Repository path: workshop/experiments/E-20260808d-carriage-decoupled/design.md · rendered 2026-09-09

E-20260808d — does perceived-source-carriage survive when form and markedness come apart?

ARM-sense-overlap step 2. Frozen before dispatch. Nothing below was written after a number existed.

1. The question, and why it is not a writing step

Step 2 was scoped as write the verdict into wiki/goodness-senses.md. The verdict it has to carry is RS-20260807f §3: blind perceived-source-carriage correlates with naturalness at −0.975, and its responses to carried source form (+1.750) and to gratuitous oddity (+1.861) are indistinguishable — i.e. the sense D-20260802-13 created in order to be separate from naturalness is that material's naturalness with the sign flipped.

That page's own limit 2 forbids writing it as a standing relation. Carrying Schwob's litany into English marks the English, so on that passage FORM and markedness are confounded by construction, and the observed identity is exactly what a confound of that shape produces. §6.4 of that page specifies the replication owed: a source whose marked forms can be carried into unmarked English, plus an accuracy gate scored without the other senses in the same call. This is that replication. A standing sentence in the project's controlled typology, cited by every future evaluation, may not rest on one passage whose design defect is written on its own result page.

What this unit teaches about translating literature (subject rule, wiki/tracks.md): it decides whether a reader who reports that a translation carries its source across is reporting anything beyond this English is odd — which is the empirical content of the foreignizing claim, and is a question about translations and their readers, not about this project's instruments.

2. Materials

• Source. Herman Bang, Ved Vejen (1886), ¶829–852 — 503 Danish words in 24 paragraphs, the dawn movement. workshop/translations/ved-vejen/source-unit.txt. The project's first Danish.
• Why this passage decouples. workshop/translations/ved-vejen/device-census.md §3, in full: the project's unmarked / literary-contemporary register anchor A-mchugh-presence is catalogued as invisible free indirect discourse · say-only dialogue tags · texture carried by named particulars · structural one-sentence paragraphs, and Bang 1886 has all four (devices A4, A5, B7, A1/A3). Carrying these devices should therefore cost nothing on naturalness. That is not an assumption of the analysis — it is gate G1, and the experiment is void if it fails.
• Census. Sixteen Class A form-only devices (no proposition, no image) and seven Class B image-bearing features, made from the Danish alone and committed at 789ae0d before any English of this passage existed.
• Segments. Seven, S1–S7, 46–95 Danish words, mapped to source paragraphs in T-ved-vejen-R14-v1 §4.

The four arms

CARRY vs FLAT is the primary contrast, and it is the contrast RS-20260807f could not make: form manipulated, markedness held (that is G1), content held (that is G3), imagery held byte for byte (that is the Class A / Class B split).

The STRANGE edits use no word-order inversion. Danish is a V2 language, so an English fronted-adverbial inversion would be a calque of the source's ordinary syntax — precisely the defect RS-20260807f's pre-run critic caught in six of its eighteen unlicensed edits (its BLOCKING (c)). Every edit carries, in materials/arms.py, a written statement of what the Danish does at that site; all nineteen read PLAIN.

3. Panel and stages

Seats J1 = P1, J2 = P2, J3 = P5 (config/models.md). Critic and the two controls are non-panel. The lead judges nothing (charter §5).

28 items (4 arms × 7 segments) per seat per rating stage; 3 seats; 252 sense-cells plus 15 competence cells. Item ids are opaque hashes, order is seed-fixed and independently shuffled per seat, and arms are interleaved.

4. Gates — all declared here, none movable after a number exists

5. Predictions, registered

The primary is read on seat-pooled segment means, n = 7, with the 21-cell form reported alongside. Exact paired permutation over sign flips, 2⁷ = 128 relabelings, minimum attainable P = 0.0078.

• P1 (superiority). Δperceived-source-carriage(CARRY−FLAT) ≥ 1.00 with exact P ≤ 0.05. Holds ⇒ the sense has a source-specific component that survives decoupling.
• E1 (equivalence, absolute). The 90% permutation interval on that difference lies inside ±0.75. Holds ⇒ the sense is measurably indifferent to carried source form.
• E2 (equivalence, relative). |Δ(CARRY−FLAT)| ≤ 0.33 × Δ(STRANGE−FLAT) on the same sense, same seats, same items. E2 exists because E1 alone has already failed once on seven segments — RS-20260808c returned a point estimate of exactly 0.000 and an interval of [−0.777, +0.666], missing ±0.75 by 0.027, note (bkm). E2 is powered by the run's own manipulation instead of by an outside guess at what margin matters.
• P2. r(blind naturalness, perceived-source-carriage) over the 28 items ≤ −0.80. RS-20260807f got −0.975 on confounded material; this asks whether the identity survives decoupling.
• P3. Δperceived-source-carriage(STRANGE−FLAT) ≥ Δ(CARRY−FLAT) — unlicensed oddity earns at least as much source-carriage credit as really carried form.
• P4. Δaccuracy(CARRY−FLAT), scored alone, is smaller than RS-20260807f's +0.917 measured with four senses in one call. Holds ⇒ that page's G1 failure was a within-call halo. Fails ⇒ form contaminates accuracy as a property of the sense, and every accuracy figure in this repo carries it.

6. Failure criteria — declared, and not weakened after firing

• F1. G1 fails ⇒ P1, E1, E2 WITHHELD. The passage did not decouple and this run is a second measurement of the confound, not a test of the sense.
• F2. G4 fails ⇒ P1, E1, E2 WITHHELD. A sense that does not move on this material cannot be shown not to move on form.
• F3. G3b < 6 of 7 or G3c < 4 of 5 ⇒ the arms are not content-matched; everything is descriptive and P4 is void.
• F4. G6 fails for ≥ 2 seats ⇒ stages 3 and 4 are uninterpretable.
• F5. Any seat returning < 80% of its cells is dropped and the loss is declared in the result; cell completeness is reported whatever it is.
• F6. If E1 and P1 both fail and E2 holds, the result is reported as equivalent by the run's own yardstick and not certified at ±0.75, in those words, and the typology text says so.

7. Known limits, written before the run

1. The lead wrote all four arms and chose the passage, as at RS-20260807f. G3b, G5 and G7 put parity, the error scale and the census outside the lead; nothing puts the 24 flattening edits or the 19 oddity edits outside the lead.
2. The lead knew the hypothesis while writing the operator tables — though not while translating CARRY, whose log was frozen at 5999d6f before this design existed.
3. Contamination is a reachability statement, not a measurement (device-census.md §4): no English «Ved Vejen» is reachable. Both primary arms descend from one rendering, so any recall is shared and cannot produce the contrast.
4. One phrase escapes the manipulation by design: half so of gold and half of roses is a Class A doubling sitting inside Class B image B1, so it is held constant and survives into FLAT. This works against the run's hypothesis, and is declared in T-ved-vejen-R14-v1 §2.
5. F6/F7 add words. Connective explicitation and FID attribution supply subordinators and attributions that the Danish leaves implicit. G3b is the control that adjudicates them, and RS-20260807f §5 limit 4 is why this is stated in advance rather than discovered.
6. One passage, one author, one hand, one pair, three seats. Tier D is NOT PASSED and perceived-source-carriage has never been through Tier D at all. Every sentence of the result is provisional and internal-judgment-only.

8. Pre-flight cost estimate

Built from max_tokens, not from expected output (note (abc)), and priced at the worst plausible provider (config/models.md, S022 caution).

Today's ledger (UTC 2026-08-08) stands at $1.932724933 of $5.00 across S132–S134; $3.067 headroom. $1.00 fits.

9. Amendments adopted from the pre-run critic, 2026-08-08, before any rating was dispatched

nvidia/nemotron-3-ultra-550b-a55b (non-panel), max_tokens 16,000, stop, 11,484 characters, $0.0349038. Verdict NEEDS-REDESIGN, 5 BLOCKING, 8 ADVISORY. Eight findings accepted, four overruled with written reasons. The critic's own summary of the design's largest weakness — "no gate catches the confound in Finding 1" — is what amendments A1 and A2 answer.

Accepted

• A1 (BLOCKING 1, in part). Three flattening sites supplied a causal subordinator the Danish does not state (because ×2, since ×1). All three are replaced by bare and, in T-ved-vejen-R14-v1 §5. Two other explicitations were checked against the Danish and kept because the Danish asserts the relation itself.
• A2 (BLOCKING 1 + BLOCKING 5). The parity control now asks two questions, not one: propositional equivalence (gating, G3b), and separately enrichment — does either passage state outright a relation, cause, attribution or inference the other leaves for the reader to supply? The second is reported, not gating, because a positive answer is expected: the whole manipulation is about what is enacted versus what is stated, and a gate that fires by construction is not a gate. Its rate and direction are what the result must carry.
• A3 (BLOCKING 4). G1 is tightened from ≤ 0.75 to ≤ 0.50, and the design's claim is restated: carrying these devices costs less than half a scale point of naturalness, not "nothing". Justification, which the critic correctly said was missing: 0.50 is under a third of the −1.444 that RS-20260807f's FORM manipulation cost on confounded material, and above the 0.333 that RS-20260805e showed is not distinguishable from unit-level noise on this instrument.
• A4 (BLOCKING 2). P2 is void as registered — 28 items are 7 segments under 4 conditions, so the correlation would report between-arm separation, not a within-segment relation. P2 is replaced by P2′: the correlation between blind naturalness and perceived-source-carriage computed on segment-centred residuals (each item's score minus its segment's mean over the four arms), df = 20. The uncentred figure is reported alongside, labelled as the invalid one.
• A5 (BLOCKING 3). P4's cross-experiment comparison is dropped. The halo question is now tested inside this run, on the critic's own proposed design: stage 5 has the same three seats score all four senses in one call on the same 28 items, and P4′ compares Δaccuracy(CARRY−FLAT) scored alone against the same difference scored alongside three other senses, pairwise, same seats, same items. RS-20260807f's +0.917 becomes a descriptive reference and not a test.
• A6 (ADVISORY 1). G3b is raised from ≥ 6 of 7 to 7 of 7.
• A7 (ADVISORY 3). Power is stated rather than implied: with n = 7, P ≤ 0.05 requires all seven segment differences to share a sign, and E1 has already failed once on seven segments at a point estimate of exactly zero (note (bkm)). E2 exists for that reason.
• A8 (ADVISORY 4). Segments are 46–95 Danish words. Both the unweighted and the word-weighted pooling are reported, and a disagreement between them is a finding, not a choice.

Overruled, with reasons

• BLOCKING 1's main remedy — drop A4, A10, A12, A13, A14, A15, A16 from the manipulated set. Overruled. The critic's ground is that these change narrative mode, discourse structure and perspective. They do; narrative mode is form, and it is exactly the form this experiment exists to manipulate. Adopting the remedy would leave a manipulation of punctuation, italics and dialogue tags, which is not what wiki/goodness-senses.md means by style-correspondence and is not what any translator of Bang decides. The enrichment charge inside the finding is real and is answered by A1 and A2; the reclassification is not.
• BLOCKING 5's remedy — move style-correspondence to a separate or blind call. Overruled on the blind half: this sense is defined as a source–target relation and scoring it without the source is incoherent. Answered better on the substance by amendment A9: a mechanical, judgment-free manipulation check, G2b, computed by analysis/checks.py over the frozen texts before dispatch. It counts paragraph splits, one-sentence paragraphs, paragraph-initial And, bare he said, italic spans, four-dot ellipses and the doubling repetitions. CARRY 45, FLAT 7, STRANGE 7, WRONG 7 — CARRY strictly higher in 7 of 7 segments, and the residue of 7 is the same in all three derived arms, so it cannot differentiate them. This is a description of frozen materials, not a threshold on an outcome, and it is a stronger manipulation check than any rater.
• ADVISORY 2 — make G7 gating at ≥ 0.80. Overruled. A formal-device catalogue is not a closed set; an independent model naming fewer devices is weak evidence about the lead's list rather than strong evidence against it. G7 stays reported and non-gating at 0.60, as at RS-20260807f G5.
• ADVISORY 6 — read a positive G4/P3 as "the sense is oddity-sensitive, not source-sensitive". Not overruled so much as adopted as the result's interpretive frame: that is precisely the conclusion the run is built to license or refuse, and it is written into §5 of the result.

Revised ceiling: $1.40 (stage 5 adds 3 calls at max_tokens 12,000, worst case $0.40). Today's headroom is $3.067.