Repository path: workshop/regimes/R27-footing-max.md · rendered 2026-09-09

R27 — footing-max

v1.0 (frozen 2026-08-14, S186.) An R06 (lead, source-only, single pass, no return pass) whose specification instructs the translator to attempt, at every mechanically identified site where the source grades social footing, an English device that marks the same relation in the same direction — and to record, site by site, whether the attempt succeeded and what it cost.

R27 is a ceiling probe, not a recommended way to translate. Its output is expected to be worse English than its own baseline. The quantity it exists to produce is not a translation but a price list.

Why it exists

framework/v0.2 §10 rests a practitioner instruction on five hands — two published human translators, two 2026 language models, one lead rendering — and every one of them was translating without being asked to carry the source's grammatical footing. The section's sentence is therefore ambiguous in a way that matters to a practitioner:

• (a) English has no device, so the marking cannot be carried; or
• (b) English has devices, and translators, weighing them against fluency, decline to spend them.

A census of hands that were not trying cannot separate (a) from (b), however many hands it has. Only a hand that tries can, and R27 is the specification of trying. ARM-carrier failed twice to find a positional predictor and ARM-footing a third time (framework/v0.2 §10.2); none of those failures asked a translator to spend the mark deliberately.

What it does not license

An R27 output is one hypothesis-aware hand and is never evidence about how English works. It is evidence about reachability: a site carried under R27 is a site where a device existed and was found, and a site abandoned under R27 is a site where this translator, told to try and free to be ugly, could not find one. Both are facts about one attempt. A framework sentence may cite an R27 log for what was reachable at what price; it may not cite it for what English does, for which the hands that were not trying remain the only evidence.

The site list is the manipulation, not the measure. Note (bio) governs: a specification that restates the measure manufactures the result. R27's specification names the sites, so no primary of any design may be a count over those sites judged by the translator. The log's codes are self-report and are labelled as such.

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. The baseline first. An R06 rendering of the identical source span, by the same translator in the same session, is written and frozen and committed before the R27 rendering begins. The pair is the unit; a lone R27 output has nothing to price against.
2. The site list is mechanical and is frozen before the baseline. A committed pattern list is run over the source and its output is written to a file whose rows are not read until the baseline is frozen. Site selection is therefore not a translator's judgment in either direction. The extractor prints totals only.
3. Source only, as R06 §1: no published rendering of the same passage in any target language is opened before both members of the pair and both logs are frozen. Priming declared.
4. Read the site list. Then translate once, straight through. No return pass, as R06 §2.
5. At every site, attempt an English device that marks the same relation in the same direction — the referent elevated, the speaker lowered, the addressee deferred to. The attempt is made even where the result is ugly, archaic, or long. Fluency is subordinate to the attempt; that is the regime.
6. No apparatus. The carriage must be in the prose: no footnote, no gloss, no bracket, no translator's note, no capitalised honorific invented for the occasion that the English of the rendering does not otherwise support.
7. Code every site in the log, at the moment it is written, with the English named: - C-n — carried, at no cost: the device went in and the translator would defend the sentence as ordinary English saying what the source says. - C-p — carried, at a price: the device went in and something was paid. The price is written out — what became stilted, longer, more archaic, or less accurate, and how much. - A — attempted and abandoned: the English that was tried is written out verbatim in the log, with the reason for abandoning it. - N — no attempt possible: no candidate device came to mind at all. N is not A; the distinction is the whole point, and a translator that codes N where it means "I could not face it" has broken the regime.
8. The counts of C-n/C-p/A/N are self-report and are labelled internal-judgment-only wherever they appear.
9. Freeze. Commit the rendering and its log before any comparison against the baseline is written up, and before any published hand is opened.
10. File under workshop/translations/<work>/R27-v<N>/, with sites.py and sites.tsv beside it.

Outputs

translation.md with its coded log, plus sites.py and sites.tsv. No run/ — no API call produced it. Cost $0; never ledgered (charter §6, A4).

Known limitations (by design)

• Everything R06 lists, and its no-return-pass floor undiluted.
• One hand, hypothesis-aware. The translator wrote the specification. There is no defence against its own expectations beyond the frozen baseline and the mechanical site list, and neither is a defence against them at the level of how hard it tried.
• C-n versus C-p is a judgment, and it is the translator's own. A blind panel shown the two renderings could measure the price independently; that is a separate design and is what an R27 pair is for, not something the log settles.
• The regime rewards effort, and effort is unmeasured. A translator who tries harder carries more. The log therefore prices this attempt, and a later attempt that carried more would not refute it.
• A and N are both failures and only one of them is informative. N says nothing was available; A says something was and was rejected. Where the honest code is uncertain, the regime requires A with the tried English written out, so the reader can disagree.

Changelog

• 2026-08-14 v1.0 (frozen, S186) — created for ARM-honorific-hands step 2, to give framework/v0.2 §10 the one thing its five-hand census could not supply: a hand that tried. Frozen on creation, because the pair it was minted for is written in the same session (README.md's rule would otherwise bar the comparison it exists to make — same reasoning as R06 v1.0).