Repository path: workshop/regimes/R01-single-pass.md · rendered 2026-09-09

R01 — single pass

v1.0 (frozen 2026-07-24). The baseline regime: one model, one call, whole text, no revision. Every comparison needs this floor. Frozen at the v0.2 wording (piloted, prompt-tightened) so the first recorded comparison (E-20260724-r01r02-selfrevise) runs against a fixed spec.

Roles

• Translator — one panel model, bound at run time from config/models.md; the run record logs the resolved slug and version date.

Human involvement

None. (The regime family for human-in-the-loop variants will branch from R03+, not here.)

Procedure

1. Assemble the source: full text from the canon directory (source.txt), plus title, author, year from the manifest.
2. One chat-completion call, prompts below verbatim, parameters below.
3. Save translation + full raw request/response JSON + cost into the translation record (workshop/translations/ layout).

Prompts (frozen at version bump; v0.2 wording)

System:

You are translating Japanese literary prose into English for publication. Produce a complete, faithful, publishable English translation. Output only the translated prose itself: no title, no author line, no date, no heading of any kind, no notes, no commentary. Begin directly with the first sentence of the body text and preserve the source's paragraph breaks.

User:

Translate the following work into English. The bibliographic details below are provided only as context — do not reproduce them in your output. Translate the body text that follows them.

Title: {title}
Author: {author}
First published: {year}

{source_text}

Parameters

• temperature 0.7; top_p default; max_tokens ≥ 2.5 × source character count (headroom for reasoning-style models); "usage": {"include": true}.
• Single attempt; a truncated or refused output is recorded as a failed run and rerun once, with both attempts kept.

Outputs

translation.md (per workshop/translations/README.md) + run/ raw JSON + cost.

Known limitations (by design)

No self-checking, no terminology memory, no skopos conditioning. The regime exists to measure what one pass buys.

Changelog

• 2026-07-23 v0.1 — initial spec (draft). Piloted only; freeze before first recorded comparison.
• 2026-07-24 v0.2 (draft) — prompt tightening from pilot observation 3 (E-20260723-pilot-kumonoito, title-heading leak). System prompt now forbids title/author/date/heading explicitly and directs the model to begin at the first body sentence; user prompt labels the bibliographic block as context-not-to-be-reproduced. Behavioural intent only; no wording is frozen and nothing recorded has run under v0.1. R02's draft call inherits this prompt.
• 2026-07-24 v1.0 (frozen, S010) — the v0.2 wording (prompts and parameters above) is frozen verbatim as v1.0 for the first recorded comparison E-20260724-r01r02-selfrevise. No wording change from v0.2; this bump only removes the "draft" status so a comparison can be recorded against it. Future changes require a version bump and re-freeze; runs are logged with the resolved regime version.