Repository path: workshop/experiments/E-20260811-floor/design.md · rendered 2026-09-09

E-20260811-floor — how far into a page of English prose does the reader learn which country it comes from?

ARM-idiom-reach step 2, respecified by RS-20260810z §7.3. Frozen before the pre-run critic saw it. The translation limb T-mikan-R06-v1 was frozen at commit 49e154a, before this design existed, and its translator's log lists the choices its translator noticed making before any detector in code.py had been run over it.

1. The question, and why it is not method work

framework/v0.2 §7 has spent five sessions measuring what a located device adds to a placeless baseline. RS-20260810z found that the baseline is not placeless: two judges placed the lead's R22 rendering — a rule whose entire content is use nothing a reader could place — and named apologise among the markers. A spelling. Every difference §7 has reported is a difference from a floor nobody measured.

The question. How much of ordinary, competent English narrative prose — written under no register instruction at all — is nationally located, how early, and can a reader tell?

One sentence on what this teaches about translating literature (continue-prompt.md §4.5): it teaches whether "write English that no reader could place" names an act a translator can actually perform, or an instruction that is void on the first page. The object is the English language as a translator has to write it, not any project statistic. The corpus is twelve published translators and authors whose national variety is documented; the project's own arms are reported beside them and are not one side of the primary comparison (note (bly)).

2. Declared deviation from the arm's step-2 wording, with the reason

The arm page says step 2 is "the floor measurement — what share of ordinary competent English narration two judges will place with no register manipulation at all — plus a loc instrument shown to read standard-spelled located English."

The loc instrument is not used here, and NAT replaces it. loc asks would a reader place this English? — a question with no documented right answer for any item, so the instrument can only ever be checked against another instrument, which is what failed at S155 (G3b, one judge of two). NAT asks was this written by a British or an American hand? — a question with a documented right answer for every gate item, so the judges are validated inside this run against the bibliographic record rather than against each other.

The cost, declared: the floor produced here is not on §7's loc scale and may not be subtracted from any loc figure. What it licenses is a statement about placement by nation, which is the channel RS-20260810z §1 actually caught the baseline leaking through.

3. Materials

3.1 The gate corpus — twelve hands, national variety documented

Ground truth GT is the orthographic tradition the hand wrote in, taken from the bibliographic record (hand's nationality and career, corroborated by place of first publication), fixed before any text was measured. BR includes the Commonwealth tradition.

Seven BR, five US; three of the twelve are originally English rather than translated, so the floor is not an artefact of translation. BR1/US1 are the same Turgenev text in two hands — the matched pair, which makes content unable to explain any difference between them.

US5 is the one case where hand and publisher disagree (American hand, Tokyo publisher). GT is scored on the hand, as declared above, and the disagreement is named here so a reader can discount it; G1 is also reported with US5 dropped.

3.2 Reported beside the gate, with no ground truth and no gate

3.3 Extraction, declared before measuring

Provenance headers, # comment lines and Project Gutenberg boilerplate are stripped by the rules in run.py:body(). The measured span is the first 3,000 words of the body, or the whole body if shorter. Lead artefacts are taken from the plain-text spans files already frozen by earlier sessions (r04-spans.txt, r06-spans.txt) or from between ## The translation and the next ## heading. No text was inspected for spelling before its entry in §3.1 was written.

4. The detector (code.py, frozen with this file)

Two inventories, written from reference knowledge of British/American splits, not from any corpus text.

• ORTH — spelling proper. Both members of a pair are the same word, so a difference between two texts cannot be a difference of subject, period or register. Primary and gated.
• LEX — word choice and four morphological preferences (towards, learnt, fortnight, gotten). Reported, never gated: a nineteenth-century American may write railway without error.

The asymmetry rule. Each entry declares which side is diagnostic. -ise is British-marked and -ize is not American-marked (Oxford spelling); sidewalk is American-marked and pavement is not British-marked. One-sided entries contribute on one side only. A character span is claimed by the first rule that matches it, so no token is counted twice.

Per text: n_marks, br, us, fmp (word index of the first mark), density (marks per 1,000 words), verdict (BR/US/TIE/NONE), purity (= max(br,us)/total).

5. Gates — checked before any prediction is read

6. Predictions, with bars fixed before the run

Both directions are informative and the design is indifferent. If P1 fails, the floor is not immediate and §7's placeless baseline is defensible as written. If NAT1 fails — judges mostly answer CANNOT TELL — then the marks the detector finds are below a reader's threshold, and the framework's baseline survives at the level of reading even though it fails at the level of the page. That would be the more interesting result and it is registered as such.

7. Procedure

1. corpus — build materials/corpus.json from §3 by the declared extraction rule; run the detector; write analysis/mech.json. No API.
2. items — build NAT items (2 per text × 18 texts = 36; the first qualifying window of 3–6 sentences and 60–120 words starting at word 200, and the first starting at word 1,000; for texts under 1,200 words, at word 100 and at the midpoint; for the matched pair BR1/US1, passage 1 is the first qualifying window from word 1 in both, so the two are the same Turgenev) and SWAP items (sentences of 15–45 words containing exactly one invertible ORTH mark and no LEX mark; up to two candidates per gate text, then round-robin across all twelve texts to 16, so every gate text contributes one before any contributes two — without the round robin the first six texts fill the stage and the American-published texts never appear). Fixed shuffle seed 8156. Arm and source labels are never shown to a judge.
3. critic — one independent pre-run critic pass over this file and code.py (openai/gpt-5.6-terra), adjudicated in writing before any judge is dispatched.
4. nat — 3 blocks × 12 items × 3 judges = 9 calls.
5. swap — 2 forms × 3 judges = 6 calls, 16 items each. Form V1 shows pairs 1–8 as published and 9–16 swapped; V2 the complement. No judge ever sees both members of a pair in the same call, and the two calls are independent dispatches.
6. verify — recompute every reported number from the raw bodies; mutation tests on the detector; G3 checked character by character.

Judges. L1 = nvidia/nemotron-3-ultra-550b-a55b and L2 = z-ai/glm-5.2 — the two seats RS-20260810z used, so the floor is read by the hands that read S155's arms; L3 = google/gemini-3.6-flash, a third seat added because S155 showed two seats can disagree about the object. The critic model is not a judge.

Judgement is not parallelised; every dispatch is sequential and every raw body is written to runs/ before anything is computed from it. A dead body is rotated into runs/discarded/ and never overwritten (note (bhd)).

8. Pre-flight cost estimate, built from max_tokens (note (abc))

Declared ceiling: $0.75. Today's UTC ledger (2026-08-11) is empty; the cap is $5.00.

9. What this design does not do

1. No loc figure and no comparison with RS-20260810z's numbers. §2.
2. No claim about register. Nation-marking and register-marking are different channels; this run measures the first and says nothing about the second.
3. The publisher, not the translator, often fixes the spelling. House style is real and is not controlled for. It does not weaken P1 — a text located by its publisher is still located, and a translator who wants placeless English cannot get it from a publisher either — but it does mean G1's success is evidence about texts, not about translators' habits.
4. Twelve texts is a small corpus and it is English-only on the target side. fmp medians are reported with the full ordered list, not as point estimates.
5. X6's P4 count depends on the log being honest about what its translator noticed. The log was frozen before code.py was written, which is the only guarantee available; the lead wrote both, which is stated rather than defended.

10. What was known before the freeze — declared, not tidied away

This design was written in full, including every bar in §6, before run.py existed and before any number had been computed. The mechanical stage was then run as a dry run before this file was committed, and three things happened that a reader is entitled to know about.

1. Two rules in code.py were false positives on their face and were deleted. The -our/-or stem list contained terr, which fired on terror — a word with no British variant — and the -ise stem list contained surpr and advert, which fired on surprise and advertise, neither of which has an -ize form in any variety. Both deletions reduce the mark count and push fmp later, which makes P1 and P3 harder, so the correction runs against this session's expectation and not with it. A negative-control list of 45 non-split words (error, horror, mirror, doctor, author, actor, sailor, tailor, emperor, exercise, enterprise, compromise, promise, precise, otherwise, story, glory, memory, factory, history…) now returns zero marks and is re-checked in verify.py.
2. Missing splits were deliberately NOT added. The same negative-control pass showed the inventory has no entry for parlour/parlor, and it plainly lacks others (judgement, skilful, instalment, fulfil). Adding them would have raised sensitivity in the direction that helps P1, after the fact, so none was added. The consequence, stated as a bias direction: n_marks and density are lower bounds and fmp is an upper bound. Every figure this run reports is conservative with respect to P1.
3. G1 was restructured, and this is the substantive change. As written, G1 required the detector to recover each hand's documented nationality. The dry run showed that Isabel Hapgood — an American, published by Scribner in New York — wrote neighbourhood, grey, colour, favourite, woollen, rumours, theatre and honourable, and the same for her 1886 Crowell Cloak. Either the American hand wrote in the British tradition or the scans are of a British-market printing; the design cannot tell which, and neither reading is a detector failure. The original G1 was therefore not a test of the detector at all but a test of an assumption about hands that this session had no right to make. It is demoted to M4, reported with every mismatch named, and replaced by two gates that do not depend on the bibliographic record: internal split-half stability, and concordance with the words the judges themselves quote. No bar in §6 was changed, no text was added to or removed from §3, and the jury limb had not been dispatched when this was written.

What this costs. The mechanical limb of this run is not blind — its numbers were seen before this file was committed, and it should be read as a measurement with a pre-written analysis plan rather than as a pre-registered test. The jury limb (NAT, SWAP, G1b, M5) is pre-registered: not one of its calls had been made when the freeze commit was taken.