Repository path: workshop/regimes/R33-footing-directional.md · rendered 2026-09-09

R33 — footing-directional

v1.0 (frozen 2026-08-15, S192.) R06, R27 and R28 are three points on one line: how much of a source's grammatical footing to carry. R33 is not on that line. It answers a different question — whose footing a paragraph is allowed to mark — and it leaves the quantity to fall out of the answer.

Why it exists

RS-20260814h §6 found that the maximal regime R27 delivers a relative grading whose direction is not the source's: blind seats named the woman as the higher party in the paragraph where her ladyship occurs ten times, and the man in the paragraph the narrator's his lordship was pleased to dominates — the same rendering of the same scene pointing different ways in different paragraphs. framework/v0.2 §10.6 drew the consequence in one sentence: "a rule for how much to carry is the wrong shape of rule; what is wanted is a rule about whose footing a paragraph is allowed to mark."

No such rule has ever been written down here, and this is it. Writing it is the point; whether it can be executed, and whether its execution reaches anybody, is what the run that follows measures.

Procedure

1. The baseline is frozen first, and the skeleton is the baseline's. R33 is written only where an R06 rendering of the identical span by the same hand already exists, frozen and committed. Outside the material it marks, R33's text is identical to that baseline, modulo line wrapping — the constraint R28 §2 imposes and for the same reason: a freely rewritten rendering would confound footing treatment with every other sentence a second look would have changed.
2. Read the frozen site list grouped by source paragraph, not site by site. This is the whole difference from R27 and R28, which decide at a site and never look up. For each paragraph the sites establish a grading: some parties raised, some lowered, some left ungraded.
3. The direction rule. The paragraph's English may raise the parties the paragraph's marks raise and lower the parties they lower. It may not place a mark that raises or lowers a party the paragraph's marks leave ungraded, and it may not mark a relation in the opposite direction from the one the paragraph's marks establish.
4. The dominance rule, and it is the operative one. Score each party in the paragraph: +1 for each token that raises that party, +1 for each token that lowers a different party. The party the paragraph's marks place highest must end with the highest score; it may tie only with a party the same paragraph's marks also place above everyone else. A paragraph whose scores put somebody else on top is a regime breach and must be rewritten or the marks dropped.
5. The assertion constraint, inherited from R28 §4 unchanged. Carry where the English device does not assert something the source does not assert; decline where it does. A deferential auxiliary (deign, be pleased to, vouchsafe, see fit) only where the source's predicate is something its subject did on purpose (framework/v0.2 §10.5.3); no device that requires a clause to be added.
6. The budget, declared as a number before translating: at most +8% on the baseline's word count. R27 ran to +21.1% and RS-20260814h established that 70% of what that cost was the cost of the words as words. A directional rule that pays R27's bulk has bought nothing.
7. Record the decision at every paragraph — the grading read off the sites, the marks placed, the dominance scores, and every mark declined with its reason. A paragraph passed over silently is a regime breach.
8. No apparatus: no footnote, no gloss, no bracket, no translator's note.
9. Freeze. Commit the rendering and its log before any evaluation of it is designed.
10. File under workshop/translations/<work>/R33-v<N>/.

The one thing this regime knows it cannot do, written before it was executed

A source that marks footing by inflection can change direction inside two sentences; English marks by lexis, and lexis persists. Where a paragraph graded one way sits between two paragraphs graded the other, R33 will obey the rule in each paragraph separately and a reader who takes in all three at once may still see only the dominant direction. The rule is per-paragraph because the source's marks are per-clause; it is not claimed to control what a longer stretch reads like, and the design that follows this regime puts exactly that stretch in front of blind judges.

What it does not license

An R33 output is one hypothesis-aware hand and is never evidence about how English works. R27 §"What it does not license" applies unchanged. The paragraph gradings are read off a site list and glosses the same agent wrote (E-20260814b, frozen 2026-08-14, independently audited at 64 of 67 by P3), and the counts are self-report: internal-judgment-only wherever they appear, and note (bio) governs — no primary of any design may be a count over these marks judged by the translator.

The regime names a direction and a length budget, and it names no goodness sense. That is deliberate and is the thing a critic should check hardest: a regime specified as "mark it where the English still reads well" would restate naturalness, and any run comparing it against R06 on naturalness would be measuring its own instruction.

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.

Known limitations (by design)

• Everything R06 lists, everything R27 lists, and everything R28 lists.
• The skeleton constraint cuts both ways, exactly as in R28: it makes the comparison attributable and it means R33 prices insertion, not composition.
• The gradings are the translator's reading of the source's marks. They are derived from frozen glosses by a stated rule, and the rule can be checked; the glosses themselves are one hand's philology, and two of the 67 are flagged in RS-20260814b §6 as genuine cruxes.
• A per-paragraph rule cannot see a scene. See the section above; this is the regime's declared weak point and not a discovered one.

Changelog

• 2026-08-15 v1.0 (frozen, S192) — created for ARM-footing-price step 2, to write the rule framework/v0.2 §10.6 named and could not point at. Frozen on creation, as R27 and R28 were.