Repository path: workshop/translations/szent-peter-esernyoje/collation.md · rendered 2026-09-09

«Szent Péter esernyője» — collation of the two reachable witnesses, Part I

Why this exists. R05 §1 requires the whole cleaned source to be committed once, before translating, and spans to be indexed into it. Choosing which text to commit is therefore the arm's first decision, and the project has been here before: ARM-atelier-cycle translated four spans of Canth from a posthumous reprint before discovering that the reprint carried five demonstrable corruptions (S100, S105, S110). This collation was run before a word of chapter III was translated, so that the same discovery could not arrive at span D.

The two witnesses

Both are free. R is a transcription of a printing, from page images; M is a modernised digital text of unstated provenance. Mikszáth died in May 1910 and the Révai collected edition is the last of his lifetime. That is the reason of principle; §3 is the reason of evidence.

Neither is clean, so the policy adopted is eclectic — R is the copy-text, M is followed where R slips (R05 D1). This is the same policy collation.md for Canth reached at S100, for the same reason.

§1 Method, and its declared limit

Word-level difflib alignment after a normalisation that folds case, cz→c, vowel length, and punctuation, so that pure orthographic modernisation does not appear as a divergence; word-division differences (kajla szarvú / kajlaszarvú) are then separated mechanically by testing whether the two readings are identical with spaces removed. What survives is inspected by eye.

The limit, stated because the Canth precedent measured its size. At S105, nine of sixty-two apparent Canth divergences dissolved when checked against the page image, three of which would have been published as real variants. No page image has been consulted here. R is itself a transcription, so a divergence in this table may be R's transcriber, R's compositor, or Mikszáth. Where the reading changed an English word it is marked ⚑ and the argument is given; nothing in this file is offered as a statement about the 1895 first edition, which the project has not reached.

§2 Counts, Part I whole

The S149 corrections, and they go both ways. Span E withdrew two rows that were never substantive (¶137, ¶211 — §6) and added three that the method could not or did not report (¶244's comma, ¶252's stop, ¶251's dropped ezüst). Net for chapter IV: 9 → 11. Net for chapter III: 5 → 4. Neither number was arrived at by running the instrument again; both came from reading the two witnesses side by side while translating.

The S144 correction. Span D's translating found a ninth substantive divergence in chapter IV (¶208, a dropped comma in R) that the method in §1 cannot find, because the normalisation folds punctuation before aligning. The count above is therefore a floor on punctuation divergences, not a measurement of them, and the same blindness applies to chapters I–III, which have not been re-read for it.

Two divergences per thousand words are substantive. For comparison, the Canth collation found 59 divergences in 13,522 words, 4.4 per thousand, between a first edition and a posthumous reprint.

§3 The substantive divergences, chapter by chapter

⚑ marks the ones that would change an English word.

Chapter I — ¶ numbers corrected S154

Chapter II — ¶ numbers corrected S154: they were M's, not the copy-text's

Chapter III — the chapter translated this session

Chapter IV — ¶ numbers approximate at S138, pinned exactly at S144 by span D (log D21)

The approximate numbers were wrong by seven to fifteen paragraphs in every case, which is why §5 declared them approximate rather than publishing them. The corrected table:

Only three of the nine fall in span D (¶147, ¶170, ¶171, plus ¶208's punctuation). Five are span E's, and are carried on register.md §Unresolved as span E's first required business.

§4 What the modernisation costs, measured rather than assumed

The reason this collation was run at all is the suspicion that a modernised text quietly removes Mikszáth's marked forms — and in particular his archaic narrative past (mondá, kérdé, ismétlé, mondák), the device the S127 translator's log recorded itself as having under-counted.

Measured across the whole novel, the suspicion is largely wrong, and the number is worth having. A census of the narrative-past forms in both witnesses returns 103 in R and 101 in M.

Inventory qualification, added S144. These 103 are not all of Mikszáth's marked forms. RS-20260809e censused R whole against the full paradigm — 3sg definite -á/-é, 3pl -ának/-ének and -ák/-ék, ikes -ék, the archaic passive and the closed copula — and accepts 181 (142 in the 3sg definite slot alone). Neither figure is wrong; the inventories differ, and S138 published no occurrence-level list, so they cannot be reconciled. The R−M claim below is unaffected, being about the difference between witnesses on the forms it counted. M loses two: ismétlé (III ¶127) and mondák (II ¶10); ch. IV's hoznának for hozanak is a third instance of the same hand, giving three in the whole book. M preserves kérdék, mondá, kérdé, kiáltá, mutatkozék and the rest untouched.

So: the modernised text is substantively faithful on the device, and its editor's hand is inconsistent rather than systematic — which is arguably worse for a reader trying to see the pattern, and is nearly irrelevant for a translator, since three sites in 103 will not change how the layer is handled. This is an honest null on the question the collation was run to answer. What the collation did buy is the four ⚑ readings in chapter III, one of which (¶131) changes the shape of a sentence, and the eight in chapter IV, which span D now starts from rather than discovers.

§5 What is not claimed

• Nothing here is a statement about the 1895 first edition (Légrády), which has not been reached.
• No page image has been consulted; the Canth precedent puts the false-positive rate of an unchecked diff at roughly one in seven.
• ~~The chapter IV table's ¶ numbers are approximate~~ Discharged S144: span D indexed chapter IV against R's own paragraphing and pinned all eight, which had been out by seven to fifteen paragraphs. One of the eight (¶171) also had its referent misread at S138 — the epithet is God's, not the priest's.
• Punctuation divergences are not counted at all (§2, the S144 correction). The alignment folds punctuation by design, so §2's totals are a floor — and the floor rose twice more at S149 (§3, ¶244 and ¶252). Four of the four punctuation divergences this arm has found were found by an eye on the text, none by the instrument.

§6 Two rows this file published and had wrong (S149)

collation.md §3 asserted, twice, a divergence in which R prints an interjection and M prints the third-person pronoun ő — at ch. III ¶137 (Oh! / ő,) and again at ch. IV ¶211 (Óh / ő), the second explicitly described as "the ¶137 corruption a second time, same class."

Both readings of M were checked against MEK-00954 directly at S149 and neither is what M says.

The divergence in both places is Oh/Ó — a final h and a vowel length. That is orthographic modernisation of exactly the kind §1's normalisation folds, and the class of corruption this file asserted twice does not occur in this text at all. The likely mechanism is that S138 read Ó off the raw M side of a diff as ő; it is not recoverable and does not matter. Both rows are struck and §2's counts corrected.

§7 The chapters I–II gate, run punctuation-inclusive for the first time (S154)

R05 §1 makes the copy-text gate the first business of every span, and span F is the first span to sit on chapters I–II. §1's method folds punctuation before aligning and §5 records that all four punctuation divergences this arm had found were found by an eye. So ¶1–¶47 were re-collated whole-stream and punctuation-inclusive — R and M as token streams with explicit paragraph- boundary tokens, orthographic modernisation folded (case, cz→c, vowel length, en-dash/hyphen, guillemet orientation) and nothing else.

52 differences in 1,381 words: 24 punctuation, 18 word-level, 9 word-division, 1 paragraph division. The published table for the same two chapters carries 27 aligned divergences. The punctuation class is dominated by M's comma habits and is not otherwise interesting; the four findings that matter are in §3 above and in D62. Three of them are corrections to what this file published — an off-by-one ¶, three ¶ numbers taken from the wrong witness, and a misstated M form — and they are the same class as the two rows withdrawn at S149: a machine result written into prose by hand, and never re-read. That is now three sessions in a row in which the gate's own record was the thing the gate caught.

What is worth carrying forward. This is not an instrument fault — the sieve did its job and flagged a real difference at both sites. It is a transcription fault at the point where a machine result was written into prose by hand, and it survived two sessions and one downstream design because nothing re-read the witness. The other seven ⚑ readings in the chapter III and IV tables were re-checked against M at S149 and all seven stand.