Repository path: workshop/experiments/E-20260725-svidanie-audit/design.md · rendered 2026-09-09

Frozen design — auditing Garnett and Hapgood on the one passage the comparative evidence is about

Frozen 2026-07-25 (S024), before any published English rendering of «Свидание» was fetched. Verifiable in git: T-svidanie-R04-v1, landmarks.json and this page are committed before the first retrieval of Garnett or Hapgood.

1. Why this runs, and how it is wired to the study limb

SR-20260725-parity-evidence closed the project's best-configured Tier D candidate — Turgenev in Garnett and Hapgood — under exclusion code X2: comparative evidence exists, at passage level, naming both, and it ranks. The document is soundslikewish.org/?p=7860, and the passage it compares is the opening of «Свидание», not «Бежин луг» — the parity remark it contains ("Her choice of 'not … nor … nor' is as good as Hapgood's 'not … not … not'") is about the four-fold negation, which occurs in that paragraph and nowhere else in the sketch.

That search record's own §9 names the observation that would reopen the candidate:

A second, independent assessment of Garnett vs Hapgood contradicts the blog's ranking. Would put case 1 back to genuinely unresolved rather than excluded.

S023 produced two straws in that direction and did not weigh them: on «Бежин луг» the single factual divergence found ran against the blog's direction (Garnett turns «угловатую» into "square"; Hapgood keeps "angular"), and Hapgood's «лапти» → "lindenbark slippers" was the most materially precise rendering in the run. Neither was a designed test.

The wire, in one sentence: the lead's blind translation of the «Свидание» opening supplies the source-derived landmark list that tests, on the very passage the project's only comparative-reception document assesses, whether that document's ranking of Garnett above Hapgood is corroborated or contradicted on factual accuracy.

What this wire can and cannot deliver, said before the run rather than after. The blog ranks on prose merit; this audit measures factual damage. They are different dimensions, and a flag count cannot refute an aesthetic judgment. What an accuracy result can do is supply one independent, non-aesthetic assessment of the same two renderings of the same paragraph, which is strictly more than the project has now (one unrefereed blog post). If it runs against the ranking it makes case 1 less settled than SR-20260725-parity-evidence recorded; it does not by itself reopen the gate, and this page will not claim it does.

2. Question

On the opening paragraph of «Свидание» — the one passage about which comparative criticism of these two translators exists — does either Constance Garnett's or Isabel Hapgood's rendering carry source-anchored factual damage, and does the distribution of damage run with or against the only published ranking of the pair?

3. Materials

4. The instrument

tools/audit_landmarks.py v3 against landmarks.json — 17 landmarks, typed, accept patterns frozen verbatim. Types: one order (C3, the narrator's aspen → settle → sleep sequence), six values (C1 the date, C7 the winter sun, C9 the absent birds, C14 the summer evenings, C15 the low boughs, C16 the hunters' sleep), seven conjunctions (C4 white silk, C5 overripe grapes, C6 new snow, C8 the red-or-gold young birch, C10 the tomtit's steel bell, C12 grey-green metallic, C13 round leaves on long stalks), one cooccur with declared rivals (C11, the pale-lilac bound to the trunk and not to the foliage a few words later), two presences (C2 birch, C17 the dog).

v3 exists because v2 produced two bad observations in S023, both logged and neither fixed then (E-20260725-turgenev-filter/verification.md §7). Both are fixed here and the fixes are demonstrated in runs/instrument-regression.out, which runs the two offending landmarks verbatim from the frozen v2 spec:

Compound relaxation does not suppress over-matching — it makes it visible: bark inside barking reports 'barking' and is adjudicated by hand. That is deliberate. A filter meant to disqualify must not silently prefer the vocabulary its author happened to think of.

v3 also adds a precision-probe gate. Every presence, value, conjunction and cooccur landmark must carry at least one accept self-test tagged "probe": "precision" — a more specific or less obvious correct rendering than the default one. Writing the lindenbark case is what would have caught B11 before the run. cooccur is included precisely because B11 was a cooccur; a gate that skipped the type the lesson came from would be decorative (note (u)).

What the gate does not establish, stated because the limit is the point (note (z)). The same author writes the landmark and its probe. The gate catches type errors, not world errors: an auditor who cannot imagine the word cannot write the probe for it. It forces the attempt and records whether it was made. The fix that would actually close this is organisational — landmark spec and accept sets written by different parties — and this project has one author.

Self-test state at freeze: 68 cases pass across 17 landmarks, 16 of them precision probes. The v2 spec still passes its own 52 cases unchanged under v3 (spec_version gating), so the matcher changes are non-regressive.

Input preparation. Each text is reduced to the opening paragraph alone — no volume apparatus, no translator's preface — matched at its endpoints (the sentence about sitting in a birch grove in September; the sentence about the sleep known to hunters).

5. Controls and exclusions

• Positive control, run before the freeze: T3 scores 17/17. Source-derived by construction. Per note (u) it cannot reveal a matcher defect; the self-tests do that work.
• Negative control: the v2 spec's 52 self-tests, re-run under v3 unchanged (regression), plus the two-landmark regression fixture reproducing defects (a) and (b).
• Deliberate exclusion: no landmark touches the four-fold negation or the season-words inside it. That sentence is the single feature of this passage the lead was primed on before translating — SR-20260725-parity-evidence §4 quotes the blog's report of how each translator handled it. The priming is declared in full on T-svidanie-R04-v1 and cannot be undone; keeping the instrument off it is the only remedy available.

6. Predictions, recorded before the texts were fetched

7. Failure criteria, and what this cannot show

• If both texts pass 17/17, P4 fails a second time on a second sketch. That would be a real finding about published Turgenev translation and is written up as readily as any other, and it would put further pressure on note (s) — "published is not checked" — in the direction of published is more often checked than S022 suggested.
• A flag is not an error. Every flag is adjudicated by hand against the Russian, with the source quoted verbatim and the observed value recorded.
• Accuracy is not the dimension the blog ranked. P6's outcome, in either direction, is evidence about factual damage and about nothing else. It does not settle who is the better translator and this experiment will not say that it does.
• n = 1 passage, one sketch, one pair. Fourth consecutive session on which this is the operative caveat.
• 17 landmarks over 468 words is a filter, not a quality measurement. A translation can be wrong where the list does not look — and the list is deliberately blind to the one sentence about which comparative evidence exists.
• The lead's translation is contaminated (high) and primed on one feature. T3 is a positive control for the instrument, never a standard the published texts are scored against.