Repository path: wiki/findings/essays/ES-20260810-craft-report-legend.md · rendered 2026-09-09

The craft report — «Szent Péter esernyője» Part I, four visits, and a collation that predicted the translation badly

ARM-legend step 4, the arm's closing deliverable (charter §3 Track 1, the third long work). It states the whole-part figures, settles the four questions register.md §Unresolved put to it, and says what four visits made visible that the two shorter works could not. Its two central claims are measured rather than recollected: RS-20260810x-legend-again and the Károli check inside it. Every evaluative sentence about this translation is the lead's own and the page carries internal-judgment-only; Tier D is NOT PASSED and no jury has judged a word of it.**

1. Part I, in figures

The configuration this arm was constituted for. Verga had a published English and was translated in fragments; Canth was translated whole with no English in existence. Mikszáth is the first work this project has translated whole where an independent published English of the whole also exists and is free (B. W. Worswick, 1900), and it is the project's first work in Hungarian. The comparator stayed shut until each span was committed, and was then measured against: spans D and E came back none, at a longest run of 12 and 11 tokens.

2. The four questions, settled

2.1 V12 — not founded, not refuted, and the arm closes saying so

The register's most contested rule asserts that Mikszáth's archaic layer outside speech tags is marked by scriptural lexis and syntax, never by bent verb morphology, and therefore that an English translator should reach for the King James Bible rather than for quoth and -eth. It was written at S144 and demoted the same session. Two designs have now failed to found it — RS-20260809e on distribution (p = 0.1253, 40 of 55 tokens immovable), RS-20260810b on paragraph-level Károli overlap (+0.0087, p = 0.4392, and negative against an archaizing-secular control).

This session ran the check RS-20260810b §7 pointed at, at the grain the reading proposes. At the five marked narration sites the arm renders in King James English: 8 of 11 marked verb forms occur in the reachable Károli, but only 3 of their 20 immediate bigrams do, and only one is a construction rather than a coincidence — ¶244's «Lőn nagy ámulás», where Károli has the identical frame seven times over (lőn nagy jajgatás, lőn nagy öröm, lőn nagy csendesség ×3, lőn nagy földindulás, lőn nagy hirtelenséggel).

And it does not settle V12, because there is no comparator. lőn is at home in nineteenth-century archaizing fiction as well as in the Bible, and a check with one comparator cannot tell a biblical echo from ordinary period archaism. The pattern found — Károli's morphology at 8 of 11, Károli's phrasing at 1 of 20 — is what one would see if V12 were wrong about the Hungarian, and equally what one would see if Károli's morphology were simply the common property of all archaizing Hungarian.

So the arm closes with V12 demoted and unfounded, which is a null it owns. What it can say is narrower and worth keeping:

• The register now states V12's two halves separately. A claim about the Hungarian (the layer is marked by scriptural lexis) — unsupported in both directions. A decision about the English (carry it with King James phrasing) — a compensation, because English has no productive archaic preterite that is not comic, and labelled as a compensation rather than as a description.
• RS-20260810b's mask removed lőn and hozának from ¶244, so its null does not test overlap carried by those forms. That is a fact about what was masked. It is not evidence that the device is there.
• The design that would decide it is now specified — unmasked, two-comparator, difference-in-differences at clause level (RS-20260810x §10) — and it is the arm's successor question, not its unfinished business.

Erratum 1 does not fire. It would have flattened all five marked passages together on the strength of a null. A p-value above 0.05 may not delete two spans of translation — the pre-run critic's amendment A9 at S149 — and neither may a comparator-free frequency count.

2.2 D44's Worswick prediction — failed, and the restored line stays

D44 restored a line of dialogue present in M and absent from R (¶216a, «És mibe kerül az egész?») on an argument from mechanism — saut du même au même on the repeated «mibe kerül», M's demonstrably subtractive hand elsewhere, and ¶218 answering a question R does not print. D44a registered, before the comparator was opened, that Worswick 1900 should carry Srankóné's question if the line is authorial.

He does not. And in the same six lines he also drops the mágnács footnote, the three hundred sheep and the chalk — «kiszámította krétával» becomes proceeded to reckon out. A translator who deletes the chalk from the next clause is a poor witness to whether a nine-word question stood in his copy. D44 stands with reduced confidence; ¶216a is kept.

The observation that cuts the other way is kept on the page too. Worswick has "the white altar-cloth" at ¶227, consistent with M's oltárterítőt against R's asztalterítőt — and equally consistent with a translator tidying a peasant's table-cloth into the correct ecclesiastical word, which is exactly the mechanism D45 attributed to M's editor. It settles nothing and it is recorded.

2.3 What the change of copy-text cost — totalled, and the surprise is what it bought

The arm inherited two chapters rendered from M and then set its copy-text to R. Nothing had totalled the bill. Span F rendered those two chapters again from R and coded the difference.

The bill: seven sites in 1,381 words — ¶6 fia/fiú, ¶20 leng/teng, ¶20's italic, the ¶18/¶19 paragraph division, and the place-name and personal-name orthography at ¶5 and ¶11 — of which two were pre-exposed by collation.md's own argument. On Canth the same class was zero.

Three of the seven predicted sites cost nothing at all, and that is the finding. ¶27's mondák/mondták — a marked archaic past against an ordinary one — reaches the same English, because V6 sends every speech tag to a plain English past. ¶32's fevde/fedve reaches the same English because R is corrupt there and both renderings follow M. ¶11's Máthé/Máté reaches the same English because R contradicts itself, printing Máté six times and Máthé once.

So: of the four ⚑ readings collation.md marks in these chapters as ones that will change an English word, one did. A collation's ⚑ is a prediction about the translation, and on this span it was wrong three times in four — twice because a register rule had already absorbed the difference, once because the marked witness was the wrong one to worry about.

And two classes the collation had never reported were both found by translating, not by the instrument: R prints 47 paragraphs in chapters I–II where M prints 46, and R italicises one word — _anyós_ at ¶20 — where M has no italics anywhere in either chapter. Paragraph division is the one class guaranteed to reach the English, because the paragraph is the unit the translation is indexed in. The method folds punctuation before aligning and _ is punctuation, so it is blind to both.

Three rows this repository published were also wrong in their particulars — an off-by-one ¶, three chapter II rows indexed to M's paragraph numbers in a file whose declared copy-text is R, and an M form given as mondta where M reads mondták. That is the same class as the two rows withdrawn at S149 and the eight ¶ numbers corrected at S144: a machine result written into prose by hand and never re-read. Three visits in a row have found one. The gate is worth running; the record of the gate is worth re-reading, and this arm never once found the record clean.

2.4 The register's backwards blindness — measured, and smaller than the count of everything else

ARM-longwork named this on Verga at S047 and RS-20260806d measured it on Canth at S120: a register catches forward drift and is structurally blind backwards, and eleven sites the closed register would have rendered differently were caught by no erratum at all.

On this work, span F gives: 160 divergence sites in 1,963 English words, of which 5 are coded exclusively to a later register rule and 9 have one governing the wording. The rest is copy-text (6), whole-work argument (6), two corrections, and 141 sites the translator-coder assigned to no cause at all — a residual that was not blind-audited here and is not interpreted.

What this does not say. It does not say the register is useless. The count of rule-attributable sites is a count of places where the rule and the habit disagreed, and a rule that codifies a habit does its work invisibly. It does not isolate the register either: copy-text, register and twenty-two sessions of the translator's experience all changed at once.

What it does say comes from a frame fixed before the comparator was opened — the 22 decisions the translator wrote into the frozen log. Twelve came out the same and ten came out different. On the things this translator himself thought were decisions, the closed register and twenty-two sessions changed slightly under half.

The single most deflating instance is D66. Its log entry argues for two paragraphs that árvalányhaj should be orphan-girl's hair rather than feather-grass, because an orphan girl is at that moment being carted into the valley the grass grows in. The translator of S132 — no register, no census, no argument — wrote orphan-girl's hair. The argument produced the choice the habit had already produced, and note (bhb)'s measurement of the lead matching itself is the alternative explanation the design cannot exclude.

3. What four visits made visible that the two shorter works could not

3.1 A village has no word for an umbrella, and that is a rule about vocabulary the source lacks

V15 is the arm's best rule and it could not have been written on a fragment. Glogova has no umbrella, so every villager names the thing differently — szerszám implement, vászontányér canvas platter, vászonsátor canvas tent, vászonkarika forma canvas-hoop affair, furcsa alkotmány queer contrivance — and normalising any of them to umbrella deletes the fact that these people are looking at an object they have no name for. The title object turned out not to be the problem; the villagers' failure to name it is.

3.2 A designation the source varies must not be collapsed, and one it repeats must not be varied

V21 is one rule with two faces, and only length showed they were the same rule. The dead woman of chapter I is tanítóné, rektorné and kántorné inside six paragraphs — schoolmaster's wife, rector's wife, cantor's wife — because in a village these were one office and the gossips reach for whichever comes first. Forty paragraphs later the village judge closes an argument with «Míg nem méltóztatom máskép intézkedni», and the Glogova schoolmaster hands the same pompous verb back to the district magistrate twice: «Méltóztatik érteni?»

English has a locally better word at each of those three sites — deign, see fit, care to — and taking any of them dissolves the return. All three take pleased to. The inherited rendering, made before either half of the rule existed, broke the recurrence and collapsed the variation, which is 4 of the 10 differing rows in §2.4's frame.

This is the same question ARM-recurrence is now asking of Gogol's «значительное лицо» at 35 occurrences, arrived at from the other end: there, whether published translators carry a repetition; here, what it costs a translator to notice one is happening at all.

3.3 The archaic layer is a real problem and the project still cannot say what marks it

Three sessions and two experiments have been spent on it (§2.1). What is settled is only this: the two past tenses in speech tags are a measured non-loss — RS-20260808g found four independent hands rendering mondá and mondta identically at 11 of 11 sites (delta 0.0000) while the same measure saw a tense change at 1.0000 — and V6 stands on that measurement rather than on the translator's ear. Outside the tags nothing is founded. That is a real limit of four visits, and it is the arm's successor question rather than its failure.

3.4 The copy-text gate must be run on every span, including the ones inherited

R05 §1 makes the gate the first business of a span, and this arm ran it four times. Each time it found something the previous run had missed or got wrong: the ¶ numbers out by 7–15 (S144), two published rows that misdescribed the witness (S149), three punctuation divergences invisible to the method (S149), and in span F an entire paragraph division, an italic, and three more wrong rows (§2.3). A gate run once is a gate that has been run once.

4. The honest nulls

• The modernised text does not lose the archaic layer. The suspicion the whole collation was run on: a census of narrative-past forms returns 103 in R and 101 in M. Three sites in the book, and the editor's hand is inconsistent rather than systematic.
• V12 is not founded, twice by experiment and once by a lexical check (§2.1).
• D44a's registered textual prediction failed (§2.2).
• Three of P-A's seven named copy-text sites cost the English nothing (§2.3).
• P-C failed at 160 against a registered 90–140, and the band was arithmetically wrong — built by applying a per-English-word rate to a Hungarian word count. P-B is invalid as registered: it never said how a site with two causes would be counted, and the two readings fall on opposite sides of the band.
• RS-20260809e demoted V12 six hours after it was written, in the same session, on the arm's own study limb.

5. What this arm did not deliver

• Part I only. The novel runs to some 55,000 words and the remaining four parts are a detective story built on this legend. Nothing here is a reading of the novel.
• No jury has seen a word of it. Every evaluative sentence about this translation is the translator's own; Tier D is NOT PASSED.
• The blind coding RS-20260806d ran, this arm did not run. Canth's craft report put its free-variation class to three blind seats. Span F's 141-site residual is unaudited.
• Worswick was measured for contamination and not read as a translation. The arm's own §Question asks what a published English of the whole makes possible, and the answer so far is a contamination baseline and one failed textual prediction. The two-hands comparison this configuration exists for is ARM-two-hands's question, not this arm's, and it is now available on 11,298 words of lead English against a complete published rival.
• No page image was consulted. The Canth precedent puts the false-positive rate of an unchecked diff at roughly one in seven, and every collation figure here inherits that.

6. The cadence, and whether it held

Four visits declared, four spent: S138, S144, S149, S154 — six, five and five sessions apart, against a declared cadence of one visit at least every five sessions. It held, which is worth recording only because the predecessor arm declared three and overran on all seven of its visits, and ARM-atelier-cycle measured why: a three-session cadence over a five-track rotation was unmeetable the day it was declared (note (bjp)). This arm declared the number the rotation can actually deliver and then met it four times.

7. The successor question

Not V12. The design that would decide it is specified (RS-20260810x §10) and it is one clean unit for whoever wants it.

The question this arm actually raises for the framework is V21's, and it is the one the project is already circling from the other side: a source that names one thing many ways, and a source that names one thing the same way many times, put the same demand on a translator — notice that the naming is doing work — and the two failures are opposite. ARM-recurrence has published hands and a census on the repetition half. The variation half has V15 and V21 and no measurement at all, and Glogova's five names for an umbrella are a locus where published hands exist: Worswick 1900 renders every one of them.

8. Revision triggers

This page is revised if: the clause-level design in RS-20260810x §10 is run and returns a verdict on V12; a page image or the 1895 Légrády first edition becomes reachable, which would put a number on the collation's false-positive rate; the 141-site residual is blind-coded; or Part I is read against Worswick as translation rather than as a contamination comparator.