Repository path: workshop/experiments/E-20260807f-carriage-or-strangeness/design.md · rendered 2026-09-09

E-20260807f — is perceived-source-carriage about the source, or about the strangeness?

Frozen before dispatch. No API call had been made when this page was first committed. Amendments made after the critic pass are numbered A<n> at the foot and dated; nothing above them is edited in place.

1. Question

wiki/goodness-senses.md carries two senses that both go up when a translation is pulled toward its source, and it has never asked whether they are two things:

• style-correspondence — marked formal features of the source receive functional equivalents rather than silent flattening. A verified source–target relation.
• perceived-source-carriage — a marked departure from target-language norm reads as a deliberate attempt to carry over a feature or effect of the source. A reader's attribution.

RS-20260807e (S129) is why the question is live now: the foreignizing arm came out highest of seven arms on both, 6.000 and 6.722, while accuracy did not move at all. The two senses moved together and nothing in that design could separate them.

The question, in one sentence: does a reader's sense that a translation is carrying its source over track the source's actual formal features, or does it track markedness in the English?

Subject-rule sentence (wiki/tracks.md, continue-prompt.md §4.5): this unit teaches whether the impression that a translation "brings the reader to the author" is bought by carrying the author's forms or bought by being odd — which is the empirical content of the foreignizing claim, and which decides whether two of the eight senses of good are one sense under two names. That is about evaluating translations, not about this project's apparatus.

What is new against E-20260802f (S092), which is the predecessor. That run manipulated licensed against unlicensed strangeness at site level, on Homeric verse, and measured attribution of individual departures: real licences attributed at 15 of 15, manufactured oddity at 0.60 without the source and 0.40 with it. perceived-source-carriage did not exist as a sense then — it was created out of that run by D-20260802-13. This run asks the question in the sense's own scoring form (a 1–7 whole-rendering score), crossed with a second factor the predecessor never had (whether the source's forms are carried at all), on prose, in a new language pair, alongside style-correspondence on the same items.

2. Materials

Source. Marcel Schwob, «Paroles de Monelle», from Le Livre de Monelle (Paris, 1894) — the four consecutive litanies de la destruction / de la formation / des dieux / des moments. Public domain, Project Gutenberg #53374, read in French. 500 words, segmented into six segments (../../translations/monelle-paroles/source-fr-segments.txt). The project's first work in French symbolist prose and its first Schwob.

Why this passage. It is a text whose difficulty is almost entirely formal. Its eighteen catalogued devices (../../translations/monelle-paroles/device-census.md) — anaphora, jussive chains, a six-fold epigram template each member of which is a figura etymologica, verset paragraphing, unreduced keyword chains — can each be removed without changing what a single sentence asserts. A passage where form and content can be varied independently is what a 2 × 2 on this question needs, and this project has not had one.

Base rendering. T-monelle-paroles-R04-v1, the lead's close translation, with R06-v1 frozen first per R04 §2a. Both committed at fcdd9c2 before any comparator was read and before this design existed.

Contamination: suspected, measured, with a declared priming event — ../../translations/monelle-paroles/contamination.md. Nine fragmentary OCR lines of the Meloney 1929 English were seen before translating; measured overlap 27 shared 7-grams, 5 twelve-grams, longest run 16 tokens, against two independent published pairs at 6 / 0 / 10 and 23 / 6 / 17. The design's validity does not turn on independence from Meloney — every contrast is within-lead and matched — and §7 carries the one place it does bite.

The four arms (materials/arms.md, frozen by commit before dispatch), a 2 × 2:

CC is the R04 translation verbatim. FC applies eighteen flattening operators, one per catalogued device, in the spirit of R14. CS applies eighteen unlicensed markedness edits — six odd prepositional governments, six coined hyphenated compounds, six inversions — placed where the source is formally plain and of kinds that answer to nothing in the French. A calque would be source-licensed and is therefore excluded by construction. FS applies the same eighteen edits at the same propositional sites in the flattened wording.

Nothing in the 36 edits changes what the text says. That is the binding rule and gate G1 checks it.

3. Seats and blinding

Three seats, non-Anthropic, the Tier D jurors of S020/S034/S086 so the instrument is extended rather than replaced (config/models.md panel v1): J1 = P1, J2 = P2, J3 = P5. Slugs logged in runs/ as provenance.

• 24 items, ids I01–I24, assigned by sha256 of the (segment, arm) pair so the id order carries no information about either factor. materials/labels.json is never sent anywhere.
• Order shuffled per seat under a fixed seed, with no two items of the same segment adjacent.
• No arm labels, no regime names, no mention that arms, factors or an experiment exist. Seats are told the set contains renderings from different sources, in no order, and that the same French may appear more than once.

Two stages, because the two senses are asked under different information by their own definitions.

• Stage A — source present. Each seat sees the French segment and one English rendering, and scores style-correspondence and accuracy, integers 1–7.
• Stage B — source absent. Each seat sees the English alone and scores perceived-source-carriage and naturalness, integers 1–7.

Stage B's source-absent condition is required by perceived-source-carriage's own entry (the score must state whether jurors had the source text) and is the condition in which the sense is most at risk of being pure markedness-attribution — which is what this run is for. Stage A is dispatched first; Stage B's prompt never mentions that a source exists.

4. The senses, quoted verbatim to the seats

The strings sent are the first paragraph of each entry in wiki/goodness-senses.md, with the naturalness string in its revised, target-text-only form (D-20260802-13; RS-20260805e requires that a design say which string it used) and its register anchor named as unmarked / literary-contemporary.

5. Gates, checked in this order, all registered before dispatch

• G1 — accuracy parity. max − min of the four arm means on accuracy (Stage A) must be ≤ 1.00. Above that, the arms are not propositionally matched, every downstream contrast is confounded with content, and the primary is withheld.
• G2 — the STRANGE manipulation took. Blind naturalness (Stage B), pooled mean(CC, FC) − mean(CS, FS) must be ≥ 1.00. Below that, nothing about strangeness is licensed by this run.
• G3 — the FORM manipulation took. style-correspondence (Stage A), pooled mean(CC, CS) − mean(FC, FS) must be ≥ 1.00. Below that, either the flattening operator did not remove the devices or the sense cannot see them, and both readings are reported; the primary is withheld.
• G4 — leak screen. No item text may contain an arm name, a device name, a translator's note, or the words arm, rendering set, experiment. Mechanical, in verify.py.
• G5 — the device census is not the lead's word alone. One non-panel seat sees the six French segments and no English at all and is asked to name each segment's marked formal features. The independent list must cover ≥ 60% of the eighteen device families, scored per segment and pooled, by a coding rule fixed here: a device counts as covered if the independent answer names the same formal property of the same segment, whatever vocabulary it uses. Below 60%, the FORM factor is the lead's stipulation and the result says so.

6. Registered predictions

Every one of these is registered in the direction that would flatter the two senses. The lead's expectation is that all four hold.

• P1 — style-correspondence is discriminant. Its FORM effect mean(CC, CS) − mean(FC, FS) is ≥ 1.00, and its STRANGE effect |mean(CS, FS) − mean(CC, FC)| is < 0.75. Fails if unlicensed strangeness moves the sense that is defined on source form by 0.75 or more.
• P2 — perceived-source-carriage responds to the source. Its FORM effect is ≥ 1.00. Fails if carrying eighteen of the source's own formal devices does not raise the sense that is named after carrying the source over.
• P3 — perceived-source-carriage responds to strangeness. Its STRANGE effect is ≥ 1.00. Expected from RS-20260802f's 0.40–0.60 attribution of manufactured oddity.
• P4 — the decisive cell comparison. mean(CC) − mean(FS) on perceived-source-carriage is ≥ 1.00: a rendering that carries the source's forms and adds no gratuitous oddity reads as more source-carrying than one that carries none of them and is merely odd. Fails, and the failure is the finding, if FS ≥ CC — the sense would then be bought by strangeness alone.

Descriptive, not a prediction: the item-level Pearson correlation between the two senses across all 24 items, reported with its factor decomposition. A value ≥ 0.70 with both factors varying is reported as the two senses are behaving as one on this material, and is not a threshold anything turns on.

7. Known confounds and limits, declared before the run

1. Length is confounded with FORM. Flattening supplies the connectives English wants and the flattened arms are 12–14% longer (591 and 581 against 519 and 515). Every effect attributed to FORM is also an effect of 70 words. Reported, not repaired; the analysis reports the four cell means so a length reading stays available.
2. Explicitation. The flattening operator supplies causal connectives (since, because) where the French is asyndetic. The relations are implied by the source, but making them explicit is the nearest the operator comes to touching content, and G1 is the check on it.
3. The lead wrote all four arms, chose the passage, catalogued the devices and registered the predictions. G5 puts the device catalogue outside the lead; nothing puts the edits outside it.
4. One passage, one author, one language pair, one hand. Nothing here generalises past French symbolist prose without replication.
5. The contamination bite, from contamination.md §4: to the extent CC is remembered Meloney, the CARRIED pole is less purely the lead's craft than the design's language implies. Bounded at five twelve-grams in 519 words.
6. Tier D is NOT PASSED and perceived-source-carriage has never been through Tier D at all (config/models.md). Every sentence of the result is provisional.

8. Failure criteria

• F1 — any gate G1–G3 fails → the primary (P2, P4) is withheld and the run reports what its gates measured.
• F2 — any seat returns fewer than 24 rows in a stage → that block is re-dispatched once; if it fails again the seat is reported as incomplete and every figure is recomputed without it.
• F3 — G5 below 60% → the FORM factor is reported as lead-defined and P1's discriminant claim is downgraded to a statement about the lead's own catalogue.
• F4 — if accuracy and style-correspondence correlate above 0.90 across items, Stage A is reported as a single judgment wearing two names and P1 is withheld.

9. Analysis

analyse.py computes every reported number from runs/*.json and nothing is transcribed by hand. verify.py imports nothing from analyse.py, recomputes every figure independently from the raw bodies, re-derives the item→arm mapping from arms.md rather than from labels.json, runs the G4 leak screen and the G1 parity arithmetic, and includes mutation tests that corrupt a stored score and assert the checks fire.

Exact permutation tests where the cell counts permit: the FORM and STRANGE main effects are tested by exhaustively relabelling the 24 items' factor assignment within seat, reporting the exact two-sided P.

10. Pre-flight budget

UTC day 2026-08-07; config/budget.md shows $2.783760945 of $5.00 spent across S125–S129, so headroom is $2.216239055. Declared ceiling for this run: $0.90.

Worst cases are built from max_tokens at the most expensive plausible provider, per note (abc). Snapshots either side; per-response usage.cost is primary and the key delta is the cross-check.

Amendments

Pre-run critic: nvidia/nemotron-3-ultra-550b-a55b (non-panel), stop, 22,122 characters, $0.0297342, provider Together. Verdict NEEDS-REDESIGN, 8 BLOCKING. Six accepted in whole or in part, two remedies overruled with reasons. No scoring call had been dispatched.

• A1 (from BLOCKING (d) — the best finding in the pass, accepted whole). The design measured perceived-source-carriage only source-absent, so its FORM effect could not be told apart from seeing the source suppresses markedness attribution. A third stage is added: Stage C, same 30 items, source present, perceived-source-carriage alone, dispatched last. The sense is now measured under both information conditions on the same items with the same seats, which is what the question needs and what RS-20260802f did at site level.
• A2 (from A1). Two predictions are added, P5 and P6 below.
• A3 (from BLOCKING (c), accepted in part). The critic went through the eighteen strangeness edits and found all six odd prepositional government edits traceable to the French preposition at that site — saturated of ← saturé de, upon them ← sur eux, unto ← à twice. That is correct and it made a third of the STRANGE factor source-licensed, which is the one thing the factor may not be. All six are replaced by broken / unresolved constructions — a kind the passage contains nowhere, since its syntax completes every construction it starts, and the kind RS-20260802f measured as the worst case for false attribution (5 of 6 codings read a manufactured anacoluthon as faithful). The critic's verdicts on the compounds and the inversions are overruled: French has neither Germanic noun-compounding nor English predicate fronting, so skin-gatherers, up-buildings, between-thing ties, tomb-wailer, over-/under-goodness and all six inversions are markedness English supplies out of its own resources. Its arguments there invent French forms not in the text (a cleft at S3 that Schwob does not write, a sur-/sous- prefixation the passage does not use).
• A4 (from the same finding, accepted). Edit 2 moves off distortion-mirrors, which shadowed the ordinary French miroirs déformants, to shape-warping mirrors.
• A5 (from BLOCKING (e), accepted). An accuracy-parity gate over arms built to be propositionally equivalent cannot fail, and a gate that cannot fail is not a gate. A fifth arm WRONG is added — CC with one content error per segment and no formal change — and G1 is re-registered with a power condition. WRONG is not a cell of the 2 × 2 and enters no FORM or STRANGE figure. Items go from 24 to 30.
• A6 (from BLOCKING (e) on G3's circularity, accepted). G3 used style-correspondence — one of the two target senses — as the sole manipulation check for FORM. G3b is added: a mechanical device count in verify.py, which counts anaphora openings, repeated keyword chains, paragraph-initial connectives, jussive Let openings, template members and figurae etymologicae in each arm and asserts the flattened arms carry strictly fewer. It uses no seat and no judgment.
• A7 (from BLOCKING (b), accepted as reporting). Every effect is computed within segment (each segment supplies all five arms, so the design is paired) and the analysis reports the item-level correlation of word count with each sense, so a length reading of any FORM effect stays visible. The confound is not repaired and §7.1 stands.
• A8 (from BLOCKING (g), accepted in part). Order is re-randomised per stage as well as per seat, so nothing correlates across stages. The remedy of adding 24 filler passages is overruled: it doubles the spend against a leak that lets a seat see there are variants but not which factor is which, and the factors are not nameable from the items alone. Declared as a limit, as RS-20260807e declared the same one.
• A9 (from BLOCKING (f), limit accepted, remedy overruled). An independent translator must produce at least one arm per cell would destroy the design: the arms differ only by a matched operator, and an independent hand would reintroduce every difference the matching removes. G5 puts the device catalogue outside the lead, G3b puts the manipulation check outside judgment, and the parity call puts propositional equivalence outside the lead. Nothing puts the edits outside the lead and §7.3 says so.
• A10 (from BLOCKING (h), accepted). Binding on the result page: its headline may say what this manipulation did on this material, and may not say that two senses are or are not one thing. One passage, one author, one hand, one pair.
• A11 (budget). Stage C and the sixth item per stage raise the declared ceiling from $0.90 to $1.20. Headroom on the UTC day at the time of the amendment is $2.216239055; the revised worst case is $1.17.

Re-registered gates and predictions (superseding §5 and §6 where they conflict)

• G1 — accuracy (Stage A): max − min over the four 2 × 2 arms ≤ 1.00, and mean(four arms) − mean(WRONG) ≥ 1.00. The second clause is the power condition: if the gate cannot separate six planted content errors from none, its first clause is uninformative and the primary is withheld.
• G3b — mechanical device count, verify.py: for each of the six segments, the flattened arms must carry strictly fewer counted device tokens than the carried arms. A segment that fails is named and excluded from every FORM figure.
• P5 — perceived-source-carriage's FORM effect is larger source-present (Stage C) than source-absent (Stage B), by ≥ 0.50. The seat that can check the French should be able to tell carried form from invented oddity; the seat that cannot, should not.
• P6 — perceived-source-carriage's STRANGE effect is smaller source-present than source-absent, by ≥ 0.50. This is RS-20260802f's 0.60-without / 0.40-with, in the sense's own scoring form. Fails if seeing the French does not reduce the credit that unlicensed oddity earns.