Work on this with me.

For classicists

How does machine commentary fail, and in what patterns?

The argument about automated scholarship is being conducted almost entirely without evidence, in both directions. What does not exist is a close description of how this kind of commentary actually fails when somebody qualified reads it: how often, on what sort of material, and in what recognisable patterns. That is a measurement problem with no instrument, and building the instrument is the harder half. It needs a scheme that separates an invented fact from a defensible reading pushed past its evidence, from an anachronism, from a claim that is perfectly true and useless — and, hardest of all, it has to handle the notes' claims about intent, tone and humour, where the notes assert the most and "wrong" is an argument rather than an error. Once the categories exist, the rates are arithmetic.

Intertextuality is the natural stress test, because an allusion is exactly the claim a fluent system can manufacture and a reader without the languages cannot check. The rule here is therefore absolute: the model may never assert that a passage quotes or alludes to anything, only explain a connection handed to it from a list a person compiled — which is why the site shows 1 quotations rather than several hundred. Lift that rule over passages whose intertexts are already well established in the scholarship, and you learn what share of genuine allusions the model finds and how much invention arrives with them — a measure of what the caution costs, and a finding that would inform every project of this kind.

The corpus makes the sampling tractable: 11,399 notes over 30 works spanning roughly a thousand years, every note held separately by kind, so author, genre and period can be varied deliberately rather than taken as they come. The result would be the first error taxonomy for machine commentary on classical texts, with measured rates attached.

For education researchers

Is the apparatus a scaffold, or a crutch?

The cognitive load literature says scaffolding helps novices. The same literature says the same support can hurt a reader one step further on — the expertise reversal effect — and that friction removed is sometimes learning removed: the looking-up I so resented doing may have been the desirable difficulty. Whether annotation in place is a scaffold or a crutch is exactly the kind of question that literature exists to settle, and as far as I can establish, nobody has asked it of readers meeting ancient texts in translation.

This is what the dial is for. The same passage can be served bare, lightly glossed, or fully annotated, and the level can be assigned rather than chosen. The outcomes worth measuring are the obvious three: comprehension now, retention at a delay, and the transfer test that actually matters — hand the reader a new passage with no notes at all and see what the apparatus taught them. If readers who had the full apparatus do worse on their own afterwards, I want to know that more than I want anything else on this page.

Building the conditions is straightforward and it is my job, not yours — along with consent screens, random assignment, and whatever else your protocol requires.

For psychologists

Does a supplied reading foreclose the reader's own?

Many of the notes take positions — on tone, register, intent, where the joke is — and the site flags them as arguments rather than facts. Flagging may not help. The question is whether a reader who meets the note before the passage ever really reads the passage, or whether the note's interpretation simply becomes theirs: anchoring, applied to literary judgement. The design is almost embarrassingly clean — the same interpretive note served before the passage, after the reader has committed to a reading of their own, or not at all, with agreement as the measure.

Its twin is calibration. Annotation in place may improve understanding, or it may mainly improve the feeling of understanding — readers whose confidence has grown faster than their comprehension. Metacomprehension is a measured thing, the confidence–accuracy gap is a standard instrument, and if this apparatus mostly manufactures conviction, that is a more consequential result than the one I am hoping for. It bears directly on how anything of this kind should be built from here, and it would still be worth publishing.

For translation studies

Can the machine catch a Victorian translator tidying?

Copyright drives much of this corpus into pre-1929 translations, which elevate register, moralise, and flatten obscenity and humour with some consistency. The notes make claims about precisely this, constantly — that the translator has softened a word, expanded a silence, or quietly declined to render something — and the Greek or Latin sits aligned beside every one of those claims, passage by passage.

It is the most tractable question on this page, because the claims have determinable answers: take the notes that accuse a translator of tidying, and check them against the original. It also has an audience well beyond this project. Reading Victorian translations is what most non-specialists actually do, and whether an automatic gloss can reliably flag where one has been tidied is worth establishing whatever one concludes about the rest of the method.

For HCI and science communication

What does the label "machine-written, unreviewed" actually do?

Every page on this site tells the reader, plainly, that the notes were written by a machine and reviewed by nobody. The assumption behind that honesty — mine, and increasingly everyone's — is that disclosure produces calibrated trust: told what the notes are, readers will lean on them accordingly. As far as I can find, nobody has tested whether readers of a working site see the label, believe it, or change a single behaviour because of it — how much they check, what they are willing to cite, how far they let a note settle a question.

The label is mine to vary: prominent or buried, repeated beside every note or stated once in a footer, reworded to any strength a design calls for, within a study readers have consented to join. Every archive, encyclopaedia and news organisation now attaching provenance labels to machine-written text is guessing at this. A measured answer would travel far beyond classics.