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

Regimes

A regime is a fully specified way of producing a translation (charter §3): single-pass, draft-plus-revision, multi-agent role structures, skopos-conditioned variants, retrieval-augmented variants, and human-in-the-loop designs written as protocols. Comparing regimes is the workshop's basic experiment.

One file per regime, id R##, type: regime, versioned in-page (v0.1, v0.2, … with changelog). A regime spec must be executable without further judgment calls:

• Roles — by panel role, never by model slug (binding happens at run time from config/models.md; the run record logs the resolved slug as provenance).
• Prompts verbatim — frozen; any change bumps the version.
• Parameters — temperature, token limits, sequencing.
• Procedure — every step, including what is carried between steps.
• Outputs — what is saved (always: full raw request/response JSON, cost, resolved models).
• Human involvement — where a human may enter, what they see, what they may change; or "none".

Status frozen means citable in experiments; draft regimes may be piloted but not compared for the record.

Standing caution on within-pair contrasts (S082, from RS-20260801e-lead-carryover). When one continuous agent produces both arms of a regime pair in one session, the arms are not independent renderings, and the measured departure widens the gap: on the two matched pairs where it could be measured, the second arm sat two to three times closer (shared 7-grams) to an independently written first arm than to the arm beside it. Any regime contrast measured on a same-session pair is therefore an upper bound on what the regime change buys, and a design or result quoting such a contrast states this. (An unregistered typicality check complicates the size, not the direction — see the result's §3.)