Methodology

Edgar Gibson’s 1894 Cassian has four holes in it. They are filled here, and this page is how you decide whether to believe the filling. It is not a summary of the work; it is the working record, with the numbers, including the ones that count against us.

The translation is machine-made, with human review at every stage. That is said first because everything else on this page follows from it. A language model translating a fourth-century Latin text it has certainly seen translated before has one characteristic failure, and it is not the one people expect: not garbled Latin, but fluent English that was recalled rather than translated — a verse arriving in the words of the Authorized Version before the Latin has been parsed, a clause arriving in a modern translator’s shape. It reads well. It is often nearly right. And it is not a translation of the sentence in front of it. Every part of the method below exists to make that failure visible, and the last section of this page sets out, without softening, where it still cannot be seen.

Almost every figure on this page is computed when the page is built, out of the file that produced it — the Latin from latin.json, the English and its notes from content.json, the vocabulary from glossary.json and the mining run behind it from glossary_candidates.json, the departures from departures.json, the calibration and pass 5’s measured exposure from calibration.json — so that a number here cannot drift from the thing it describes. What is not computed is listed here, in full, rather than gestured at, because a sentence whose whole job is to tell a reader which figures to distrust is worth nothing if it leaves any of them outside its own classes. There are four sorts.

• The pages of Gibson’s own volume — the four printed folios his omission notes stand on. The transcription this edition parses carries no pagination, so there is nothing in the data to derive them from; each was read off the scan, and the table below says so.
• Five figures that exist only in the working logs, each named on this page beside the log it lives in: fornicatio’s count correlation before the mining was decontaminated (task-9-report.md); the withdrawn word count 4,297 (translation-log.md); the ten psalm references pass 5 misaligned by one, and the fourteen notes Conference XXII’s readers found false (translation-log.md); and how the calibration’s verdicts against Gibson divide between those refuted by his own English and those settled by morphology (calibration.json’s per-entry reasoning, which is prose and not a count). None of these has a field any build can read.
• The readings, page images and Latin quoted in the narrative — extingunt, praegrauata, the 600 ppi renderings, the pixel measurements of Petschenig’s letter-spacing, the section-convention check. These are what was done, written down; they are evidence, not arithmetic.
• What this page says about its own gates — that twelve checks run, that three of them compare the English against its own Latin, what each cannot see. Those are claims about code rather than about data, and check_methodology_page() holds them: change the checks without changing this page and the build stops.

Citations follow one convention throughout: work, then book in Roman numerals, then chapter in Roman, then Petschenig’s section in arabic. Institutes VI. VIII. 1 is book VI, chapter VIII, section 1; Conference XII. V. 2 is Conference XII, chapter V, section 2. The Roman forms are Petschenig’s own, so a chapter is IIII and VIIII rather than IV and IX, exactly as the reading pages head them. Every locus named below was checked against latin.json.

Gibson did not explain himself. In place of one book of the Institutes and two whole conferences his volume prints a single sentence each; over the first half of one chapter of Conference II it prints Cassian’s Latin, untranslated, with a footnote under it. Those four sentences are the entire record of the decision, and they are reproduced here word for word, with the page each stands on.

All four are about sex. That is not an inference from their subject matter; it is the whole of what they have in common, and nothing else in the 1894 volume was touched. Cassian is not discussing marriage or morals in them. He is treating the body’s involuntary motions as diagnostic — evidence about the state of a man’s heart — and he is exact about the detail because in his scheme the diagnosis depends on it. The exactness is what the 1894 volume leaves out. Why it does, Gibson does not say: his four sentences give no reason, and this page does not supply one for him.

The cost of the omission is not only that four texts are missing. Institutes VI and Conference XII carry Cassian’s clearest statement that chastity is given and not achieved — the doctrine of grace whose fullest statement is Conference XIII, the conference Prosper of Aquitaine answered in his Contra collatorem. Prosper’s charge ran the other way from the sentence above: what he attacked was not that Cassian made grace the giver, but that he left the will an initiative of its own, which is the reading for which the label semi-Pelagian was coined centuries later. Either way a reader of the English Cassian has had to argue about his theology of grace with two of its primary statements removed from the file.

The base text is Michael Petschenig’s, in the Vienna Corpus Scriptorum Ecclesiasticorum Latinorum: De institutis coenobiorum (CSEL XVII, 1888) for Institutes VI, printed pages 113–127; Conlationes XXIIII (CSEL XIII, 1886) for the conferences — pages 333–360 for Conference XII, 614–636 for Conference XXII, and page 62 for the chapter of Conference II. Those ranges are read off the established text itself, not typed beside it: a page range in prose is a claim nothing checks, and one of them was wrong for a week for exactly that reason.

Petschenig for three reasons. He is the only editor who collated the manuscripts of both works and printed an apparatus that lets a reader see what he rejected — which is what makes it possible for this edition to say, below, that six of its readings rest on a single manuscript and two are the editor’s own conjecture. He is the text every modern study of Cassian cites by page and section. And he is public domain: published 1886 and 1888, and Petschenig died in 1923.

There is no later critical text. The 2004 CSEL volumes, which carry Gottfried Kreuz’s name beside Petschenig’s on the title page and are catalogued as an editio altera supplementis aucta, are a Neudruck: the publisher, the Austrian Academy, states of both of them that the reprint “brings Petschenig’s prolegomena, text and apparatus in unaltered form” — Prolegomena, Text und Apparat Petschenigs in unveränderter Form. What Kreuz made new are the indices, revised and redistributed across the two volumes. So there is no second text to prefer, nothing for the two to differ about, and no copyright question about the Latin at all: the Latin printed in the 2004 volumes is the public-domain Latin of 1886 and 1888, which is what was read here off the scans of the original printings. The only thing in those volumes that is in copyright is the apparatus of indices, and this edition uses none of it.

An earlier version of this paragraph said the opposite — that Kreuz is “the better text on any scholarly account”, that it was left aside for copyright, and it reasoned about “where the two would differ”. Every clause of that was false, and it was the sentence in which this page explained its own base text. It is corrected here rather than quietly replaced, because a reader who read it once should be able to find out that he was misinformed.

Migne’s Patrologia Latina 49 — Alard Gazet’s text of 1846 — is kept alongside as a second witness. It is older and less careful, and it was used the way a second witness is used: consulted where a reading looked strained, never preferred silently, and never allowed to overrule the page in front of us. 61 places across the four texts were logged as genuine divergences between the two, each with the reading chosen and why. A fifth witness turned up unexpectedly: the Latin Gibson himself printed on page 317, which differs from Petschenig in four substantive places — changes of grammar or of sentence division, counted out of latin.json, and not counting the nineteenth-century spellings it also normalises.

Not one word of the Latin was taken from a transcription somebody else typed. Every page was rendered from the scan at its native 600 ppi and read in horizontal bands. That is a method chosen against a failure, not for thoroughness: at the resolution a document reader renders a page by default, u and n, c and e, and m against ni are not reliably separable in Petschenig’s face, and the CSEL XVII scan is glyph-coded — a compressed ink layer painted through a stencil built from a symbol dictionary — so the letter on the screen may be a letter the compressor chose. What is excluded from the transcription is as deliberate: the apparatus criticus, the scripture-reference band, the marginal line numbers and the marginal section numbers. Petschenig prints his scripture quotations letter-spaced, and the spacing was measured in pixels rather than judged by eye — ordinary type runs 1 to 7 pixels between letters and his spaced type 10 to 20 — after eyeballing it failed three times at three-times magnification and once put Cassian’s own antithesis inside a psalm.

The transcription was then diffed, word by word, against a machine OCR of the same images. The OCR is a net, never a source. Where the two disagreed, the disagreement was settled by looking at the image again; the image always won.

What the collation earned, plainly. 1,449 disagreements sounds like 1,449 catches. It is not. Across all four texts the diff corrected the transcription once. On printed page 118 of CSEL XVII the eye had read extinguunt; the OCR read extiugunt; the page, re-read at 600 ppi, plainly prints extingunt, with one u — and Petschenig’s own apparatus at that line records extinguunt as the reading he rejected. The transcriber had silently normalised the page to the commoner form. That is the class of error that survives a careful review, because it produces a real word in the right place, and the only thing that caught it was a machine that was, in that one spot, right.

Everywhere else the diff confirmed and did not correct. Conference XII returned 695 disagreements and corrected nothing: nine readings were argued and all nine resolved in favour of what had been read off the image. That is a weaker result than Institutes VI’s, not a stronger one. The eight calibration chapters of Conference XI returned 202 disagreements with nothing adjudicated at all — the net caught nothing. The honest statement is that the warrant for this Latin rests on the pages having been read at 600 ppi with Petschenig’s own line numbers as the completeness check, and the diff is a second opinion that has earned its keep exactly once.

Once, the machine was right and the reader was wrong in the other direction. The first word of CSEL XIII printed page 62 was transcribed praegranata; the OCR read praegrauata, and a glyph-level dump against a confirmed u and a confirmed n on the same printed line showed the three profiles are not separable in that face — the raster genuinely cannot decide it. It was settled on five grounds, none of them the image: there is no Latin verb praegranare; praegrauo is Cassian’s own word elsewhere in both volumes; Migne reads praegravata at exactly that place; Petschenig’s apparatus reports nothing there, so he saw nothing worth reporting; and the sense requires it — a mind weighed down is what is forced to fall asleep. The word belongs to the chapter before the one this key establishes, so no published Latin changed. It is recorded because it is the counter-example to the rule that the image always wins, and because it is why two further filters exist.

Petschenig’s section convention. The paragraphs of this edition are his marginal sections, and his numbering has a trap in it: it restarts at 1 in every chapter, and section 1 is never printed — the chapter numeral stands in its place, so the first digit a reader sees in the margin is a 2. That was not assumed. The section numbers in this edition’s data were written by a counter, so the coincidence had to be checked: seven page images across both volumes and all five established keys were rendered and the margins read directly, twenty section boundaries in all, with no divergence — and at Conference II. XXIII the marginal 2 sits exactly where Gibson’s English resumes, which confirms that the paragraph he printed in Latin is Petschenig’s unprinted section 1.

Cassian returns to a small technical vocabulary across both books, and a word he rendered one way for thirty years of pages ought to be rendered the same way in the four he skipped. So the vocabulary was not decided from scratch: it was mined from Gibson’s own English, in the thirty-three texts his volume does put into English.

The mining works because Petschenig prints a capitula list at the head of every book and conference and Gibson translates those same capitula as his chapter titles. Pair chapter n of one with chapter n of the other and you have two kinds of evidence: a title pair, in which one Latin clause faces its English, and a body pair, in which a chapter of Latin faces a chapter of English. Of the 759 chapters that can be aligned at all, 725 aligned on the title — 95.5 per cent — and 651 on title and body together, of which 635 survive a decontamination step described below. Title evidence decides wherever it exists, because a capitulum is one clause and cannot swallow its neighbour; body evidence is circumstantial and is reported as circumstantial, with the count correlation printed beside it. All four denominators are counted out of glossary_candidates.json, the mining run’s own artefact, when this page is built.

The decontamination matters more than it sounds. Where the OCR destroyed a chapter’s closing numeral, that chapter’s words fall into the previous chapter’s window, which goes on being scored against its own small piece of English. Conference V collapsed that way: twenty-seven chapters into three windows, one of them pairing 3,852 Latin words against 49 English. Sixteen such windows were dropped. No decision moved; the correlations sharpened — fornicatio stands at 0.92 in the run as it survives. Its value before the drop, 0.41, is the one figure in this section with no file behind it: it is recorded in task-9-report.md, the report of the round that did the dropping, and is quoted from there.

One rendering is a deliberate choice against Gibson. The table shows integritas carnis rendered “integrity” against a count of nought: the word does not occur in his volume. He had no fixed rendering for the noun — his 22 occurrences of it carry 23 renderings, twelve “perfection”, six “uprightness”, five something else — and his commonest one collides. At Conference XII. II. 5 the Latin reads ad perfectionem puritatis … nisi etiam mentis addatur integritas, and “the perfection of purity, unless perfection of mind be added” is not a translation but a collapse. “Purity” and “chastity” are already spoken for by puritas and castitas — Institutes VI. XX’s own capitulum puts two of the three in one clause — and “uprightness” will not govern what Cassian governs with it: integritas corporis nostri, the capitulum of Institutes VI. XXII, is not “the uprightness of our body”. What settles it is that Cassian glosses the term himself, inside one of the texts this rendering governs:

aliud enim est continentem esse, id est ἐγκρατῆ, aliud castum, et ut ita dicam in affectum integritatis uel incorruptionis transire, quod dicitur ἁγνόν Institutes VI. IIII. 1

Which makes integritas the condition Cassian names ἁγνόν — the state above mere continence, and the word the page above prints, his neuter adjective and not the abstract noun a lexicon would give — and not a word about moral rectitude. The entry was first decided as “uprightness” and re-decided when review showed its load-bearing claim false: “uprightness” occurs eleven times in the volume and only six of them render integritas. Its confidence is recorded as low, and the reason is written out: there is no title pair, no precedent for the word itself, and Cassian does not always distinguish integritas from puritas. What is high is the negative finding.

And the expectation the mining overturned. This work began assuming that the vocabulary of the suppressed texts — inlusio for a nocturnal emission, pollutio, gastrimargia — would have no Gibson precedent, because the subject had been cut. All three do. Not one of the 22 terms came back empty: inlusio stands in five of his capitula and he renders it “illusion” twenty-one times; gastrimargia is “gluttony” fifty-three times. Even the physical sense of inlusio, which no capitulum carries, is printed three times in his own body text — quadam somni inlusione contaminet, “should by some illusion in a dream pollute”. What the Victorians suppressed was the four texts, not the vocabulary.

Departures are declared at the place they happen. Where a glossary term stands in the Latin, the fixed English is owed; across the four texts that is 220 obligations, of which 191 are discharged and 29 are declared as departures, each keyed to the exact text, chapter, section and term, each carrying its reason. By term: uirtus 17, continentia 8, affectus 3, integritas carnis 1. A declaration that excuses nothing is itself a failure, so blanket permissions cannot accumulate — and the declaration is checked against the page it describes, because one of them once quoted a phrase the page had stopped printing.

Before a word of the four texts was written, the method was run against a conference Gibson did translate, blind. Conference XI, chapters I to VIII — 23 of Petschenig’s sections, 1,749 Latin words — was established from the page images and translated without opening his English at any point. The Latin artefacts were scrubbed first: nothing in the transcription, the collation or the variant notes glosses a single word into English, and the agent doing the translating was not told it was a calibration. Only afterwards were the two Englishes laid side by side.

109 places differ. 57 of them are Gibson’s own looseness — he paraphrases, and the difference is not a disagreement about what the Latin says. 37 are stylistic. 2 are excluded, and the exclusion is a disclosure rather than a convenience: this edition’s style sheet quotes two clauses of Gibson’s English from this very conference as evidence for two typographic rules, so at those two places the translator had seen his words. One matches him by direction; at the other the translator, having seen it, deliberately chose otherwise. Twenty-nine words of 2,836, worth 0.1 of a percentage point on the overlap figure below — the two clauses in and out, counted in calibration.json and differenced here.

That leaves 13 semantic divergences — places where a reader of one English and a reader of the other come away believing different things about what Cassian said. Twelve resolve against Gibson. One resolves against us. None is unresolved.

Three of the twelve, to show what kind of thing they are. At Conference XI. VI. 1 faciunt homines a uitiis temperare is “make men refrain from faults”, not “enable men to control their faults”: temperare with a and the ablative is to abstain from a thing, there is no accusative in the clause for “control” to govern, and the conference says the same thing twice more in its own words. At Conference XI. VII. 3 a relative clause hangs on gradum in his English and on filiorum in ours, and the verb settles it without any appeal to sense: credunt is plural and gradum is singular. At Conference XI. VII. 6 he renders nostrae salutis gratia as “for the grace of nothing but our salvation”, taking a causal idiom for the noun — and his own next clause refutes him, because the answering limb has obtentu doing the identical job and he renders that “for the sake of”. Four of the twelve are refuted by Gibson’s own English elsewhere on the same page; seven are settled by morphology or by Cassian’s usage. Those two are counts of a kind of reasoning rather than of a field: calibration.json records how each entry was settled in the entry’s own prose, so they are read off semantic_divergences and not computed from it — and they do not partition the twelve, since more than one argument settles some of them.

The twelve were then handed to an adversarial reviewer briefed to take Gibson’s side, which re-parsed each from the full period rather than from the clause the diff quoted, checked every one against Migne, and verified the established Latin against the OCR at all twelve loci so that no textual escape hatch was left unexamined. Twelve upheld, none overturned. It also found one more, which is what makes the semantic count thirteen and not twelve, and it is against us: at Conference XI. VI. 3 receperunt had been rendered “received back”, which asserts a restoration the sentence does not state. (The diff files that entry as VI-15, which is its serial number in the run and not a locus.) That entry is worth more than the twelve, because it names the error this method will actually produce at scale — not misparsing, which the wooden pass catches, but silently deciding what the Latin leaves open. A sixth pass exists because of it.

The baseline that makes the overlap figure mean anything. Our English and Gibson’s share 26.7 per cent of our words inside runs of five or more words identical to his. On its own that number says nothing at all: two people translating the same sentences out of the same Latin into the same register will agree. So it was measured again against text he wrote but we did not translate — all 189,202 words of his other conferences — where the figure is 5.0 per cent, and against single unaligned conferences one at a time, where it is between 0.2 and 1.6 per cent. About five points of the agreement is register. The remaining twenty-two is translating the same sentences. Had the English been recalled from his, the overlap with his other conferences would have risen toward the same level. It does not.

Which measurement, and by whose tokenizer. The three figures in the sentence above come from two different runs, and the page says which is which rather than printing them as one. 26.7 per cent (against his Conference XI) and 5.0 per cent (against all 189,202 words of his other conferences) are the adversarial review’s own. The 0.2–1.6 range is this desk’s reproduction, over four named conferences and a 195,206-word comparand; the review’s own single-conference range was 0.4–0.9. The reproduction counts chapter titles as well as body text, which is why its denominator is 2,924 words against the 2,836 the contamination arithmetic uses, and why calibration.json ends up carrying three counts of one quantity — our overlap with the Conference XI he did translate, at 26.3, 26.7, 27.3 per cent — differing by tokenizer scope and by nothing else. The load-bearing figure, 5.0 per cent, reproduces exactly. This is printed rather than tidied away for the reason the word counts below are printed twice: a figure quoted without its method is a figure nobody can reproduce.

Every section — 159 of them across 56 chapters — was taken through six passes in order, and the log records what each one caught. The first four are the loop the calibration was translated under. The last two were added after the calibration was diffed against the 1894 volume, because between them they name the two failures this method actually produces: harmonising a verse toward a familiar English, and silently deciding what the Latin leaves open.

1. Wooden pass. A literal gloss following the Latin word order, every finite verb parsed by person, number, tense, mood and voice, every participle given its agreement, every ablative absolute and every deponent named as one.
2. Draft, under the style sheet and the glossary.
3. Self-check. Back to the Latin clause by clause against the draft, hunting negations, conditionals, comparatives, pronoun antecedents and subject changes.
4. Scripture. Every quotation rendered as Cassian quotes it, from Petschenig’s letter-spaced span itself and not from the surrounding prose, with the reference in a note.
5. Scripture register audit. Mechanical: our rendering and the same verses in the Authorized Version reduced to word lists, inflection folded, aligned, and the Version’s material outside Cassian’s clause trimmed away. Only substitutions are flagged, and a flag is a question, not a verdict — the translator answers it from the Latin and the answer goes in the log. What this pass cannot see is the subject of the next section.
6. No silent additions. Every English word with no counterpart in the Latin is either removable without loss, or logged. Read the finished English against the Latin word by word and name, for each English word, the Latin word it renders. What is left over is scaffolding (free), removable (take it out), or an assertion — and an assertion belongs in a note or nowhere. An unbracketed clarification is a bracketed one with the brackets taken off, and harder to see.

Then, outside the numbering and after all six, a fresh-eyes review of every chapter: one reader per chapter, given nothing but that chapter’s Latin, its letter-spaced spans, our English and our notes — no glossary, no style sheet, no other chapter, no published English of Cassian — and asked for dropped clauses, smoothed-over difficulties, confident nonsense, inverted negations and wrong antecedents, quoting the Latin for each, and told not to comment on style. Fifty-six chapters, fifty-six readers. It is in the loop and not beside it, because it is the only check that catches the complete form of the failure this whole method is built against.

Four of its catches changed what a text claims. Each is cited by book, chapter and section, and each locus was checked against latin.json before this page was written — the first three are all in Institutes VI, at its chapters VI, VIII and VIIII.

Institutes VI. VI. 1 — a concession the Latin refuses. nulla etenim uirtute tam proprie carnales homines spiritalibus angelis … aequantur had been rendered “so nearly”: the draft had read propius, the comparative of prope, where the page has proprie, “in the strict and proper sense”. The sentence claims that by chastity carnal men are made equal to the angels properly so called; the draft conceded a shortfall, in the one chapter arguing that chastity is a gift. It read as perfectly good English, and nothing but the Latin would have caught it.

Institutes VI. VIII. 1 — a conditional turned into an assertion. sine qua et illi … non poterant interesse, et qui … nequeunt coronari is a correlative pair under one sine qua. The draft let it govern the first limb only, leaving the second standing as a flat claim that athletes cannot be crowned — the opposite of the argument, which is that they cannot be crowned without it. The same finding fixed the antecedent: qua is feminine and agrees with integritatis, not with the neuter munus.

Institutes VI. VIIII. 1 — the angels deleted from a claim about them. sanctorum tamen angelorum ipsiusque omnipotentis dei scientiam, quam nulla subterfugiunt secreta: quam is accusative feminine and its antecedent is scientiam, the knowledge of the angels and of God together. The draft’s “from whom no secrets escape” quietly attached the clause to God alone.

Conference XII. V. 2 — a proof-text arguing the opposite of its purpose, and it was Gibson’s fault that it was. Proverbs xvi. 26, uim facit perditioni suae, stood as “forcibly bringeth about his own destruction” — which is Gibson’s own English of the same verse at Conference VII. VI, archaised at the verb (he prints “brings”), taken from him rather than derived. uim facere with a dative is to do violence to a thing, the Greek reads ἐκβιάζεται, forces away from himself, and Cassian uses the idiom that way inside this same conference at Conference XII. VIII. 2, where inferre uim naturae is to force nature. The verse is adduced to show that hard labour profits the man; as printed it argued that his labour destroys him. It now reads “forcibly driveth away his own destruction”, with a note carrying the idiom, the Greek, and the fact that Gibson is departed from. This is the failure the whole method exists to prevent, appearing exactly where the calibration predicted it and being caught.

No in-copyright translation was consulted at any point, including to check a single difficult sentence. Two are named here precisely because they are the ones it would have been easiest to reach for. Boniface Ramsey’s complete Conferences and Institutes (Ancient Christian Writers 57 and 58, 1997 and 2000) is the standard modern English of both books. Terrence Kardong’s Cassian on Chastity (1993) translates exactly the three whole texts restored here — Institutes VI and Conferences XII and XXII — and is therefore the single most tempting shortcut in existence for this particular job. Neither was opened.

The reason is not piety about copyright. It is that the only way this English can be checked is against the Latin, and a translation made with a modern one open beside it is a paraphrase of somebody else’s decisions wearing the appearance of an independent reading — and no reader could ever tell the difference. The only English control used anywhere in this work is Gibson’s own, on the passages he did translate, which is public domain and which the calibration above measures against.

No facing-Latin edition. The Latin is not printed alongside the English, except for the one paragraph Gibson himself printed in Latin, which is kept beneath our rendering of it. The reason is that this edition’s Latin is a transcription made for it, and publishing it beside the English would present a working text as though it were an edition. The base text is named, page by page, so that a reader can open Petschenig.

No smoothing. Cassian is difficult in Latin — long periods, subordination stacked four deep, an argument that turns on a single particle — and he is left difficult in English. Where a sentence is hard because the thought is hard, it stays hard. Where his terms are technical they are held to one English word, even where a looser word would read better, because the argument is built out of the distinctions and a reader who cannot see the distinctions cannot follow it.

No verdict. This desk does not tell you whether Cassian is right that chastity is given rather than won, whether Theonas is right about guilt and consent, or whether Gibson was wrong to leave these texts out. The four texts are here so that the question can be asked with the evidence in front of you.

This is the section the page exists for. Everything above describes instruments; this describes what they do not reach. It is written on the principle that a check whose failure modes are not written down is a check nobody can weigh.

Six readings in Conference XXII rest on a single manuscript. Petschenig prints them against every other witness, and in five of the six against the printed editions as well. They are named here because a reader is entitled to know which sentences would read differently under another editor.

Two readings are not in any manuscript at all. They are Petschenig’s own conjectures, marked scripsi in his apparatus, and both fall in the same section — Conference XII. XI. 4, the passage on what wakes the body in sleep, which is exactly where the conference’s argument is decided.

• iugi stillatione (printed 352) — “by a continual dripping”, where the manuscripts read iugis titillatione, a continual titillation, or instillatione. Whether the night’s stirring is mechanical pressure or arousal is the question the chapter is about.
• tantum for the manuscripts’ tamen (printed 353) — every manuscript and every printed edition reads tamen. “Only through sleep” and “nevertheless through sleep” are different claims about involuntary emission, which is the point at issue in the whole chapter.

The English follows Petschenig’s printed text at both, because that is the base text this edition declares; the notes on those paragraphs name the conjectures as his. The second of them is load-bearing beyond its own sentence: it is the word whose scope the cold check below turned on, and the reason a published sentence was changed after publication.

The word counts do not reproduce, and both are printed. The English of the four texts is 20,536 words counted as whitespace-separated tokens, and 20,441 counted as runs of A–Z letters — which splits self-love in two and drops the ligature out of Chæremon. The gates use the second; this site publishes the first. Neither is wrong, and a figure quoted without its method is a figure nobody can reproduce. One earlier count in this record, 4,297 for Institutes VI, was correct on the day it was taken and went stale through seven later word removals — which is why every count on this page is computed when the page is built rather than remembered. That figure is itself an example of what this page cannot compute: it describes a state of the text that no longer exists, so there is nothing to count it out of. It survives in translation-log.md §4, which withdraws it and shows the walk down, and it is quoted from there.

The semantic divergence is at Conference XII. XVI. 2, and this edition lost it. The sentence is sed cum dormienti tantum per sopitae mentis incuriam conmotio carnis obrepserit, and what is at issue is the scope of tantum — whether it limits when the motion comes or how it arises. The page had printed “in sleep only”, taking tantum with the dative it stands next to; the cold reading took it with the per-phrase that follows. Both parse. What settled it is the editor’s own apparatus: at Conference XII. XI. 4 Petschenig prints tantum against every manuscript and every printed edition, and defends that conjecture with a bare cross-reference to this sentence — and what he printed there is tantum per soporem, so the parallel he is invoking is the collocation itself. The English was changed, and the note on that paragraph now carries the whole argument, including what a reader who wants to disagree would have to weigh: the two passages stand or fall together, since the tantum at XI. 4 is the editor’s word and not any manuscript’s.

What the sample says, and what it does not. One semantic divergence in 705 Latin words is a measurement of the method, not a clean bill of health. Three chapters are three chapters — 5 per cent of the restored text — and taken at face value the figure puts roughly one such place per 700 Latin words: not a mistranslation, but the English deciding something the Latin declines to decide. That is the error class the calibration’s adversarial reviewer was briefed to look for and the fresh-eyes readers kept catching, and no gate on this page can measure its rate. It is also why the warrant offered here is the readers and not the checks.

Two smaller results, which are about this page rather than about the text. At Conference XII. IIII. 4 the cold translator rendered de sua uirtute as “power” — a departure from the glossary’s own default that this edition had already declared at that locus in departures.json, reached independently and with none of this edition’s reasoning in view. And at Institutes VI. XII. 1 it rendered uiderit as “hath looked on a woman” — the Authorized Version’s verb for a Latin word that does not require it. That is a fresh instance of the blind spot at the head of this section, produced in a chapter drawn at random by a reader who had never seen our page: the pull toward a received English is a property of the job, not of one translator.

Gibson’s translation is public domain — published 1894, and he died in 1924. The Latin base texts are public domain: Petschenig’s two CSEL volumes and Migne’s Patrologia Latina are nineteenth-century publications. The 20,536 words of English made for this edition, and the 305 notes under them, are dedicated to the public domain under CC0 — no rights reserved. Copy them, print them, put them in your own edition, correct them, sell them. No attribution is required, though it is welcome.

That is the point of the exercise. There has been a modern English of three of these four texts since Kardong in 1993, and of all four since Ramsey’s Conferences in 1997 — Kardong covers Institutes VI and Conferences XII and XXII, and Ramsey’s Conferences completes the set with the chapter of Conference II; his Institutes followed in 2000. What there has not been is one anybody could copy. The translation is not the contribution. The licence is.

Corrections are wanted. If you read Latin and something here is wrong — a mood, an antecedent, a scripture rendered from memory rather than from the span, a note that decides more than the grammar does — please write, quoting the Latin: xeb@hey.com. Anything corrected will be corrected on the page with the reason recorded, in the same way the errors on this page are recorded. An edition that publishes its own limits and then does not answer them is only advertising.