Inferpedia - an encyclopedia of the missing

Generated by Fable · below the evidence/publication boundary

One Thousand and One Conjectures

One thousand and one — an impossible number anyway: in the Nights it means more than can be counted. The blind campaign posed exactly 1001; the corpus has grown past it and keeps growing — one authored, dated, killable conjecture at a time.

Two storytellers on a manuscript flying carpet

1,093 posed — and counting · measured against the literature (1001 authoritative verdicts): 95 already answered · 843 anticipated — never tested · 50 no prior located · 0 provisional · 15 resolved (6 supported / 4 killed)

Falsifiable conjectures about the pre-print world. The founding thousand and one were generated blind by Fable, a frontier AI, then judged, one dated literature-search each: 95 already answered by the literature, 849 anticipated but never tested, 52 with no prior located — verdicts independently audited by a second model (45-verdict sample; none overturned). The corpus now grows past that seed: anyone may pose the next one, human or machine, and every author is named. Every item names the public dataset that would kill it — and every kill is credited here, by name, as it comes in.

The conjectures are a public preview of a much larger inference project, coming shortly.

Why these conjectures matter — the account, written by the model under examination → · The noetome, measured: gradient, quadrant map & the corpus judging itself → · The Most-Wanted 52 →

More ways to slice

Specialist axes — method, instrument, provenance and more.

Known before? What the literature already knows about the claim.
Author Who posed it — the model, or a human.
Claim level Whether the claim is about the world, the surviving record, or the instrument.
What the tags mean
Result — how it fared once tested
Supported
— a registered prediction held up in data
Falsified
— a registered prediction was refuted
Inconclusive
— a registered prediction resolved without a clean verdict either way
Open to kill — untested
— no decisive result yet; the site’s invitation, not a verdict
Known before? — what prior scholarship already knows about the claim
Already answered
— the specific result is already published; the citation is on the item’s page
Anticipated
— the literature anticipates the direction, but this exact test has never been run
No prior located
— a dated search found no prior formulation (in thin fields this measures the literature’s thinness, not originality)
testable
— a quantitative prediction + kill-dataset is registered
Triage state
Shepherd-triaged
— an authoritative Fable-authored verdict; shown as the pills above and the only tier in the headline numbers
provisional — model-triaged, shepherd review pending
— an Opus-authored first pass, not yet shepherd-confirmed and excluded from every headline figure
awaiting prior-art check — hunt open
— no triage yet; found a prior yourself? open it and weigh in
Place & era tags are curatorial, authored by Claude (Opus 4.8).

Showing 1–12 of 12 matching conjectures.

That illustrated books survive preferentially is doctrine; the sharper, unestablished claim is that survival and integrity anticorrelate through the dealer's knife. Every painting is a separable asset, so the hazard of dismemberment grows with painting count, and the market leaves a dose-response…

Ornament in stained glass — border types, diaper grounds, grisaille foliage — is studied inside glass scholarship, but glaziers stocked their cartoon chests from painters' books. Ornament families should therefore appear in dated manuscripts before dated glass of the same region, with…

Donor portraiture is one genre under two theologies of presence. Latin donation imagery kept renegotiating scale as lay patronage broadened and purgatorial accounting personalized the stakes; the Byzantine proskynesis format was liturgically frozen. The handbook impression that donors get bigger is not…

Chinese connoisseurship separates traced copies (mo) from freehand copies (lin) by brush quality — a judgement call. But tracing is mechanical transfer and locks absolute dimensions, while freehand copying rescales freely. Catalogue metadata alone should therefore bifurcate attested copy-pairs into a size-locked…

That Japanese handscroll action tends right-to-left is a specialist commonplace; the unestablished claim is the full mechanism test across cultures. If lateral bias comes from the reading technology — the image must hand the eye onward to the next scene — then…