Repository path: workshop/regimes/R05-serial-long-work.md · rendered 2026-09-09

R05 — serial long-work translation (source-only, cumulative)

v0.1 (draft, 2026-07-27.) A variant of R04-lead-close for a work too long to translate in one session. Everything in R04 holds unless this page overrides it; the overrides are all consequences of one fact — the translation is frozen in pieces, and the pieces are separated by weeks and by other work.

R05 exists because R04's central discipline is a single freeze: draft, self-revise, write the log, freeze, and only then let anything downstream see it. That discipline has no opinion about what happens when there is a second sitting, and the differences turn out to be structural rather than administrative.

What R04 supplies unchanged

• Roles. Translator is the lead agent, as a labeled subject; the lead never judges its own translation.
• Source only. No published rendering of the work into any target language is read before the relevant span and its log are frozen. Priming, if it occurs, is declared on the artifact.
• Draft, then self-revise against the source, within the sitting.
• Log at translation time, decisions and alternatives, not quality claims.
• Contamination declared with a one-line basis.
• Cost $0, never ledgered.

What R05 adds

1. Spans are defined by index into a stored source, before translating. The whole cleaned source text is committed once, and each span is named by paragraph range into that file. A span boundary is chosen at a break in the narrative, not at a word count. This makes the unit of work reproducible and stops the source from drifting under the translation across sessions.
2. The artifact is append-only. One translation.md per work per regime-version; each span is added, and nothing already frozen is revised except through a numbered, dated erratum at the foot of the page. A silent repair of span 1 while translating span 4 would destroy the only evidence the arm is after — what the translator did before knowing what was coming.
3. The log is cumulative and continuously numbered. Decisions are D1, D2, … across the whole work; span 2 begins where span 1 stopped and nothing is renumbered. The log is the object of study, and a log restarted each sitting cannot show a decision being made in span 1 and paid for in span 4.
4. A binding register is maintained — register.md in the work directory — carrying the terminology, naming and voice decisions that bind later spans, each citing the log decision that created it, plus an explicit unresolved section listing questions the next span is required to settle. Revisions to the register require an erratum. This is the instrument the short-unit regimes never needed, and producing it is part of the translation work, not bookkeeping around it.
5. A cadence, not a deadline. The arm declares how often the work must be visited; under-visiting is the alarm. R05 does not specify the interval — the arm does.
6. Decisions made by drift are recorded as such. Where a policy was adopted while drafting rather than chosen (register V4 in the first instance), the log says so. This class of entry is only visible in serial work: in a single sitting every such policy is still in view and gets tidied before the freeze, which is precisely how the short units have been hiding it.

Outputs

translation.md (append-only, cumulative log, errata section), register.md, and the stored source. No run/ — no API call produced any of it.

Known limitations (by design)

• Everything R04 lists, in particular that self-revision shares the drafter's blind spots.
• The freeze is only as good as the discipline. Nothing mechanically prevents a later session from editing an earlier span; the erratum rule is a convention, and a session that breaks it leaves no trace except in the diff. A later instrument could check the earlier spans byte-for-byte against the commit that froze them, and probably should.
• Contamination is measured once, at selection, and then assumed. The gate runs on the first span. A later span could sit on a much more heavily reproduced passage — a famous scene, an anthology piece — without anything re-firing. ARM-longwork should re-measure at least once mid-work.
• The register can become an alibi. A decision written into the register looks settled, and a later span that ought to reopen it has an instrument telling it not to. The unresolved section is a partial answer; it is not a complete one.