Repository path: workshop/regimes/R10-anchored-register.md · rendered 2026-09-09

R10 — anchored register target

v1.0 (frozen 2026-07-31, S074.) An R06 (lead, source-only, single pass) whose target register is a named Tier 1 naturalness anchor, supplied to the translator as that anchor's feature catalogue decomposed into numbered propositions and frozen before the source is read for translation. Every logged decision records which numbered proposition it acted on.

Why it exists

The project has three Tier 1 naturalness anchors and has never used one as a target. They were built as derivation aids — descriptions of what a register looks like — and every use of them so far has been as a comparand after the fact. R07 targets a rule set (Venuti's, turned round into instructions); R10 targets a catalogue of observed features of real prose, which is a different object: R07's rules are prescriptions with no exemplar, R10's propositions are descriptions with an exemplar attached.

The regime also makes a catalogue testable. If a translator can work to a catalogue and an independent reader cannot then find the catalogued features in the result, the catalogue is not usable as a target, and that is a measurement rather than an opinion. Two R10 runs over the same source unit against two different anchors form a matched pair: source, translator, session and length are held fixed, and only the target catalogue varies.

Roles

• Translator — the lead agent, as a labeled subject (charter §3, A4). No panel call, no API cost.
• Judge — never the lead (charter §5).

Human involvement

None.

Procedure

1. R06 §1 unchanged. Source only. The translator reads no published rendering of the same passage into any target language before the translation and its log are frozen. Priming, if any, is declared.
2. The target is named by anchor id, and its catalogue is decomposed into numbered propositions which are frozen and committed before the source is read for translation. The commit is the freeze. A proposition may not be added, deleted or reworded once translating has begun; a proposition discovered to be missing is recorded in the log as missing.
3. Translate once, straight through. No return pass.
4. The log is written as the draft is written, decision by decision.
5. Every logged decision carries a target code, assigned at the moment the decision is made: - A (aimed) — the decision was taken in order to satisfy one or more of the numbered propositions. The proposition ids are recorded. - X (contrary) — a numbered proposition bore on the decision and was not followed, because something else outweighed it. The proposition ids and the reason are recorded. - N (none) — no proposition in the target list bore on the decision at all. A decision is logged when the translator was aware of two or more live renderings at the moment of writing. The log therefore under-counts total choices by construction, exactly as R07 §5 does, and the code rates are rates over contested sites only.
6. The translator additionally records, before translating, which propositions have an OPPORTUNITY in this source unit — i.e. whether the source supplies any material on which the proposition could be satisfied or violated at all. This list is frozen with the log. It exists so that a proposition found absent from the result for want of material is not confused with one the translator failed to install.
7. Freeze before anything else happens. Commit the translation, the log and the opportunity list before any comparison, scoring or evaluation is designed.
8. Contamination declared with a one-line basis, measured where a freely reachable published rendering exists (CLAUDE.md, standing rule).
9. File under workshop/translations/<work>/R10-<anchor-tag>-v<N>/.

Outputs

The translation, the decision log with target codes, the frozen opportunity list, the contamination measurement, and the target anchor id and claim-file hash.

What R10 is not

• It is not a claim that a translation should aim at any particular register. A-mchugh-presence says in its own text that it does not license "maximize naturalness", and A-mansfield-garden-party says its naturalness is period-relative. R10 executes a target; it does not endorse one.
• It is not calibrated. Tier D has not passed. A translation produced under R10 is a labeled subject, not a demonstration that the target was worth aiming at.
• The pairing is the point. A single R10 run measures nothing: the translator's own report that it installed a feature is internal-judgment-only. The matched pair against two anchors is what makes the target codes falsifiable, because a feature installed for one target and not the other should separate the two texts to a reader who has never seen either.

Change log

• 2026-07-31 (S074) — created and frozen at v1.0, for the matched pair T-karen-R10c-v1 / T-karen-R10p-v1 (E-20260731f-catalogue-reach).