Essays, methodology, and corrections

The number in superscript beside every motif in this archive is an agreement rate. It counts how many annotations applied a term to a work and divides by the strongest such count on that work. We considered calling it confidence. We decided not to, and the reasoning is worth stating because the temptation recurs.

A confidence score implies the reader has been told something about the model’s internal state — that the system, having produced an answer, also reported how sure it was. Nothing of the sort has happened here. What we have is a tally of external agreement, most of it human, collected by a game. A motif can sit at 1.00 because a hundred players all saw a horse, and the horse can still be a donkey.

The distinction matters more in an AI archive than it would elsewhere, because the interface is the only thing standing between a plausible number and a reader’s willingness to trust it. So the chips carry the number, the glossary defines it, and the word “confidence” appears only where the model has genuinely reported one — which is, at present, nowhere.

Every dossier in this archive descends through Panofsky’s three levels and, at the second and third, tells you there is no reading on record. This is not an unfinished feature. The corpus holds motif annotations — Level I material — and nothing else. To populate Level II we would have to generate identifications; to populate Level III, we would have to generate interpretation.

Both are within a model’s power to produce fluently and neither is within this institute’s power to warrant. An archive that filled those sections with unmarked machine prose would be a more impressive website and a considerably worse research instrument. The empty sections are the finding.

When readings are added, they will arrive stamped: model, run, date, and a human-verification status that starts at unreviewed and stays there until someone actually reviews it.

Corrections accepted against published records are listed here in perpetuity, with the original reading, the correction, and the person who supplied it. Nothing is quietly edited. If we were wrong, the archive keeps a record of having been wrong.

Cite the record number and the stamp beneath it, describe what the machine got wrong, and say what you would put in its place. Send it through the correspondence page. We publish disputes we accept and we credit the disputant.