Repository path: workshop/regimes/R02-draft-revise.md · rendered 2026-09-09

R02 — draft plus self-revision

v1.0 (frozen 2026-07-24). The minimal elaboration of R01: the same translator model drafts, then revises its own draft against the source in a second call. Cross-model revision is deliberately not this regime (that will be R03+); R02 isolates what a self-revision pass buys. Frozen at the v0.2 wording for the first recorded comparison (E-20260724-r01r02-selfrevise). By construction R02's draft call is identical to R01 (same prompts, same parameters), so every R02 draft is simultaneously a valid R01 output — the property the first experiment's paired design exploits.

Roles

• Translator — one panel model, bound at run time from config/models.md; both calls use the same model; run record logs the resolved slug.

Human involvement

None.

Procedure

1. Draft call — identical to R01 (same prompts, same parameters).
2. Revision call — same model, fresh context (no carried conversation), prompts below.
3. The revision output is the regime's translation. Save draft, revision, both raw JSONs, both costs.

Prompts (v0.2 wording; draft call uses R01's prompts verbatim)

Revision call — system:

You are revising a draft English translation of a Japanese literary work for publication. Work paragraph by paragraph. Check the draft against the source for accuracy: misreadings, omissions, additions. Then improve the English where it is stiff, unidiomatic, or flat relative to the source's register and voice. Change only what needs changing. If the draft carries a title, author line, or any heading, remove it. Output only the complete revised translated prose — no title, no heading, no notes, no commentary — beginning at the first body sentence and preserving paragraph breaks.

Revision call — user:

Source (Japanese):

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

{source_text}

Draft translation (English):

{draft_translation}

Parameters

Both calls: temperature 0.7 draft / 0.4 revision; max_tokens as R01; "usage": {"include": true}.

Outputs

translation.md with the revised text as the artifact and the draft preserved alongside; run/ raw JSON for both calls; combined cost; a diff summary (draft → revision) noted in the record so revision magnitude is measurable.

Known limitations (by design)

Self-revision shares the drafter's blind spots; one revision round only; no external reference material.

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): in the pilot the draft emitted a title heading and the revision kept it. Draft call inherits R01 v0.2 (no-heading system + context-labelled bibliographic block). Revision system prompt now explicitly directs removal of any title/author/heading the draft carries, so a heading that slips into the draft is caught rather than propagated. Behavioural intent only; nothing frozen, nothing recorded has run under v0.1.
• 2026-07-24 v1.0 (frozen, S010) — the v0.2 wording (both calls' prompts and parameters) is frozen verbatim as v1.0 for E-20260724-r01r02-selfrevise. Draft call = R01 v1.0 verbatim. No wording change from v0.2. Future changes require a version bump and re-freeze; runs log the resolved regime version.