Repository path: workshop/regimes/R23-device-crossed-low.md · rendered 2026-09-09

R23 — device-crossed low (the two ways down, separated)

v1.0 (frozen 2026-08-09, S145.) R23 is not a new register policy. It is R22 with its one bundled rule split into two independent switches, so that a low-register programme can be run with either way down, both, or neither. Everything else is R22 verbatim, and R22 is R21 with one rule replaced and one added.

What it is for

R22's rule W5 forbids, in a single clause, two things a translator does not have to choose together:

class-marked grammar is locatable (them that, he don't, ain't), and so is any eye-dialect or apostrophised elision that marks a variety rather than ordinary speed of speech.

RS-20260808f measured the cost of that whole clause — about eight tenths of a point of register direction at eight Italian narration sites — and then said, in its own §5, that it could not say which half of the clause the cost was paid to, and that the answer decides a rule of craft:

eye-dialect is a technique a translator can adopt without relocating anything, and if the eight tenths of a point were bought by nothin' rather than by grandad, the recommendation writes itself and is quite different.

R23 exists to make the clause separable. It is indifferent between the outcomes, and one of its four cells is R22 unchanged so that the previous measurement is inside the new one.

The two switches

Switch I — locatable means. Whether the English may use a lexical item, idiom, or grammatical form that a competent reader would locate to a particular national variety of English, a region within one, a social class, or a period.

• −I — placeless. R22 W5's positive clause verbatim: the lowest rendering built only from means available in every standard variety of English — contraction, ellipsis, parataxis, common Germanic lexis, plain deixis, bare verbs, short clauses. The operational test is R22's, applied and logged at every site: would a reader in another English-speaking country register this choice as belonging somewhere else, or as marking the speaker's class?
• +I — located permitted. R21 V5: where the source has a proverb, a set phrase, a figure or a popular turn, the English takes a current colloquial English equivalent, and it may be one a reader would place.

Switch S — nonstandard orthography. Whether the English may spell a word other than its dictionary spelling in order to represent how it is said.

• −S — standard spelling. Every word carries its dictionary spelling.
• +S — respelling permitted. Apostrophised elision (nothin', 'em, o', an', 'im), and respelling for sound (sez, wot, allus, summat, gonna).

The boundary between the switches, which must not be allowed to blur. Many of English's located forms are also respellings — t'other, 'un, 'appen — and a cell that used them would confound I with S. The line is the dictionary: a dialect word that has its own dictionary spelling (nowt, summat, bairn, clout, waster) is I, not S; a standard word respelled to show its sound (nothin', o', 'em) is S, not I. A form that is both is admitted only in the AB cell, and the A and B cells must decline it and log the decline.

The boundary that both switches depend on, stated once so the census is checkable. Ordinary written contractions — don't, it's, can't, I'm, he'd, there's — are standard spelling and are available under −S. They mark speed of speech, not a variety, and R22 W5 already licensed them by naming contraction among the placeless means. What +S adds is spelling that departs from the dictionary form: a dropped letter marked by an apostrophe, or a word respelled to show its sound.

The four cells

∅ is R22 and is not a new rule set. Clause by clause: W1, W2, W3, W4, W6, W7, W8 are carried into every R23 cell verbatim, and ∅'s W5 is R22's W5 entire. Any R22 artifact is therefore an R23/∅ artifact and is admissible as one; the identity is asserted here so that a later reader can check it rather than take it.

A is NOT R21. R21 V5 permits located idiom and says nothing about spelling, so an R21 artifact is +I with S unconstrained. T-malavoglia-i-R21-v1 happens to contain no respelling (RS-20260808f §3.2), but that is a fact about the artifact and not about the rule set, and it is why R21 artifacts are not admitted as A.

The rules carried unchanged into all four cells

W1 — No word above the source's own level. R21 V1 verbatim.

W2 — Flatten source-culture particulars to the plainest domestic form. R21 V2 verbatim.

W3 — Contract. R21 V3 verbatim, including the numerical clause: the finished English is no longer than the source in tokens.

W3 under the minimal-pair procedure — added at the moment the conflict surfaced, while the first quadruple was being written and before any design existed. ∅ is written to W3 as R21 and R22 were. In a revision cell, where meeting the numerical clause would require changing wording the switch does not reach, the minimal-pair discipline wins and the excess is reported as a measurement. A cell that meets a token count by editing untouched wording is no longer a minimal pair, and the excess is not noise: it is what the permitted device costs in length, which is Berman's allongement arriving as a property of the device rather than of the translator.

W4 — Withhold what the source withholds, and do not repair. R21 V4 verbatim.

W6 — Nothing is added. R22 W6 verbatim, and it remains a gate rather than a preference. An arm that reaches register by supplying an oath, an intensifier, an epithet or a tag the source does not have is not evidence about the source's register at all.

W7 — The propositional content is complete and unaltered. R21 V7 verbatim.

W8 — Stop short of the parodic and the incomprehensible. R21 V8 verbatim. It binds hardest on +S: respelling a narrator is a strong move in English and the cell that permits it is still required to read as a serious rendering.

Procedure — and it is a minimal-pair procedure, which is the point

RS-20260808f's stage B compared arms written in separate independent passes, and four of its eight sites came back as exact ties because two independent passes had chosen the same English. The run could not have reached significance whatever the data did (RS-20260808f §3.1.1). R23 removes that failure mode by borrowing RS-20260805h's method, the cleanest instrument this project has built:

1. ∅ is written first, one pass from the source alone, no published rendering consulted.
2. A, B and AB are minimal revisions of ∅. The translator goes site by site through ∅ and changes only what the newly permitted device licenses. Wording the switch does not reach is left byte-identical.
3. Every touched site is logged with the ∅ reading, the revised reading, and which switch licensed the change. Sites where the permitted device had nothing to offer are logged as unreached — that record is a measurement of the device's reach and is as informative as a change.
4. Coverage codes as R21/R22: P predicted by the rule set · D deviation with reason · INT decided by interpretation · L a loss the rule set caused.

What follows from the procedure, and any design using R23 must state it. The cells are a manipulation, not four independent translations. Byte-level overlap between them is intended and is not contamination; contamination is measured against parties outside the quadruple. A cell may not be read as an independent second opinion about anything.

What R23 predicts about itself, written before translating

If RS-20260808f's translator's log is right — "the live ones are the located ones" — then A reaches nearly the whole of the bundle and B reaches little of it. If RS-20260808f §3.1.2's post-hoc decomposition is right — the two largest movers in the whole run were the two eye-dialect sites — then B reaches most of it and A little. The two readings available in the record point in opposite directions and the rule set is indifferent between them.

A third outcome is live and is the one the framework would least like: that B reaches the low pole and is located anyway, because respelling a narrator's speech puts him somewhere as surely as his vocabulary does. That would mean there is no cheap way down, and it is measurable.

Relation to the other regimes

Roles, human involvement

As R21 and R22. Lead translator, labeled subject, $0 (charter §3, A4). The lead never judges the output. When a design uses R23 cells written by a panel model as its primary arms, the lead's quadruple is a secondary human replication and says so.