Repository path: workshop/experiments/E-20260815e-fluent-carriage-2/design.md · rendered 2026-09-09

E-20260815e — carriage against non-fluency, on a manipulation that survives a parity control

ARM-fluent-carriage step 2. Frozen before any call is dispatched. The three arms and the rebuilt translator's record were frozen at fffce5f7, before this file existed.

This is a rebuild of the manipulation, not a replication of the result (ARM-fluent-carriage §Steps, registered before step 1's numbers existed). Step 1 built the three arms, passed 342 mechanical checks, and then lost its whole carriage half to a content-parity control dispatched last. Note (boa) is the standing consequence and this design is its first application.

1. The question, unchanged

When carriage and non-fluency point at opposite renderings, which one does the judgment follow?

RS-20260814f §7 named the missing arm — fluent and source-ward — and step 1 built it. What step 1 could not do is read the answer, because the texts it contrasted were not propositionally identical.

2. What changed, and why each change is forced

(i) Three of the seven device classes are dropped. RS-20260815b §6(i): an independent auditor, having caught 3 of 3 planted factual errors, held that flattening a suspension invents its completion, that a reduplication asserts more than a single token, and that varying a held keyword changes the degree asserted. A device is form-only if removing it changes nothing the text says, and the test is the auditor's to apply, not the translator's. A1, A3 and A4 are therefore not manipulated; they stay carried in all three arms.

(ii) What is left is a purely syntactic manipulation, and that is a sharpening. The dropped classes are lexical and punctuational; the four that remain are all word order and sentence shape — mimetic manner (6 sites), the piled period (10), the attribution standing alone (12), the withheld subject (5). 33 sites. Whatever this run measures, it measures about syntax, with every word of the vocabulary and every mark of the punctuation held identical.

(iii) A7.1 is excluded. R29-v1 records it FAILED — English cannot leave a finite main clause subjectless — so the carrying arm does not carry it. Step 1 flattened it anyway, which put a difference between the arms that was not a carriage difference. Registered here: a FAILED site is never flattened, asserted by checks.py C-8d.

(iv) The oddity arm retires the two operators that produced step 1's defects. O6 dangling modifier — it reversed agency at S2 and garbled referents at S5, and it does so by construction, because a modifier that dangles necessarily re-attaches to a different subject. O7 hedge insertion — basically, sort of, needless to say are semantically empty and pragmatically not, and the parity auditor was right to flag them (§6(ii)). What remains — nominalisation, stilted collocation, adverbial displacement, heavy circumlocution — adds no proposition and moves no agency. A committed filler blacklist stops the retired words returning through another operator.

(v) The parity control is dispatched FIRST, as a gate, and nothing is judged until it passes.

3. The three arms

FF2 is NOT a zero-carriage arm and the design does not treat it as one. The 29 dash suspensions, the 8 reduplications and the 11 tokens of the held keyword are carried in FF2 exactly as in FS. The FS/FF2 contrast is a syntactic dose contrast, not an ablation — 33 sites against step 1's 84 — and every reading below is stated at that strength. Registered consequence: a null on this manipulation is weaker evidence than a null on step 1's would have been, and is reported as such rather than as an absence of the effect.

Orthogonality is mechanical, not claimed: FS/FF2 differ only at flattened spans, FF2/ODD only at edit sites, the two sets disjoint. 507 checks, 0 failures, 9 of 9 mutation tests caught. Eighteen dropped-class sites sit inside a rewritten span and are not exempted: each is asserted to survive in kind — dash present, doubling verbatim, held-keyword count unchanged over the same stretch of story (C-3b).

Length ratios FF2/FS 1.0269, ODD/FS 1.0532, ODD/FF2 1.0256; declared tolerance ±6% whole-text, ±10% per segment, asserted by checks.py (C-7). Worst segment S3 at 1.090.

4. Segments and seats

Seven segments, the same cuts as step 1 (analysis/segments.py, STARTS unchanged), so the two steps are directly comparable. Alignment is computed by difflib and asserted.

Seats: P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, P3 x-ai/grok-4.5 — the same three as ARM-carriage-direction and as step 1, so the instrument is imported unchanged. P5 excluded on note (bne). P4 moonshotai/kimi-k3 runs the parity control and the pre-run critique and takes no judging cell.

Conditions. 7 × 7 × 3 = 147 cells, one call each.

Prompts are step 1's verbatim, so nothing in the instrument moved between the steps. Presentation order fixed per cell by sha256("E-20260815e|" + cond + "|" + seg + "|" + seat) mod 2. The word Japanese appears in no dispatched prompt in any condition, nor the author's name, the title, any regime id, or any of carriage, flatten, device, fluent, odd — asserted by the runner and by the verifier; a violation voids the run.

LG is not re-run, and the omission is declared here rather than discovered later. Step 1 measured the language leak at 9 of 9 cells — total, and arm-invariant, the stated cue being content that is byte-identical across the arms (RS-20260815b §7). The dropped classes make the arms more alike lexically than step 1's, so the leak can only be more arm-invariant, not less. Nine calls saved; the step 1 figure is imported and cited, not re-measured.

5. The gate that runs first

PAR — content parity, dispatched before any judging call. P4, which takes no judging cell, is given the three arms per segment and asked whether they state the same things. Five planted factual errors — place, referent, negation ×2, quantity — are the positive control, one call each, against the undamaged FS text.

• If P4 does not catch 5 of 5 plants, its parity verdicts carry no weight and the run proceeds with parity unestablished, reported exactly as step 1 reported it.
• If P4 catches the plants and reports a real content difference on a segment, that segment is excluded and every statistic is recomputed on the remainder.
• If it reports a real content difference on 4 or more of 7 segments, the manipulation is not content-neutral and NOTHING IS JUDGED. The arm closes retired and the finding is that this project cannot build a content-neutral carriage manipulation on this material — which RS-20260815b §6 already half-says and which is worth saying plainly for one call's price rather than a run's.

This is the whole procedural point of step 2 and it is registered before the first call.

6. Hypotheses and what each predicts

• H-CARRIAGE — the seats track which way the translation is marked. Predicts FS on AB and FS on AC; makes no prediction on AD, where neither arm's syntax carries.
• H-LITERALITY — the seats choose whichever text reads more like a translation, without reading for direction. Predicts ODD on AB and ODD on AD. On AC it predicts FS, and this is stated plainly because it is the design's honest limit: source-ward syntax and translation-sounding syntax are the same surface. AC therefore does not separate the two hypotheses; AB does, and AB is the primary for that reason.
• H-LENGTH — the seats choose the longer text. Predicts ODD on AB and AD and FF2 on AC. Registered because both manipulations add words.
• H-QUALITY — the seats read "followed the original" as general merit. Predicts FS on AB, and is indifferent on AC iff G2 shows quality parity.

What AC is for, given that it does not separate H-CARRIAGE from H-LITERALITY. Step 1 measured AC at 21 of 21 with all seven classes varying. Here only syntax varies. If AC still fires, the sense responds to sentence shape alone; if it does not, step 1's 21 of 21 was carried by the lexis and punctuation that this version holds fixed — and that is a substantive statement about what the sense reads, in either direction. It is registered as a co-primary on that ground, not on separation.

The diagnostic table, written before the numbers exist:

7. Registered statistics, bars and failure criteria

The unit is the seat, not the cell (E-20260813h's critic, imported): seven cells from one model are not seven draws. Each seat contributes a count out of 7; a primary is stated on all three jointly. Bars are step 1's, unchanged — a bar lowered after seeing a weaker manipulation is bar-shopping, and the two steps must be comparable.

P1 and P1′ cannot both fire. If neither fires the result is indeterminate, reported as indeterminate with the width written down.

What voids the run.

• F0 — PAR fails on 4 or more of 7 segments. Nothing is judged; the arm closes retired. (This is the gate that runs first and it is the reason step 2 exists.)
• F1 — G2 fails. If the seats tell FS from FF2 on quality at better than chance, FS is not the fluent-and-source-ward arm the design needs, P2 is confounded with quality, and P1 is reported but may not be read as separating carriage from non-fluency.
• F2 — G3 fails. If ODD is not read as the worse-written arm, the oddity operator did not operate and P1/P3 are void.
• F3 — a PAR flag on 1–3 segments. Those segments are excluded and every statistic recomputed; both the full and the excluded figures are reported.
• F4 — a banned string reaches a dispatched prompt. Voids the run.
• F5 — position lock. Any seat answering the same slot letter in ≥ 20 of its 21 carriage cells is reported as position-locked and excluded from P1, P2, P3, which are recomputed.
• F6 — checks.py non-zero at run time. Voids the primary. (It exits 0 at freeze.)
• F7 — truncation. Note (bph): finish_reason != "length" is a kept-ness condition in the runner and an assertion in the verifier. A truncated body is re-dispatched, never counted.

Registered in advance about what a firing P1 does NOT license. It does not establish that the FS arm carries anything — the census is a count, not a perception (RS-20260813g §10), and a blind auditor already cut step 1's self-assessment from 0.976 to 0.80. It does not establish quality; Tier D is NOT PASSED and every evaluative word here is internal-judgment-only and provisional. And three models sharing a training distribution are not three readers: note (bpi) measured two of these seats agreeing at a 33-character contiguous run on an unrelated task, so unanimity is what a shared projection looks like and this run cannot tell that apart from perception.

8. Money

Pre-flight, built from the cap each request permits and not from an assumed length (note (abc)).

UTC-day headroom at freeze: $1.528904 (2026-08-15 stands at $3.471096 of $5.00 after six sessions). The ceiling fits inside it with $0.18 to spare; if the gate stage overruns, the AS condition is cut first (21 cells, ~$0.14) and the cut is reported.

Note (bny): every reasoning-capable seat gets an explicit reasoning: {"max_tokens": N} and never an effort pin; the two keys cannot both be sent. Note (bph): the reasoning budget is held at about a fifth of the content cap, never at or above it. Note (bnx): every body is appended to runs/bodies.jsonl as it returns and a re-invocation resumes. Note (bof): no key-usage delta is reported as a cross-check; per-request billed cost is the ledger.

Lead translation is free and is not ledgered (charter §3, A4): the rebuilt arm, the 507-check verifier, the alignment and the analysis cost $0.