Repository path: workshop/experiments/E-20260726-ovid-period-form/design.md · rendered 2026-09-09

Frozen design — period against form against translator identity: four translators of one poem, crossed

Frozen 2026-07-26 (S029). Nothing in §§1–9 is written after seeing a single overlap number. See §4 for the exact, and weaker-than-S028, statement of what the freeze covers.

The wire, in one sentence. The translation limb — 120 Latin hexameters of Ovid rendered blind, in four thirty-line windows fixed by a length-only rule — supplies the subject whose closeness to each of four published translators the study limb reads against a reference distribution built from those same four translators across all fifteen books, in which period and form are fully crossed, so that "the lead sits near a modern voice" can for the first time be told apart from "the lead sits near this one translator" and from "the lead sits near translations in this one form".

1. Why this runs

NEXT.md action 1, top action, third attempt at the same confound, and the first attempt with materials that can actually resolve it.

• E-20260725c-contamination-sweep (S026) found the lead the most central text in six cells of six, and could not distinguish H-lead (centrality is a property of this translator) from H-period (any 21st-century English sits near the centroid of a cell of Victorians).
• E-20260726-period-control (S027) tried to put a modern translator in the room and returned a null: the only freely reachable post-2000 translation of a suitable work was machine output. Note (mm).
• E-20260726-genealogie-period (S028) succeeded in putting a real, verified, human modern translator in the room — Ian Johnston, 2014, against Horace Samuel, 1913 — and the strong period hypothesis failed by reversing sign between essays of one book. Its own amendment A1, forced by the pre-run critic, recorded why that result could not settle anything: one translator per period, so "closer to Johnston" and "resembles Ian Johnston specifically" were the same measurement.

S028's A1 named the fix and the requirement: ≥2 public-domain translations and ≥2 freely readable modern human translations of the same complete unit. This design meets it, and adds a factor S028 could not vary at all — form — because in S028 every text was prose.

2. Questions

• Q1 — is "period" a dimension this instrument can see at all, among published translators? Across fifteen books, do the two same-period published pairs overlap more than the four cross-period pairs? This question does not involve the lead and is the precondition for interpreting anything about the lead. If the answer is no, then H-period was never an available explanation of the sweep's centrality results, and that is a finding about eight sessions of work.
• Q2 — period or year-gap? "Period" as a binary is a coarsening of a continuous quantity. The six published pairs have six distinct year-gaps (10, 71, 78, 88, 149, 159). Does overlap decline monotonically with the gap?
• Q3 — period or form? Two of the six pairs share form (prose–prose, verse–verse) and four do not. Is same-form overlap higher than same-period overlap, and is either separable from the other in the two gap-matched contrasts available (§5.5)?
• Q4 — the lead's profile: period-shaped, form-shaped, or translator-specific? Against each of the four translators separately, is the lead's closeness to the matched book unusual relative to the lead's closeness to that same translator's other fourteen books? Four excess scores per target passage, sixteen in all.
• Q5 — does S028's rank-based replacement for CI behave on a fourth work, fifth language pair, and a four-text cell? Note (nn).

3. Materials

One poem, fifteen books, four published English translations plus the Latin plus the lead — a fully crossed 2 × 2 of period and form.

The crossing, stated as the six published pairs. Every pair is labelled by its period relation, form relation and year-gap. These labels are fixed here, before measurement.

Why these four and not others. Both contemporary translators had to be free, human, and cover the same complete work. Of Johnston's corpus only his Lucretius and his Ovid are still served on web.viu.ca; his Homer pages redirect to johnstoniatexts.x10host.com, which 404s on every content path tried (this is the second session to lose time to that host — S028 recorded it as serving an empty index; it is worse than that, it is gone). Of the works Johnston and Kline share, Ovid has fifteen books against Lucretius's six, so Ovid carries more than twice the reference distribution. Riley and More were then chosen to complete the crossing: one prose, one verse, both public domain, both machine-readable without OCR.

Note (d), applied and paid off. Every bibliographic fact in the table above was verified from an actual tool result this session, not from prior belief. Two priors were wrong: Johnston's Homer is unreachable, and johnstoniatexts.x10host.com is not merely an empty index.

Copyright hygiene (charter §7, rule 5). kline and johnston are in copyright. Both are fetched to a scratch directory outside the repository, reduced to counts, and never stored whole in the repo. What this experiment commits is derived statistics plus brief attributed excerpts. Both consultations are logged in wiki/base/consulted.md.

Encoding, per note (pp). johnston's pages are Word exports declaring windows-1252; kline's declare utf-8. Both declarations were read off the bytes this session and each file is decoded by the charset it declares. No file is decoded with errors='replace'; a decode error is a hard failure, not a degradation. S028 lost every apostrophe in a whole text to exactly this.

4. The freeze, stated exactly — and it is weaker than S028's

S028 could say that nothing English had been fetched. This session cannot say that, and will not imply it.

What was fetched before the freeze: the Latin TEI; Brookes More's complete English TEI; both Project Gutenberg Riley files; Johnston's Ovid table of contents and his Book 1; Kline's Ovid table of contents and his Book 1. All four English texts of Books 4, 9, 11 and 14 therefore sat on disk, unread, while the translation was made.

What was displayed and read before the freeze: file headers and bibliographic metadata; CSS class inventories and structural markers; per-book line and token counts; Johnston's Book 1 opening (~500 words) and Kline's Book 1 opening (~200 words), for the note-(mm) humanity gate. Book 1 is excluded from the targets by construction precisely so that this reading cannot reach a target.

What was not displayed: any English word of Books 4, 9, 11 or 14, from any of the four translators.

Why this is weaker, and what the residual risk is. The freeze rests on the discipline of not printing bytes that were already local, rather than on the bytes being remote. That is checkable in the session transcript but not in git. The honest statement is: the target English was available and was not looked at. A reader who does not credit that should discount Q4, which is the lead-dependent question; Q1–Q3 and Q5 do not depend on it at all, because they involve only published texts, and no choice in this design was made after seeing an overlap number.

The humanity gate, discharged (note (mm)). On Book 1, before target selection:

• kline renders In nova fert animus mutatas dicere formas / corpora as "I want to speak about bodies changed into new forms", and carries a section apparatus keyed to explicit Latin line ranges ("Bk I:1-20 The Primal Chaos"), a dated copyright notice (© 2000), a stated non-commercial licence, and print and audio editions. Prose, by sense, correct.
• johnston renders the same as "My spirit drives me now to sing about / the forms of things changed into new bodies", and carries a signed preliminary note declaring an editorial policy — "the numbers without brackets refer to the English text; those in square brackets refer to the Latin text" — with a first-posted date of 2010 and a revision date of October 2011. Verse, by sense, correct.
• Both are dated years before any language model could have produced them, and both are attached to a named person with an institutional affiliation and a large, consistent corpus. Verdict: human, both. internal-judgment-only; the evidence is quoted so a reader can disagree.
• riley (1851) and more (1922) are print provenance and need no gate; none is claimed.

5. Procedure

5.1 Extraction, frozen

• lat. From the TEI: split at <div type="textpart" subtype="book" n="N">; take the text of each <l n="...">.
• more. From the TEI: split at the book divisions; take all <l> text within, discarding <note> and <milestone> elements. Perseus's subtype="card" divisions are not used for slicing (see §5.4) and their boundaries are ignored.
• riley. From each PG HTML: keep only <div class="maintext">; then delete every <div class="footnote">, <div class="intro">, <div class="contents">, <div class="advert">, <div class="fig">, <div class="mynote">, <div class="mynote plain">, and every <p> whose class is synopsis, explanation, mynote, smaller, or poem. Split at the <h4>/<h5> heading whose text matches ^BOOK [IVXL]+$; discard FABLE headings but keep the prose under them.
• kline. From each per-book page: delete the site chrome (everything before the Contents block and everything after the last content paragraph), delete the table-of-contents block itself, and delete every line matching ^Bk [IVXL]+:\s*[\d\-–,\s]+ (his section headings, which repeat Latin line numbers and would inject digit strings and title words).
• johnston. From each per-book page: delete the bracketed preliminary note, the bracketed book synopsis, and the endnote block; delete the floating marginal section headings, which in this Word export appear as runs of two or more consecutive ALL-CAPS words inside a paragraph of ordinary prose; delete bracketed Latin line numbers \[\d+\] and unbracketed standalone integer tokens (his English line numbers and footnote markers).
• lead. The four ## Book N, lines 51–80 blocks of translation.md, up to the next ## or ---, with the leading and trailing ellipsis characters removed.
• Common. Strip tags, unescape entities, collapse whitespace; then delete standalone digit tokens, applied symmetrically to all texts.

Structure was inspected before the freeze — heading lines, CSS class names and markers only, never target prose — which is how the five splitters above could be written down rather than tuned afterwards.

5.2 The statistic, frozen

Exactly the metric of tools/ngram_overlap.py, unchanged from E-20260725c-contamination-sweep §4 and used unchanged by S027 and S028: NFC normalise, map curly quotes to straight and em/en dashes to space, lowercase, delete everything outside [a-z0-9'], whitespace-split. For a pair of texts, count shared n-gram types for n ∈ {4, 5, 6, 7}, expressed as shared types per 1000 tokens of the shorter text. Headline n = 5.

5.3 The free parameter, bracketed rather than chosen (note (kk))

• V-none — no proper-noun exclusion at all. This is the headline variant, because Ovid is saturated with proper names, the lead's translator's log records a deliberate decision to retain all of them, and under V-none cell composition is irrelevant by construction.
• V-frozen — the R1 ∪ R2 name rule exactly as tools/ngram_overlap.py computes it.
• No third variant is declared. Note (ii)'s typographic trap — a sentence-ending character outside the exemption list .!?:" — is checked for empirically (F6) rather than pre-empted with a guessed variant, because four genres in, the failure has never been the same character twice.

5.4 Alignment: the book, and nothing finer

The unit is the book, for every published–published measurement. Book boundaries are unambiguous in all five texts and are the one division every edition of this poem agrees on. No sub-book alignment is attempted anywhere in this design. This is deliberate: sub-book alignment defects cost S027 27 of 42 pairs and forced a declared post-hoc realignment, and Perseus's card divisions, Kline's line-range headings and Johnston's bracketed Latin line numbers are three different and mutually non-nesting schemes. Refusing to align below the book removes the entire failure class.

The consequence is accepted openly: the lead's 120-line translation is compared against whole books, not against the corresponding 120 lines. §5.5 is built so that this does not invalidate the comparison.

5.5 The lead's statistic: excess over each translator's own baseline

A short text compared against a long one yields a rate that is not comparable to a long-against-long rate, so the lead is never ranked inside the published–published distribution. Instead, for each published translator T and each target book B ∈ {4, 9, 11, 14}:

• compute lead(B) ~ T(B) — the matched value; and
• compute lead(B) ~ T(b) for all fourteen b ≠ B — the mismatched values, which measure what the lead's wording shares with this translator's English in general, on content it is not translating.
• The statistic is the rank of the matched value among all fifteen, reported as "k of 15", k = the number of the fifteen values strictly greater than the matched value, ties counting as not greater (the direction conservative against the lead). k = 0 means the matched book is the single closest.

This is the same length, the same lead text, and the same translator on both sides of the comparison, so translator-specific vocabulary, form, register and text length all cancel. It is note (nn) — stop dividing, measure the distribution — applied per translator rather than per work.

Sixteen rank observations (4 books × 4 translators) is the lead's whole evidential contribution, and it is the weak part of this design, as three sections were the weak part of S028's.

5.6 The lead's form is declared, not discovered

The lead translated in verse, decided and written into translation.md before the Latin was read. The lead is therefore form-matched to more and johnston and form-mismatched to riley and kline, and this is a confound on Q4 that cannot be removed, only handled: §5.5's per-translator baseline absorbs it to the extent that a translator's form shows up in their mismatched books too, which is total for form and partial for form-linked diction. Any Q4 result in which the two verse translators pattern together is uninterpretable between form and period, and will be reported as such.

5.7 Controls

• C1 — the same-book requirement (note (p): a case the instrument should pass). For each published pair, matched-book overlap must exceed the median mismatched-book overlap of that same pair. Mismatched pairing is book i against book i+1, cyclic over the fifteen. Two translations of the same book must share more than two translations of different books; if they do not, the statistic is measuring English and not translation, and every number below is void.
• C2 — book length is reported, never assumed. Token count of all five texts for all fifteen books is tabulated. Any published pair whose token counts differ by more than 2× in a given book is excluded from that book's reference values, and the exclusions are named and counted.
• C3 — the lead's four target windows are not length-matched to each other (200/193/193/193 Latin tokens; English lengths will differ more). All lead figures are per 1000 tokens of the lead's text, and the four windows' English token counts are reported.

6. Predictions, registered

Registered before any overlap number existed. P5–P7 are the lead's fourth registered self-estimate; of the first three, two missed in the flattering direction and one held at the weakest level its wording allowed (note (jj)). Per note (qq), every prediction below is per-unit or per-pair, and none is a prediction about a mean over units.

• P1 (extraction sanity). All five texts yield exactly 15 books. riley books I–VII come from #21765 and VIII–XV from #26073 with no book appearing twice or missing.
• P2 (structural alignment). Pearson r between lat book token count and English book token count is ≥ 0.90 for each of riley, more, kline, johnston across the fifteen books. Below 0.90 for any text, that text's extraction is not established and it is dropped with the failure reported.
• P3 (C1, the floor). Under V-none at n = 5, every one of the six published pairs has matched-book overlap above its own median mismatched-book overlap, in at least 13 of 15 books.
• P4 (Q1, the period test proper). Under V-none at n = 5, in at least 11 of 15 books, both same-period pairs (kline~johnston, riley~more) rank above the median of the six pairs.
• P5 (Q3, form). Under V-none at n = 5, in at least 11 of 15 books, the two same-form pairs (riley~kline, more~johnston) rank above the median of the six pairs. P4 and P5 are registered as independent predictions and both may hold, or neither.
• P6 (Q2, gap monotonicity). Under V-none at n = 5, the Spearman correlation between year-gap and overlap across the six pairs is negative in at least 11 of 15 books.
• P7 (Q4, the lead is period-shaped). Under V-none at n = 5, the mean rank k (§5.5) over the four books is lower for kline and johnston than for riley and more — i.e. the lead's matched-book closeness stands out more against the contemporary translators' baselines than against the historical ones.
• P8 (Q4, the lead is matched-book-sensitive at all). Under V-none at n = 5, k = 0 — matched book strictly closest — in at least 12 of the 16 book × translator cells. If this fails badly, the lead's translation is not detectably of the book it translates, and P7 is uninterpretable regardless of how it comes out.

7. Failure criteria, pre-committed

• F1. Any text failing P1 is dropped, and if that leaves fewer than two translators in either period the cell is void and the session reports the extraction failure and no overlap numbers.
• F2. Any text failing P2 (r < 0.90) is dropped on the same terms.
• F3. If C1 fails (P3 false for any pair), no rank statement is made at all, in this session or by citation later.
• F4. If either contemporary text shows any sign of machine origin when the target books are finally read — a homograph rendering rather than a sense rendering; a fluent clause referentially absurd against the Latin; a clause present in the Latin and silently absent — the cell is void. One confirmed instance suffices. These are the three classes the S027 machine text exhibited.
• F5. If P8 fails (fewer than 12 of 16 cells at k = 0), P7 is reported as uninterpretable, not as failed.
• F6. Note (ii) check: for each text, the count of tokens whose final character is outside .!?:" yet which end a sentence. If any text's count exceeds 2 per 1000 tokens, V-frozen is reported as unreliable for that text and only V-none is used for it.
• F7 (reading rules, pinned in advance per note (r)). "Ranks above the median of six" means: the pair's value is strictly greater than the mean of the 3rd and 4th largest of the six values. An exact tie counts as not above. Spearman is computed on the six (gap, overlap) points with no tie correction; if two overlap values tie exactly the book is excluded from P6's count and the exclusion is reported.

8. What each outcome licenses, and what it does not

Nothing here is a quality claim about any of the five translations. Overlap is not merit, in either direction. The lead does not judge its own translation (charter §5); this experiment measures wording overlap and nothing else, and every figure it produces is internal-judgment-only and provisional until Tier D passes.

Extent, stated honestly. One poem, one language pair, four published translators, fifteen books, four target windows. The published–published reference distribution is 6 pairs × 15 books = 90 measurements and is the strong part. The lead's placement rests on 16 rank observations and is the weak part, and precision on the 90 does not repair the 16. "Contemporary" here means two translators, dated 2000 and 2010, one of them a professor emeritus and the other a solo digital publisher — that is not a sample of 21st-century English, and no sentence in the result may suggest it is.

9. Pre-run critic

Per charter §8, an independent pre-run critic pass on this frozen design, routed to a non-Anthropic panel model, before any measurement is run. S028's critic pass cost $0.0806 and was recorded as the best-value spend in the ledger; its two most valuable findings were rulings that predictions were unfalsifiable as written, both about tie-handling and rank direction. Per note (rr), this critic is pointed at the statistic definitions first and the argument second.

Findings and the response to each go in critic.md. Findings requiring a design change are applied by amending this file with the amendment marked and dated, never by silent edit; findings declined are recorded with the reason.

10. Amendments — added 2026-07-26 after the critic pass, before any measurement

Every amendment below was written before a single overlap number existed. analyse.py did not exist and runs/ contained only the critic response when this section was committed. Findings and the response to each are in critic.md; the numbering here is referenced from there.

A1 — §5.5's central claim was false and is replaced (critic Task E)

§5.5 said the per-translator baseline makes "translator-specific vocabulary, form, register and text length all cancel." That is wrong. Holding the translator fixed on both sides of the comparison holds those four things constant; it does not cancel episode-specific diction, named-entity density, literalness, or the interaction between any of those and the lead's particular window. The sentence is replaced by:

Both sides of the comparison use the same lead text and the same translator, so that translator's vocabulary, form, register and text length are held constant rather than eliminated. What is not controlled is episode-specific diction and named-entity density: the matched book is the one whose story the lead is telling, and shared proper names alone will raise the matched value. A2 exists because of that.

A2 — P7 must survive name exclusion or it is a name-density effect (critic Task B.6)

The critic's ruling that k = 0 "primarily demonstrates content alignment" is accepted. Consequences:

• P8 is demoted to what it always was: a check that the instrument can tell which book the lead translated. It is not evidence about period, form or identity, and no sentence in the result may use it as such.
• P7 is additionally required to hold under V-frozen (proper names excluded) as well as under V-none. If the direction appears under V-none and vanishes under V-frozen, P7 is reported as a proper-name-handling effect, not a period effect — which is a substantive finding about the instrument, given that the lead's translator's log records a deliberate decision to retain every proper name.

A3 — P4 and P5 are re-run on the residual, to absorb era-correlated extraction residue (critic Task B.2, C)

The two historical texts reach the project through two print-digitisation routes; the two contemporary texts through two web sites with navigation and section apparatus. Residue shared within an era inflates that era's pair and would manufacture a false P4. Control:

• P4b / P5b. The same tests, computed on the residual: a pair's matched-book overlap minus that pair's own median mismatched-book overlap (the C1 quantity, already required). Boilerplate or apparatus shared by two texts raises their matched and mismatched values together, so it cancels in the difference; genuine same-passage agreement does not.
• Reading rule. The raw and residual versions are reported side by side. Disagreement between them is reported as a defect of the measurement and is not resolved in favour of either. A period claim requires both.

A4 — post-drop and post-exclusion computability (critic Task A, §5.5 / P7 / P8 / F1 / F2)

• If a translator is dropped under F1 or F2, its four lead cells are not computed, the pair table loses the three pairs containing it, and every count is restated as "k of N" with N printed. No percentage, proportion or percentile language is used anywhere after a drop.
• If a book is excluded for a pair under C2, that book contributes no value for that pair and the pair's denominators (median mismatched, 15-of-15 counts) are computed over what survives, with the surviving count printed beside every figure.
• If a book is excluded for a pair, the lead's rank against the affected translator is computed over the surviving books only and reported as "k of N".
• F1's non-void case: dropping a text does not re-open the freeze. Predictions are evaluated on the surviving texts exactly as written, with reduced N stated.

A5 — tie, threshold and degeneracy rules (critic Task A: P2, P3, F2, F3, F7)

• Ties are decided on exact integers, never on a rounded rate. Two overlap values are equal iff their (shared-type count, denominator token count) integer pairs are equal. No floating-point equality test is used anywhere.
• P3: "above its own median" means strictly greater. Equality counts as not above.
• P2: if Pearson r is undefined for a text — zero variance in either series, a missing book, a non-numeric or empty count — that text fails P2 and is dropped under F2. Undefined is not a pass. r = 0.90 exactly passes (F2 fires below 0.90, consistent with P2's "≥ 0.90").
• F3 fires on the per-pair condition and not on P3's aggregate: if any published pair has matched-book overlap that fails to exceed its own median mismatched-book overlap in at least 13 of its surviving books, then no rank statement is made for that pair. If two or more pairs fail, no rank statement is made at all. P3's 13-of-15 threshold is the per-pair condition; there is no separate aggregate.

A6 — P7's reading, pinned to one of the three the critic found (critic Task A)

P7 is evaluated in the strict form: the mean rank k over the four target books, computed per translator, must be lower for kline than for both riley and more, and lower for johnston than for both riley and more. Both contemporary means below both historical means. A single crossing pair means P7 fails.

Reported as secondary, not as P7: the paired-by-book version (in how many of the four books is min(k_kline, k_johnston) < min(k_riley, k_more)), and the eight-cell group means. These are descriptive and carry no registered threshold.

A7 — F4's three tests made concrete; its residual declined (critic Task A)

Applied to the four target books when they are finally read, against the Latin:

1. Homograph rendering — a Latin word rendered by an English homograph or cognate-shaped form rather than by sense (e.g. nefas as "nefarious [thing]" where the syntax requires a noun object of videre).
2. Referential absurdity — a fluent English clause that is false against the Latin it renders, in a way a competent reader of the Latin would not produce.
3. Silent omission — a clause present in the Latin and absent or garbled in the English without an editorial mark.

Adjudicator: the lead, reading the English against the Latin. One confirmed instance voids the cell. Every instance found is quoted in the result with the Latin beside it so a reader can disagree. The verdict is internal-judgment-only.

Declined, as unfixable: the critic is right that this is not a coding rule. No rubric identifies machine output; reading does. S028 declined the same finding for the same reason. It remains the weakest joint in the humanity gate, and it is why the gate is also discharged on Book 1 by an independent route (dates, licence, institutional affiliation, corpus consistency).

A8 — F6 replaced by a diagnostic that needs no sentence splitter (critic Task A)

F6 as written required knowing which tokens "end a sentence", which requires a sentence splitter the design never specified. Replaced by:

F6′. For each text, count tokens whose final character is one of …‽!?.:;"' plus any character that R1's exemption list .!?:" does not contain but which is followed in the raw text by whitespace and a capitalised word. Report the count per 1000 tokens per text. If any text exceeds 2 per 1000, V-frozen is reported as unreliable for that text and only V-none is used for it.

This is computable from the raw text with no segmentation model, and it measures the thing note (ii) actually cares about: capitalised words that a name-detector will misread because the preceding character is not in the exemption list.

A9 — the confounds that stay untested, named (critic Task B.1, B.4, B.8)

Recorded as untested and binding on how any result is worded:

1. Within-period translator habits. Riley and More may share Victorian editorial norms, biblical diction, name spellings, and an older Latin source edition; Kline and Johnston may share plain-English pedagogical convention and common dependence on the same modern commentaries. Two translators per period is better than one and is not a sample of a century. No result may be worded as a fact about a period.
2. Same-form pairs are also particular translator pairs. A verse–verse effect could be conventional formula renderings or lineation artifacts rather than form.
3. Literalness. n-gram overlap rewards literal, phrase-reusing translation. If the contemporary translators are more literal, the lead can look contemporary with period playing no causal role. The project has no literalness measure and none is invented here.

A10 — the contraction channel, carried over from S028 A2 (critic Task B.3)

Orthography — contractions, hyphenation, apostrophes, archaic spelling — can manufacture a period separation before any wording is involved. S028 found this channel live on a 1913/2014 pair and found that the lead used no contractions at all in 1,135 words.

• Diagnostic: per text per book, the count of contraction tokens (matching \w+'\w+ after normalisation, excluding possessive 's) per 1000 tokens.
• V-expand: a fourth variant applying S028 A2's fixed expansion list to all texts before tokenising, unchanged and not re-tuned here.
• Rule: if a period effect reported under V-none does not survive V-expand, it is reported as a contraction artifact and P4 is treated as failed.

A11 — the windows are not a sample of the poem (critic Task B.7)

All four target windows are lines 51–80 of their books. Book openings carry invocations, transitions, genealogies and a raised density of named entities, and lines 51–80 sit close to them. The selection rule is length-only and pre-committed, so it is not changed after the fact; the honest response is the statement itself:

The four windows are four fixed offsets in four books, not a sample of the Metamorphoses. Anything found about the lead's placement is found at one position in a book, four times.

A12 — book-boundary landmark check (critic Task B.11, Task E)

P2's correlation can pass while a book carries spillover from its neighbour — the failure mode riley's missing end boundary would produce. Added:

C4. For each text and each book, print the first eight and last eight tokens. The first tokens of book n+1 must not appear at the end of book n, and no book's extracted text may contain the heading string of another book. A failure here drops the text under F2's terms.

This is the check that catches "plausible but inflated" books, which is the corruption most likely to survive every other test in this design.

A13 — §5.1 rewritten as an implementable specification (critic Task C, Task D)

§5.1's prose is replaced by the following. Structure was inspected to write it — tag inventories, CSS class names, inline style attributes, heading strings — and no target prose was displayed.

• Global. Every file is decoded by the charset it declares (johnston windows-1252, kline utf-8, PG utf-8); errors='strict' — a decode error is a hard failure (note (pp)). Tags are replaced by a single space, never deleted in place, so no n-gram is fabricated across a block boundary. Entity unescaping precedes whitespace collapse. Digit-token deletion happens after §5.2's normalisation to [a-z0-9'], so [12], 12, and l.12 all reduce to 12 first and are then removed uniformly.
• lat, more. TEI: split at book divisions; take the full descendant text of each <l> (equivalent to itertext()), after removing <note> subtrees entirely. <milestone> and card divisions are ignored.
• riley. Keep only <div class="maintext">. Delete every <div> whose class is one of footnote intro contents advert fig mynote "mynote plain" and every <p> whose class is one of synopsis explanation mynote smaller poem center. A book runs from the heading matching ^BOOK\s+[IVXL]+$ up to but excluding the next such heading, or the end of maintext — the end boundary the critic found missing. FABLE headings are deleted; the prose under them is kept.
• kline. Keep only <div class="default-style poem">. Within it delete every <h1>, <h2>, <ul>, and every <p class="text-center"> (title, book number, the two site-links navigation lists, the section headings, the table of contents ul.content-list, the copyright line and the licence line). What remains is the translation paragraphs. No line-matching against post-parse text is used — the section headings are <h2> elements and are removed structurally, which is what the critic asked for.
• johnston. Delete <script>/<style>, then delete every <span> whose inline style declares font-size:10.0pt. In this Word export the translation is uniformly 14 pt and the 10 pt spans are exactly the apparatus: square brackets, marginal section headings, English line numbers, bracketed Latin line numbers, and footnote markers. Verified on Book 1: of 259 such spans, the number containing any lowercase letter is 0. The tool re-runs that assertion on every book and fails loudly if any 10 pt span contains lowercase prose. This replaces the ALL-CAPS heuristic, which the critic correctly said would delete genuine emphasis and miss punctuated headings.
• lead. The four ## Book N, lines 51–80 blocks up to the next ## or ---. Ellipsis characters are replaced by a space, not deleted, so no words are joined.

A14 — the denominator contradiction, resolved by assertion and check (critic Task C)

§5.2 normalises per 1000 tokens of the shorter text; C3 said "per 1000 tokens of the lead's text". These agree only if the lead's window is always shorter than the published book, which the design asserted nowhere. The tool asserts it per cell and fails loudly if it does not hold. All 60 lead cells are expected to satisfy it by a wide margin (a 30-line window against a whole book), and the assertion is reported as checked rather than assumed.

A15 — the two gap-matched contrasts, registered properly (critic Task D, Task E)

§2's Q3 promised separability "in the two gap-matched contrasts available (§5.5)" — a wrong cross-reference to a section containing no such analysis, and no statistic, threshold or direction was ever registered. Note (r) again. Registered now:

• G1: more~johnston (same form, verse; gap 88) against more~kline (different form; gap 78). Same first translator, near-equal gaps, form differing.
• G2: riley~kline (same form, prose; gap 149) against riley~johnston (different form; gap 159). Same first translator, near-equal gaps, form differing.
• Statistic: in how many of the 15 books does the same-form member exceed the different-form member. Reading rule: ties count against the same-form member. 12 of 15 or more in both G1 and G2 is a form effect at gap held approximately constant; anything less is reported as no separation. The gaps are not exactly matched (88 vs 78; 149 vs 159) and the direction of that mismatch favours the same-form member in G1 and disfavours it in G2 — which is why both contrasts must hold, and why neither alone would be reported.

A16 — the two conclusions the design was not entitled to draw (critic Task E)

§8's table is amended in two rows:

• "neither holds" no longer licenses "period is not a dimension this instrument can see, and neither is form." Failure to clear an 11-of-15 threshold with two translators per cell and fifteen correlated books is not evidence of absence. Replaced by: no period effect and no form effect is detectable at this power — four translators, two per period, fifteen books of one poem, one instrument — which removes the positive support H-period never had and does not establish its negation.
• "P5 holds, P4 fails" no longer licenses "the sweep's centrality results are a form artifact, not a period artifact." Both patterns can arise from selected-translator identity, from extraction asymmetry, or from the metric's response to formatting. Replaced by: form is the better predictor of overlap than period in this cell, which makes a form explanation of the sweep's centrality results worth testing and does not establish it.

A18 — C4 stores the verdict, not the evidence, for the two in-copyright texts (2026-07-26, post-run)

A12 specified printing the first and last eight tokens of every text for every book. Applied literally that would place ~480 tokens of kline and johnston into a committed file, in sixty fragments. Charter §7 rule 5 permits brief attributed excerpts and this is defensible, but S028's precedent for the same translator was to store none, and the verification function does not require the evidence: what C4 needs to establish is that no book's extracted text contains another book's opening.

So for kline and johnston, runs/results.json records the token count and the boolean, and res["C4_pass"]. For lat, riley, more (public domain) and lead (project-authored) the tokens are recorded as specified. C4 passes: no book of either in-copyright text contains any other book's opening. This is a reduction in what is stored, not a relaxation of the check.