Repository path: workshop/regimes/R14-matched-flattening.md · rendered 2026-09-09
Page metadata (front matter)
| type | regime |
|---|---|
| id | R14 |
| status | pilot |
| created | 2026-08-03 |
| updated | 2026-08-03 |
| links | workshop/regimes/README.md, workshop/regimes/R04-lead-close.md, workshop/translations/odnazhdy-osenyu/R14-v1/translation.md, workshop/experiments/E-20260803f-craft-carriers/design.md, framework/tierD-repaired-rules.md |
R14 — matched-content flattening (a contrast-subject regime)
v0.1, pilot, first used at E-20260803f. A regime for producing the pole this project has never
had: a translation that is competent and dead. Tier D's damage operators produce translations
that are wrong — invented details, dropped negations, wrong referents — and a jury that separates
those has learned to detect error. Nothing in the project's vocabulary produces a translation with
no propositional defect whatever and none of the qualities that make prose live.
R14 exists because the boundary between those two things is where "good translation" actually sits, and because a contrast subject built by an operator that only touches non-propositional properties is the only way to ask what those properties carry.
What it is applied to
An existing frozen R04 (or R06) rendering by the same translator, whose translator's log is
already frozen. R14 never renders from the source. It renders from the live English.
The operators, and the one rule that binds all of them
Nothing may change what the text says. No referent, sense, negation, quantity, tense, modality,
or item of content may be added, removed, or altered. The test is propositional, and it is checked by
an independent seat before any other measurement is taken (E-20260803f §Controls) — a run whose
equivalence control fails is a run about damage, not about craft, and reports itself as such.
| id | operator | what it removes |
|---|---|---|
| F1 | cadence levelling | sentence-length variance; long periodic sentences broken into declaratives of similar length |
| F2 | figure de-specification | a figure replaced by a literal statement of the same proposition |
| F3 | repetition flattening | anaphora, triples, and repeated words collapsed to one statement of the content |
| F4 | verb deadening | a concrete physical verb replaced by a generic verb plus an explicit statement of the manner it carried, so the manner is stated rather than enacted. Amended at E-20260803f pass 1 (2026-08-03): the first wording said "of the same denotation", and on the frozen text it did not hold — the waves splashed became the waves were moving, which deletes the splashing. A deadening operator that deletes is a damage operator |
| F5 | register levelling | exclamation, apostrophe, oath, address to the reader, and voiced fragments restated as plain declarative |
| F6 | connective explicitation | parataxis, dashes and ellipses replaced by explicit subordinators that state the relation the original leaves the reader to feel |
F5 changes illocutionary force while preserving propositional content, and that is a declared limitation, not an oversight. An exclamation restated as a statement says the same thing and does it differently; the equivalence control is therefore instructed on propositional content specifically, and any run using R14 must say so.
Procedure
- The
R04log is frozen and committed first. R14 may not be planned into it. - Apply the operators, recording every site: paragraph, operator id, the live text, the flat text.
- Report the length ratio. A pre-registered tolerance is declared before the operators are applied;
length is a known confound in this project's damage work (
RS-20260802-tierD-verdict: +4.5% declared). - File under
workshop/translations/<work>/R14-v<N>/withtype: translation, the site table as the translator's log, and the lineage pointing at the live text.
Known limitations (by design)
- The operator is the lead's own taste. What R14 removes is what the lead believes carries
literary life. If that belief is wrong, R14 removes something else, and the run measures that
instead. This is why the primary measurement in
E-20260803fis whether independent readers name the same categories, and not whether they prefer the live text. - It cannot be applied blind. The same agent wrote both texts.
- One text is not a validation of the operator set. Six operators on one 600-word locus in one language pair.