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

Verification — every flag adjudicated by hand against the Russian

provisional: true (Tier D not calibrated) and internal-judgment-only throughout: every adjudication below is the lead's reading of the Russian, anchored to the source text but not to an external authority.

1. Raw instrument output

Raw: runs/audit-full.out, with a SHA-256 per text.

Read the raw counts as instrument output, not as findings. Four landmarks (G4, G6, G15, G16) fire on texts that are demonstrably correct, including the lead's own — which is exactly what prediction 3 said would be evidence about the instrument.

2. Adjudication, flag by flag

Struck as instrument defects — 4 landmarks, all five texts

G15 (order: "more sentimental than kind") and G16 (order: "less inherited than acquired") — VOID. Struck.

All three translations are correct at both sites:

The landmarks are broken, and the defect is new and general: an order landmark takes the first occurrence of each item across the whole 600-word passage, but both facts are comparisons local to a single clause. G16's pivot item was \b(as|than)\b — function words that occur dozens of times before the clause under test. G15's kind item included \bgood\b, which matches "a very good one" earlier.

Why the self-test gate did not catch this, and it is the lesson of the run. Every self-test fixture is a single sentence. In a single sentence the first occurrence of each item is the one under test, so a landmark that only fails when earlier distractors exist cannot fail its own gate. The gate tested the matcher on material structurally unlike the material it would run on. This is a new failure mode, distinct from note (z): not "the author cannot imagine the word", but "the fixture is not shaped like the target".

The mechanical fix — an order landmark must declare a scope/anchor window, and its self-tests must include a long fixture with a distractor before the site — is written into NEXT.md as the next instrument action. It is not applied here, because the spec is frozen and retrofitting a matcher after seeing which texts it flagged is the defect this project keeps finding in itself.

G4 (value: died ~ten years ago) — VOID for all five. Struck. All five say "ten years". The confusable fifteen fires from "after fifteen years of marriage, he died" — a different and correct fact that falls inside the ±90-character window of the died anchor. A confusable rule that cannot tell two adjacent true numbers apart is measuring proximity, not error.

G6 (presence: born into a poor estate) — VOID for Garnett and Hapgood. Struck. Both render it correctly, in words the accept set did not contain:

• Garnett: "having been born in narrow circumstances"
• Hapgood: "having been born in a poverty-stricken class of society"

NEXT.md note (y) for the third time: the accept set encodes the auditor's vocabulary, and its bias runs against precision. Hapgood's "poverty-stricken class" is arguably the closest of the three to «в сословии бедном», since сословие is a legal estate and "class" carries that where the lead's "estate" is ambiguous in modern English.

G9 (presence: kept on short commons) — VOID for Garnett. Struck. Garnett renders «держал… в чёрном теле» twice over — "made them a scanty allowance. He treated his aunt and sister very shabbily" — and the accept set held neither "scanty" nor "shabbily". Not an error; if anything an amplification, one Russian idiom expanded into two English clauses.

G11 (value: the second year) — VOID for the lead. Struck. The lead wrote "in the very second year"; the confusable first fires from "The first of them was called Marya Dmitrievna", 1,400 characters away.

G5 (conjunction: the husband's four adjectives) — VOID for Hapgood scan A only. Struck as an OCR artifact. Scan A's OCR reads s[)lenetic and stub])orn. Scans B and C read "splenetic" and "stubborn" and pass. This is note (aa) earning its keep: a single-scan audit would have recorded a false factual finding against Hapgood.

UPHELD — one against each translator

G8 — GARNETT. «в пятидесяти верстах от О…» → "about forty miles from O——".

• Source: fifty versts.
• 1 verst = 1.0668 km, so 50 versts = 53.3 km = 33.1 miles.
• Garnett: "about forty miles". Hapgood: "fifty versts". Lead: "fifty versts".

Garnett converts to a domestic unit — a defensible strategy — and the conversion overstates the distance by about 21%. "About" hedges, but not by a fifth. A minor factual deviation, upheld. It is not a gross error: nothing in the novel turns on the distance, and the direction of the choice (domestication) is a strategy, not a slip. What is a slip is the arithmetic.

G1 — HAPGOOD. «(дело происходило в 1842 году)» → omitted entirely.

Turgenev dates the scene in a parenthesis in the second sentence. Garnett: "(it was in the year 1842)". Lead: "(this was in the year 1842)". Hapgood's corresponding sentence reads:

"In front of the open window of a handsome house, in one of the outlying streets of O * * * the capital of a Government, sat two women; one fifty years of age, the other seventy years old, and already aged."

The date is not there. Verified three ways: the string 1842 occurs 0 times in each of the three independent scans of the whole volume; forty-two occurs 0 times; eighteen hundred occurs 0 times. It is not relocated and not spelled out — it is dropped.

A minor factual deviation, upheld. The novel's action is dated by this parenthesis and by nothing else nearby; a reader of Hapgood cannot date the opening scene. It is the omission of a datum, not a mistranslation.

Not audited, and declared

G9's role binding. «держал и сестру и тетку в чёрном теле» elides its subject (the brother). The landmark was downgraded before the run from a cooccur binding test to a bare presence test, because the binding version returned a false ACCEPT on a swapped self-test — subject and rival sit almost equidistant from the attribute, and a proximity test cannot separate them. Shipping it would have been the v3 defect in a new place. Adjudicated by hand instead: all three translations attach the subjection to the brother correctly. Recorded as instrument-unaudited.

3. The count, and what it does and does not support

Upheld factual deviations in this passage: Garnett 1, Hapgood 1, lead 0.

The lead's 0 carries no weight whatever — the lead wrote the landmarks, so its own text is the one text the instrument cannot be wrong about in the flagging direction. It is reported for completeness and should be ignored.

On the Nation's claim. The 1904 reviewer, reading this novel against the Russian, found "almost exactly the same number of errors" in Garnett and Hapgood. This audit, on 18 facts of one 442-word passage 122 years later, finds one minor deviation each. That is consistent with his claim and independent of it.

Four reasons not to make more of that than it will bear:

1. n = 18 facts in one passage. One-versus-one is the least discriminating possible outcome; a passage with two errors in one text and none in the other would have been equally unsurprising. This is consistency, not confirmation.
2. The instrument was wrong 4 times for every 1 time it was right. 20 raw flags across five texts; 2 upheld. A tool with that precision is a referral mechanism, and the finding rests on the hand adjudication, not on the tool.
3. Coverage is 18 facts of a passage containing far more. The instrument tests what its author thought to test. Errors of a kind not landmarked are invisible, and the two found were found at landmarks — the sample is not random.
4. The reviewer counted over a whole novel and three sketches; this counts over 442 words. Agreement in direction across those scales is weak evidence, and the design page's own failure criteria say the instrument disqualifies rather than ranks.

What it does support: the pair is not visibly damaged in this passage on either side, and nothing here disqualifies A Nobleman's Nest on factual grounds. Note that D-20260725-07's ratification already excluded this novel from held-out use, on the reviewers' English-style asymmetry — this audit does not reopen that, and is not offered as a reason to.

4. The finding that was not planned, and matters more

The lead's "blind" translation is measurably not independent of the published ones.

Two exact coincidences turned up in the side-by-side:

• lead "…and his aunt on short commons until…" / Hapgood "…and his aunt on short commons, until…"
• lead "the necessity of pushing his own way in the world and…" / Garnett "the necessity of pushing his own way in the world and…"

Neither is a forced rendering. So the overlap was measured, over the whole passage, against the only available baseline: the two published translations against each other (independent translators, 1894 and 1903).

Shared n-grams, proper nouns excluded (runs/ngram-overlap.json):

The effect grows with n, which is the direction that matters: long shared strings are the ones chance and forced rendering do not explain. At n=7 the lead shares twice as many seven-word strings with Hapgood as the two published translators share with each other.

DOWNGRADED IN PLACE, 2026-07-25 (S026). The two sentences immediately above do not survive re-measurement under a frozen, named metric (E-20260725c-contamination-sweep, verification.md §6). The growth with n is flat under a stricter proper-noun rule (1.684 → 1.676 → 1.625 → 1.647); the multiplier is metric-dependent, 1.647 to 2.037 for this cell; and the published-pair denominator varies by ~4.5× across passages for the same two translators, so "twice" is not a quantity this project can state. The reading is also corrected: under the stricter rule this cell is symmetric (28 shared 7-grams with each published text), and across six cells the lead is the most central text in all six — the general phenomenon is centrality, not memory of Hapgood. What survives: the lead's text overlaps both published texts more than they overlap each other, CI(7) between about 1.6 and 2.0. §5's operational consequence stands and is strengthened.

Among the shared 6-grams: "into the very depths of the azure" (lead/Hapgood) — and the translator's log item 12 records the lead reasoning its way to "azure" for internal consistency with an earlier translation, in a passage where Hapgood had already written the identical six words.

What this establishes, and what it does not.

• It does not affect §2 or §3. Those compare the two published texts against the Russian; the lead's text is not evidence there and its own 0 was already discounted.
• It does not show the lead consulted anything. Freeze order is provable in git: the translation was committed at 2a0bcaa, before any fetch.
• It does show, quantitatively and for the first time in this project, that contamination: suspected on a canonical Russian text is an understatement. T-dvoryanskoe-gnezdo-R04-v1 should be read as contamination: high, and the declared basis is now a measurement rather than an assumption.
• The confound, stated: the baseline pair are two Victorians nine years apart; the lead shares a target-language era with neither, and any two renderings of the same 442 words must overlap somewhat. The direction and the growth with n are what the argument rests on, not the absolute figures.

The consequence the project should act on: lead translations are used as controls in experiments (T-bezhin-lug-R04-v1, T-svidanie-R04-v1). A control whose prose is measurably closer to one arm than the arms are to each other is not a neutral third point. This measurement is cheap — three files and twenty lines — and should be run on every stored lead translation before any of them is used as a control again.

5. Recomputation

Every number above was recomputed from the raw files by the commands recorded in runs/. The three Hapgood scans were compared token by token (76, 78 and 22 pairwise differences, all OCR noise, none landmark-decisive except G5 on scan A). The two Russian sources were compared word by word (442/442, one az.lib.ru typo). The verst arithmetic: 50 × 1.0668 km = 53.34 km ÷ 1.609344 = 33.14 miles.