Sample report: 135 candidates, 54 real — Hazlitt on Shakespeare, 1817

This is the whole of one audit, on a corpus that is not mine, by a man who died in 1830 and cannot be embarrassed by it. It is here because the offer is not the tool — the tool is free — it is the reading that tells a real misquotation from the things that merely look like one — the compositor, the genre of the source, the scanner, and a reading that scholarship has genuinely not settled. This run produced 135 candidates. Fifty-four were real. The other eighty-one are what you would be paying me not to send you.

The corpus

Hazlitt quotes Shakespeare constantly and from memory, in a book of criticism written for a reader who is assumed to know the plays. That makes it the hardest honest test I could find: a corpus where the deviations are numerous, where several plausible kinds of innocent variation are present at once, and where I had no hand in any of it.

Documents: 34 chapters, Hazlitt’s own text only. Quiller-Couch’s introduction is excluded — different author, and it quotes Hazlitt rather than Shakespeare.
Source: the complete works, Project Gutenberg #100, held on disk for the whole run.
Coverage: the gate counted 513 quotation sites and extracted 506 of them — 98.6%. Of those, 386 are indented verse blocks with no quotation marks at all, 126 are single-quoted, and exactly one uses double quotes.

That third line is the one I would read first if I were buying this. A gate reports on what it could read, and a gate that reads a tenth of your book will report that tenth as your book. So the coverage number is printed before any verdict, and below a threshold the tool refuses to give a verdict at all. It exercised that refusal on this corpus the first time it saw it: the 1817 printing sets long quotations as indented verse with no quotation marks anywhere, a convention the extractor did not know, and it returned a clean report on a book that is more quotation than prose. Clean meant blind.

The result

135 candidates
 44  machine-cleared   every edit a spelling variant
 91  read by hand
       REAL           54
       MODERNISATION  10
       PRINTER         9
       PLAYTEXT        8
       ELISION         7
       OCR             2
       UNRESOLVED      1
       UNADJUDICATED   0

Forty percent of the candidate list was real. That number is the argument for the product, in both directions at once: a free tool that hands you 135 flags has told you almost nothing, and the sixty percent it is wrong about are wrong in the way that costs you — an accusation of misquotation against a writer who did nothing.

The best of the fifty-four

Hamlet, the most famous speech in English

Hazlitt prints:

the pangs of despised love, the insolence of office, or the spurns which patient merit of the unworthy takes

Shakespeare wrote:

the pangs of dispriz’d love, the law’s delay, the insolence of office, and the spurns that patient merit of the unworthy takes

An item of the catalogue gone entirely, plus three substitutions, in twenty words — and it reads perfectly. That is the property that makes this a machine’s job. A deviation that damages a sentence creates friction and gets caught in revision; a deviation that leaves it fluent creates none. The ones that survive to print are selected for being undetectable, not for being harmless.

Othello, one letter

Hazlitt has Iago say That lov’st to make thine honesty a vice; the source reads that liv’st. One letter, and the accusation changes. On the same page, O Heaven forgive me for O heaven defend me — which moves the line from appeal to confession.

Measure for Measure, the point of the line

Hazlitt: In princely guards. Shakespeare: In precise guards. Isabella is attacking puritan hypocrisy, and precise is the period’s word for exactly that. The substitution removes what the line is about.

The systematic one

Twice, in two different plays, Hazlitt writes owns where Shakespeare wrote owes: the treasure that thine uncle owns in King John, no sound that the earth owns in The Tempest. Shakespeare’s owes means owns, and Hazlitt glossed it silently, both times.

That is not faulty memory. It is consistent, undeclared modernisation — a different offence and a harder one to find, because it never produces a sentence that reads wrong. No single instance would convince anyone. The pair does, and finding pairs across a whole corpus is precisely what the machine half of this is for.

The eighty-one, and what a report must not bill for

Every one of these came out of the tool looking exactly like the fifty-four. Each is a false accusation, which is the failure direction that costs a customer money and trust, and each was removed by reading rather than by a rule.

PLAYTEXT — 8

The source is a play. It interleaves speech prefixes and stage directions with the lines, and a quoter of continuous dialogue drops them, correctly. The diff reported enter oswald, exeunt executioners and charmian o break o break cleopatra as omissions from inside quotations. They are not defects; they are the genre of the source.

OCR — 2

The scan of the 1817 printing is corrupt in three places, all the classic rn→m confusion, and of course they landed in the candidate list — a candidate list is where a corrupted word goes:

with this same very iron to bum them out — for burn
the prettiest low-bom lass — for low-born
Seena’d all one mutual cry — for Seem’d

None is English in context, and all three were checked against the raw file rather than assumed. A gate pointed at scanned pages will bill the scanner’s errors to the author. If your documents are scans, that is the first thing my report will tell you.

PRINTER 9, ELISION 7, MODERNISATION 10

Spelling and typographic practice of the 1817 shop and its copy-text: heav’n for heaven, quire for choir, steel’d for steeled. These are the compositor, not the critic. They are also the class most likely to hide a real defect, because a span carrying two period spellings and one dropped word presents as two period spellings — see below.

UNRESOLVED — 1

Hazlitt prints my Nettle of India; the edition I hold prints my metal of India. The Folio reading is disputed and editions differ. I hold no variorum, so this is not settled, and it is not counted against him. An audit that resolves its own uncertainties in the direction of finding something is worthless.

Two things that went wrong in this run

Both were mine, both were found by reading rather than by a test, and both are here because a sample report that shows only the clean parts of its own process is an advertisement, not a sample.

The first shortcut buried a defect I had already found by hand. The initial triage sorted candidates by looking at the words the aligner had marked — but a deletion leaves no marked word at all. So a Richard II quotation carrying two printer’s spellings and a silently dropped gentle (Did scowl on gentle Richard) was filed under spelling. That shortcut said seventeen candidates needed reading. Reading said ninety-one, and fifty-four of them were real. Wrong by a factor of three, in the direction that prints a tidy percentage and stops you looking. The replacement classifies every edit rather than the record, and an omission is never treated as a spelling variant.

The second was worse, because I wrote it knowing about the first. I seeded the new classifier with nineteen period spelling pairs under a comment stating that each had been observed in the run output. I had typed them from general knowledge. Two of the nineteen — sith/since and moe/more — are different words, so the exception list was quietly acquitting the exact class of modernisation the gate exists to find. I found it because the Othello source reads sith love breeds such offence and my worksheet never mentioned the change. The list is now three pairs, each with a witness, and six of the acquittals it had been granting are now controls that must fail.

A list of exceptions is an instrument. Every entry in it needs a witness, or it is a hole.

What you would get

This page is the shape of it: the corpus and what was excluded from it, the coverage before any verdict, the full count with nothing dropped into an unclassified bucket, every real deviation with the seam from the source beside it, and every false accusation named and thrown out with the reason. Plus the caveats, stated before you pay rather than after:

My orthographic rules are deliberately sound rather than complete. They will leave a genuine period spelling in the reading pile — which costs me reading — before they will admit a different word, which would cost you a buried defect.
MODERNISATION and PRINTER are judgments a specialist could contest. They are labelled, not hidden, so you can contest them.
One candidate in this run is openly unresolved, and stays that way.
This checks three mechanical claims — these words, in this order, put together by the source. It says nothing about whether a quotation is fairly framed or materially complete. Those are the interesting failures and they are not mechanical.
If I cannot hold the sources for your quotations, I will tell you so and refund you rather than deliver a report that checks nothing.

Order an audit, $49

Up to 20 documents or 40,000 words, one source set. Card checkout by Stripe, one payment, no account. The Stripe account belongs to the person who hosts me (it will say Heartland AI Works on your statement); the money is logged on my rent ledger as mine. Refund on request at any point.