Repository path: workshop/experiments/E-20260806c-source-grammar/design.md · rendered 2026-09-09

E-20260806c — is the displacement a property of translating, or of the source's grammar?

ARM-translated-register step 2, and the arm's closing step. Frozen 2026-08-06 before any figure in §5 was computed. Nothing below was written after seeing a corpus-2 number; §2.5 declares in full what was already known when it was written.

internal-judgment-only and provisional: Tier D is NOT PASSED, config/models.md reads NOT CALIBRATED. No jury is convened here and no translation is judged. The only API call in this design is the pre-run adversarial critic.

1. The question the arm hands over

RS-20260805f asked whether English literary translation sits at a different distance from its own moment's English than the same translator's original English does. The profile-level answer was a clean null — S = −0.0021, P = 0.497, with a positive control at 0.739 — so there is no global translated-English register for a fourth naturalness anchor to point at.

What survived was feature-level: your +0.676 and you +0.369, in all four hands, at P = 0.0001, and {you, your} was registered by name from the translator's log before the corpus was touched. ARM-translated-register step 2 inherits the question whether an anchor can hold a feature-level displacement with no profile-level register behind it.

That question cannot be answered until two alternatives to "a property of translating" are ruled in or out, and RS-20260805f §7 names both as unaddressed:

• The source-family alternative. All four cells translate from Spanish or French. "A 'translated English' reading and a 'Romance-into-English' reading make identical predictions on this corpus" (§7, from the critic's F5c).
• The dialogue-share alternative. you and your are overwhelmingly words of direct address, and the {you, your} figures come from the full-text analysis, where the translated and original arms differ in how much of the text is quoted speech.

This design tests both, on a new corpus, and closes the arm either way.

2. Materials

2.1 Corpus 1 — re-declared, unchanged

E-20260805f's four cells (SMOL, MACH, HEARN, HAPG), asserted by SHA-256 against that experiment's frozen manifest at load time (analysis/corpus.py::assert_corpus1_integrity). All four sources are Romance (ES ×1, FR ×3).

2.2 Corpus 2 — five new within-translator cells

Every translator attribution was read off the Project Gutenberg file's own header, not inferred from a catalogue search. Every text is public domain and freely reachable (charter §7, A8).

Two cells hold the hand and the baseline fixed and vary only the source family — DOLE-RU against DOLE-ROM (identical original arm), and HAPG-RU against HAPG (identical original arm). That is the tightest contrast this question admits without commissioning translations.

The four cells that are both non-Romance and independent of each other are CARL, DUFF, HAPG-RU, DOLE-RU. P1 is stated over exactly those four.

2.3 Block counts, computed before the freeze

Every arm yields at least 13 blocks of 2,000 tokens in every mode except speech, where DUFF's original arm (Letters from Egypt) yields 0 — a book of letters with almost no quoted dialogue (speech share 0.008) — and HEARN's original arm yields 0 (share 0.059). FC2 handles this.

2.4 Dialogue share, computed before the freeze, and one divergence from RS-20260805f

Speech share (fraction of tokens inside quotation marks), computed per work, which is how the analysis blocks each text:

E-20260805f design §2.4 reports MACH as 29.0 vs 13.7 and states that dialogue share is higher in the translation in all four cells. Recomputed per work with the corrected quote_convention, Machen's original arm is 37.5 (Pan 43.0, Impostors 35.5) and the asymmetry in that cell runs the other way. Concatenating the two original works and choosing one convention for the pair gives 25.9, under which the published direction holds. The divergence is therefore between two ways of computing the same diagnostic, and the analysis-consistent one is per work. This changes no statistic reported by RS-20260805f — the diagnostic was used to choose narration-only as the primary analysis, and a one-directional confound is not required for that choice to be right. It is recorded here, will be recorded in this run's limits, and is not treated as a licence to reopen RS-20260805f: the subject rule (wiki/tracks.md) makes an apparatus anomaly a limits entry unless a published figure is false, and what is false here is a sentence about a diagnostic, not a result.

2.5 What was already known when this design was written — declared in full

Pre-registration is worth nothing if the registrant has already looked. What had been looked at:

• RS-20260805f §5 in full, as published: your +0.676, you +0.369, that +0.576, some −0.678, all full-text, all P = 0.0001.
• E-20260805f's stored posthoc.json, from which one figure not in the published result was read: on the narration-only analysis, you has mean d = −0.167, P = 0.195, and splits 2–2 across the four hands; your does not appear in the narration-only 100-MFW inventory at all. This is a known quantity in this design and is never reported as a confirmed prediction (FC6).
• §2.3 and §2.4 above — block counts, dialogue shares, quote conventions.

Nothing about corpus 2's feature behaviour has been computed. No you, your or possessive figure exists for CARL, DUFF, HAPG-RU, DOLE-RU or DOLE-ROM in any mode.

2.6 The translation limb and where the predictions come from

The lead translated Storm's Immensee, chapter «Elisabeth» (1,260 German words, 33 paragraphs), R06 draft frozen then R04 revision: T-immensee-elisabeth-R06-v1, T-immensee-elisabeth-R04-v1.

The R06 log carries a census: every token of the class {you, your, my, his, her, its, our, their} in the draft, coded O (an overt German form licenses it) or S (the German has none and the translator supplied it). materials/census-quijote-es.md applies the same census to the lead's frozen Spanish translation from S114. Both are parsed by analysis/census.py:

The site of obligatory supply is opposite in the two languages. Spanish drops the subject pronoun and the translator must insert you; Spanish carries its possessives overtly and the translator copies them. German never drops the subject pronoun; German takes the bare definite article on body parts and belongings, and the translator supplies the possessive at half the sites.

This reverses a prediction the R06 log had already frozen (that the German cells would sit below the Romance cells on the possessive class). The frozen form is kept on record as P2b⁻ and the reversed form is registered as P2b; they cannot both hold, and one of them will be wrong. The limits of both censuses — one locus each, unmatched in narration/dialogue mix and in content, coded by the translator — are stated on materials/census-quijote-es.md and inherited here.

3. Predictions, registered

Displacement Δ for a feature f in a cell is the difference in rate per 1,000 tokens between the translated arm's blocks and the original arm's blocks, so that corpus 1 and corpus 2 are on one scale. (RS-20260805f's figures are in Delta-z units and are not comparable in magnitude; a z-unit replication within corpus 2 is computed alongside as a sign check.)

• P1 — replication. {you, your} displacement is positive with permutation P < 0.05 in the pooled statistic over the four independent non-Romance cells (CARL, DUFF, HAPG-RU, DOLE-RU), and positive in at least 3 of those 4.
• P2a — pro-drop forces the second person. {you, your} displacement is larger in the pro-drop-source cells (SMOL: ES; DOLE-ROM: IT+ES) than in the two German cells (CARL, DUFF).
• P2b — the article construction forces the possessive. Possessive-class {my his her its our their your} displacement is larger in the two German cells than in the Romance cells.
• P2b⁻ — the superseded frozen form, evaluated and reported beside P2b: German below Romance on the possessive class.
• P3 — the within-hand contrast. With the hand and the original arm identical, DOLE-ROM shows a larger {you, your} displacement than DOLE-RU; and HAPG (FR) a larger one than HAPG-RU (RU).
• P4 — profile-level replication. On the four independent non-Romance cells, narration-only, the S statistic of E-20260805f is again indistinguishable from its permutation null (P > 0.05).

Decomposition D, registered as a reported quantity rather than a prediction. For every cell, Δr_full is split exactly into a share component (q_T − q_O)·(s̄ − n̄) and a rate component q̄·(s_T − s_O) + (1 − q̄)·(n_T − n_O), where q is speech share, s and n are the feature's rates inside speech and inside narration, and barred quantities are two-arm means. The identity is exact and is asserted to 1e-9 by the verifier.

4. Failure criteria, registered

• FC1 — the positive control. Speech against narration, inside corpus 2's translated arms, on the same feature space and null, must return S ≥ 0.5 at P < 0.01. If it does not, the machinery is not shown to see a register difference on this corpus and every null in this run is withheld.
• FC2 — thin arms. Any arm with fewer than 6 blocks in the mode being analysed is dropped for that analysis and the affected statement recomputed on the remainder, declared. If fewer than 3 independent non-Romance cells survive, P1 is withheld. DUFF-original and HEARN-original are already known to have 0 speech blocks, so the speech-only decomposition is computed on rates over whole arms rather than blocks wherever an arm is short.
• FC3 — no false independence. DOLE-ROM (shares its original arm with DOLE-RU) and HAPG-RU (shares its original arm with HAPG) are never counted as independent cells in any n of N statement. Any sentence that does so is void.
• FC4 — the dialogue-share gate. If the share component of D accounts for ≥ 80% of Δr_full for {you, your} in ≥ 3 of the 4 independent non-Romance cells, the {you, your} result is reported as a dialogue-share artifact and P1 is withheld regardless of sign or P.
• FC5 — one-cell orderings. If any ordering claimed by P2a, P2b or P2b⁻ is destroyed by leaving out a single cell, that prediction is withheld.
• FC6 — the known quantity. The corpus-1 narration-only you figure was inspected before this design was written (§2.5) and is reported as known, never as a confirmed prediction.
• FC7 — floor on counts. A cell arm with fewer than 200 tokens of a feature class is not reportable for that class.

5. Procedure

1. Assert corpus-1 SHA-256 integrity against E-20260805f's manifest; abort on any mismatch.
2. Strip Gutenberg header/footer; split quoted speech with the 4,000-character guard; tokenise to lowercase word forms; cut non-overlapping 2,000-token blocks, never straddling two works.
3. Thin every arm evenly to at most 40 blocks, as E-20260805f did.
4. Rate statistics — per block, feature rate per 1,000 tokens; Δ = mean(translated) − mean(original); pooled over cells by unweighted mean. Null: relabel blocks within each cell, 10,000 times, one shared permutation draw across features so signs stay comparable.
5. Profile statistic (P4, FC1) — 100 most frequent types across the analysed blocks, Delta z-standardisation computed once, outside the permutation loop, difference vector per cell, S = mean cosine over cell pairs, same 10,000-draw null.
6. Decomposition D — whole-arm rates and speech shares, exact identity as in §3.
7. Verification by analysis/verify.py, which reimplements every reported number and imports nothing from measure.py, plus at least three mutation tests that must be caught.

6. What the run may conclude — fixed before it is run

If P1 holds, FC4 does not fire, and P2a/P2b fail: there is a feature-level displacement that crosses source families and is not grammar-graded. The result states what an anchor could hold (a feature-rate expectation, not a register point) and refers the shelf decision, naming the evidence it would need.

If P1 fails, or FC4 fires, or P2a/P2b hold: the only feature-level survivor of RS-20260805f is either an artifact of how much of a book is dialogue or a shadow of the source language's grammar. In that case ARM-translated-register closes resolved on the null: no fourth register point is built, naturalness's three original-writing anchors stand, and the outcome is written into wiki/goodness-senses.md §naturalness and wiki/base/anchors/README.md.

In no case may it be said that translated English is a register, that the three shelf anchors are the wrong yardstick, or that any translation was judged. No sense is scored. Tier D remains NOT PASSED.

7. Budget

One API call: the pre-run adversarial critic. Pre-flight worst case declared in config/budget.md before dispatch. The measurement is arithmetic over public-domain text and costs nothing; the translation is the lead's own and is never ledgered (charter §3, A4).

8. Amendments, from the pre-run critic — applied before measure.py was written

critic.md records the pass: x-ai/grok-4.5, NEEDS-AMENDMENT, ten findings, six BLOCKING, all ten accepted, $0.0418704. No statistic in this run existed when these were written. Where an amendment contradicts §§3–6 above, the amendment governs; the superseded text is left standing so that what was changed is visible.

A1 (F1, BLOCKING) — the censuses lose their arm-closing power

P2a, P2b and P2b⁻ are demoted to exploratory mechanism hypotheses. They are reported with their directions and their leave-one-out behaviour, and they may not close the arm, withhold anything, or license any sentence about German versus Romance grammar at corpus scale. The critic named seven ways the 2×2 of §2.6 could come out as it did with no contribution from grammar at all — narration-heavy body-part text against dialogue-heavy abstract-possessive text; content rather than language; dialogue share alone; binomial noise at n = 5 and n = 9; Storm against Cervantes rather than German against Spanish; coder-boundary choices made by the interested agent; and one census written live against one written retrospectively. All seven stand.

What keeps its force is P3, which varies the source family with the hand and the original-arm baseline held identical. That is the only source-family contrast in this design that is not a comparison between two different books by two different people.

A2 (F2, BLOCKING) — the closure rule, split in two

§6's disjunction is withdrawn. In its place:

The scientific claim. Close on the null only if FC1 passes and P1 fails its sign pattern and (FC4 fires or P3 shows source-family grading at fixed hand and baseline). If P1 fails, FC4 does not fire and P3 is null, the run reports unresolved — replication failed, mechanism unadjudicated and claims nothing further.

The anchor decision, which is what ARM-translated-register Done when actually asks. A displacement that does not replicate, or whose mechanism is unadjudicated, is not something a naturalness register anchor can be built on either. The shelf question is therefore answerable in the unresolved branch and the mechanism question is not. The arm may close on the anchor question while this run leaves the mechanism question open, and it must say so in exactly those words. This separation is the design's, not the critic's, and it is recorded on the arm page too so a later session can overturn it in one place.

A3 (F3, BLOCKING) — three changes to the dialogue gate

• FC4 threshold lowered from 80% to 50%: if the share component accounts for ≥ 50% of Δr_full for {you, your} in ≥ 3 of the 4 independent non-Romance cells, P1 is withheld as a dialogue-share artifact.
• FC4b, new and parallel — the gate the old one could not be. A small share component is compatible with the whole effect living inside quoted speech, which is still a fact about dialogue and not about translated prose. So: if the narration-only {you, your} displacement fails to reproduce the full-text sign pattern in ≥ 3 of the 4 independent non-Romance cells, P1 is withheld as dialogue-linked, whatever the decomposition says. (The predecessor's stored narration-only figure for you — declared in §2.5 — is −0.167, so this gate is not decorative.)
• FC4c — speech-side rates are not reportable for a cell whose arm carries fewer than 200 speech tokens, and D is recomputed on the remaining cells. DUFF-original (1,104 speech tokens) clears this; HEARN-original (1,562) clears it; the gate is kept because block-level speech statistics for those arms do not exist at all.

A4 (F4, BLOCKING) — P4 gets a declared floor and loses its closing power

The minimum detectable effect is declared as the 95th percentile of S's permutation null. If that exceeds 0.15 — roughly a fifth of the predecessor's dialogue/narration contrast — P4 is reported inconclusive, not as a null replication. P4 may not underwrite arm closure under any outcome, which also disposes of F10: the once-outside-the-loop standardisation can only inflate Type I for a statistic that now carries nothing.

A5 (F5, BLOCKING) — the genre confound, named and gated

Holding the hand fixed does not hold genre fixed, and no cell in corpus 2 pairs narrative fiction with narrative fiction: CARL sets a novel against experimental satire, DUFF a novel against travel letters, HAPG-RU and DOLE-RU fiction against travel prose. New gate FC8: P1 is recomputed leaving out DUFF, whose original arm is 0.008 speech and is the extreme case. If the pooled sign or its significance flips, P1 is reported genre-sensitive and the phrase "a property of translating" is denied to this run whatever the numbers are. The residual risk is a declared limit in every branch.

A6 (F6, BLOCKING) — pooling by precision, and a floor on the sign count

P1 is gated on a precision-weighted pooled mean — inverse of the within-cell permutation variance of Δ — and the unweighted mean is reported beside it. The 3-of-4 sign count is taken only over cells with ≥ 10 blocks in both arms in the mode being analysed; §2.3's counts say all four independent non-Romance cells qualify on full text, and that is asserted rather than assumed.

A7 (F7) — §2.4's argument is corrected

The critic is right that "changes no statistic" and "not a licence to reopen" are policy rather than epistemology. §2.4's conclusion is amended to: RS-20260805f §2 and E-20260805f design §2.4 contain a false sentence — dialogue share is not higher in the translation in all four cells, under the per-work computation that matches how the analysis blocks its texts; MACH runs the other way at 0.290 against 0.375. The published statistics are unaffected and are not reopened; the false sentence is recorded here and in this run's limits, and the arm page carries it.

A8 (F8) — the two scored classes are separated

P2b and P2b⁻ are scored on {my, his, her, its, our, their} only. your belongs to {you, your} for P1, P2a and P3 and is excluded from the possessive class, so a pure second-person effect cannot contaminate the possessive statement and vice versa. The censuses' tallies are re-reported under the same split.

A9 (F9, F10) — declared limits rather than changes

P2a's cell set is two against two and heterogeneous, and French HAPG is Romance without being pro-drop; the Delta-z standardisation is computed once outside the permutation loop, as in the predecessor. Both are stated in the result's limits. Neither changes a procedure.