not already been seen; if any priming did occur, it is declared on the artifact.
2. Draft, then self-revise against the source in the same session. 2a. The draft is written out and frozen as its own artifact before the revision begins (added at v1.0). The frozen draft is an R06 output — R06 is this step stopped before the self-revision — so an R04 run always produces a paired R06/R04 comparison at no extra cost, the way every R02 run produces an R01 output. A draft reconstructed after the fact, or never separately committed, is not an R06 output and may not be used as one.
3. Translator's log written at translation time, recording decisions and alternatives — not quality claims — and frozen before any evaluation of the translation is designed.
4. Contamination declared on the artifact with a one-line basis: whether published translations of this work plausibly sit in the lead's training data.
5. File under workshop/translations/<work>/R04-v<N>/ per that directory's README.

Outputs

translation.md only. There is no run/ — no API call produced it, so the artifact and its frozen log are the complete record. Cost $0; never ledgered (charter §6, A4).

Known limitations (by design)

• Contamination is usually unquantified. For a heavily translated work the lead cannot establish that its rendering is independent of remembered published ones, and must not claim it is. R04 output on a famous work is a practice artifact and a source of process evidence, not an independent measurement of translating ability.
• Self-revision shares the drafter's blind spots — the same limitation R02 carries, for the same reason.
• No external reference material on the target side; naturalness is whatever the lead's own target-language competence supplies.
• The log is self-report. It records what the translator noticed itself deciding. Decisions made without noticing them do not appear, and nothing in the log is evidence that the decisions were good ones.

Changelog

• 2026-07-25 v0.1 (S017) — created, draft, for T-genji-yomogiu-R04-v1. Not frozen: freeze before the first experiment that compares R04 against another regime.
• 2026-07-27 v1.0 (frozen, S041) — frozen on the condition v0.1 set: E-20260727b-arm-identifiability is the first experiment comparing R04 against another regime (R06, lead single pass). One substantive addition, §Procedure step 2a: the draft is frozen as its own artifact before revision, which makes every R04 run a paired R04/R06 comparison by construction. No wording change affects the five translations filed under v0.1, none of which froze a separate draft and none of which may therefore be used as an R06 output.