A problem book for the pre-print world

Each entry in this book is a conjecture — not an essay, not a fact, but a wager in three parts: a join between two things specialists usually know separately; one decidable prediction; and a named, public dataset that could kill it. One item here, “Glossed books breed webs,” wagers that contamination in classical textual traditions is a curriculum signature — that the authors who lived in the medieval schoolroom, their margins thick with glosses the next copyist promoted into the text, should be precisely the authors whose editors despair of an “open recension” — and it can be checked, this afternoon, against the per-author essays in Reynolds’s Texts and Transmission. If it dies there, this page will say so.

The conjectures are questions about the pre-print world — manuscripts and recensions, glosses and workshops, survival and loss, from the Latin schoolroom to Dunhuang — generated by a large language model probed at scale, and posed here for the people best equipped to destroy them. The counter reads 308 of 1001 posed. Take the 1001 as a promise about where this ends, never as a claim of a thousand finished insights: many entries are open problems still awaiting their first prior-art check, and the whole page is built so you can tell, at a glance, which is which.

The rules of the game

Every conjecture on the index carries one of three states, and the labels mean what they say:

• Shepherd-triaged. A dated prior-art search has been run and a verdict recorded — novel (no prior located), adjacent to existing work, or already in the literature — at the corpus’s authoritative tier. Only these items feed the statistics quoted below, and only these may carry a registered prediction or a public novelty label.
• Provisionally triaged. A model-authored verdict, labelled “shepherd review pending.” It tells you what a first search found; it is not yet a claim this page stands behind.
• Frontier. No prior-art check yet. The hunt is open — these are the items where a reader with the right shelf behind them can settle something in an afternoon.

What we ask of you is what an editor asks of a collation: witness against witness. If you know the article, the edition, the dissertation that already made one of these joins, submit the prior art on the item’s page — the verdict changes, and the change is credited to you by name. If a conjecture is malformed or its kill-test wrongly chosen, dispute it; if you can sharpen it, comment; if you want to run a kill-test yourself, claim it; and if you have a better conjecture of your own, propose it. A find that overturns a verdict is not an embarrassment to this project; it is the project. Every prior located is a data point on the chart this page exists to draw. The bar each item is meant to clear is simple: a specialist reads it and feels implicated — I could check that this afternoon, and I want to.

Why one thousand and one

The number is the Nights’ number, chosen with the philology in view. Alf layla wa-layla promises not a tally but an abundance — a thousand nights and then one more, many-and-one. And famously, the promise outran the text: no early manuscript contains anything like a thousand and one nights’ worth of tales. The number generated the text, not the reverse. When European readers began to count the nights, copyists, dragomans and dealers obliged them — orphan tales with no Arabic manuscript authority behind them (Aladdin and Ali Baba entered the tradition from the oral telling of Ḥannā Diyāb), forged “complete” manuscripts, whole recensions swollen to redeem the title’s promise. Filler, fabricated to meet a promised number, wearing the costume of the genuine article.

That is this page’s named failure mode. A corpus that advertises one thousand and one conjectures sits under exactly the pressure those copyists sat under, and an unscrupulous version of this page would meet it exactly as they did. The guard is the ledger: every wave of generation runs from a prompt committed and hashed before the run; every item is imported verbatim, with its provenance recorded and the generator’s own disclaimer — that it cannot verify its novelty — preserved; and no item is presented as novel until a dated search at the authoritative tier says so. The three states above exist so that volume can never outrun labelling.

The corpus even turns the Nights’ own swelling into a test case. One conjecture here, “The rim-accretion rule,” holds that tales padded in to reach the count should ride the frame’s outer seam — spliced in at nesting depth one, told directly by Shahrazad, where insertion is cheap — while the old core carries the deep story-within-story architecture inherited from its Persian ancestor. It is killable by setting Mahdi’s edition of the Galland manuscript against the Bulaq recension as inventoried in Chauvin: our failure mode, posed as a falsifiable claim about the very book that named it.

And the 1001st conjecture will not be generated at all. It will be posed by a person, live, at Ars Inquirendi in Oxford in November 2026. The machine poses a thousand; people pose the one.

The instrument

Everything here begins with the oldest move in inquiry: put a query to a witness, then weigh the answer against the other witnesses. The witness in this case is a large language model, probed blind — no access to this site, the web, or any database; one pre-committed prompt, one reply, imported verbatim. What such a model knows, it knows the way a scholar knows things away from the library: by formation, not consultation. Call the totality it carries its noetome — a compressed, structured, and for the first time directly queryable deposit of the written record it was trained on. Most benchmarks ask whether a model can retrieve a fact or reason toward a known answer; a query for conjectures asks something rarer — whether it can generate a claim at once surprising, specific and falsifiable. Each answer is a cross-section of the noetome: a place where the geometry of the model’s knowledge says these two things connect.

Read down the index and one move recurs: take a formal apparatus from one field and transplant it onto a historical dataset in another. Preferential attachment becomes the “Manuscript Yule process”; a vocabulary-growth law becomes “Heaps’ law in clay”; Benford’s law “audits Knossos”; eigenvector centrality yields “the jurists’ PageRank.” In the humanities waves the same move runs subtler — glossing habits joined to stemmatic topology, translation chronology joined to citation behaviour — but the shape is constant: [something known] × [something else known]. That is a genuine glimpse of the noetome’s architecture. Its knowledge is organised as reusable structures — laws, methods, habits of transmission — that can be indexed and matched across domains; a conjecture is the model reading off that two of them occupy the same shape. This isomorphism-spotting is a real strength: it bridges disciplinary silos that persist only because no single person straddles both sides. It is also a real ceiling — the creativity is recombinant. It surfaces what the known already implies; it seldom forms a concept the noetome did not already contain in pieces.

Because every triaged conjecture carries a dated prior-art check, we get a novelty rate, not an impression. Of the 158 conjectures triaged at the authoritative tier so far, only 22% (35) had no prior formulation we could locate. 23% (37) already existed in the literature outright, and 54% (86) sat adjacent to existing work — 78% in all with a known prior or a close neighbour. This is the noetome’s central limitation made measurable: “novel to the model” is not “novel to the world.” The model calls a connection new precisely when it hasn’t seen the paper that already made it — the ice-core-lead / Roman-silver link (McConnell et al.) and scattered strips as risk diversification (McCloskey) were both proposed here, and both already exist. The reinvention rate measures the noetome’s incompleteness, and its inability to tell its own edges from the world’s.

The remaining 22% is the interesting direction: knowledge that was implicit in the totality of the known — both fields existed, both literatures existed — yet had never been made explicit, because no single mind held both at once. “Moerbeke’s dividend” has the shape of such a conjecture. Everyone knows Aquinas’s career chronology; everyone knows that William of Moerbeke’s literal Greek–Latin Aristotle arrived across that career. Joined, they predict that Aquinas’s Aristotle citations should convert from mediated to direct contact over his working life while his Augustine holds flat — no new fact anywhere in it, an un-drawn implication of what the field already collectively knew, and resolvable against our own judged corpus of Aquinas citation contacts. Whether it earns its place in that 22% is itself still open: its first prior-art check leans adjacent, and the verdict sits provisional, pending the shepherd — the essay’s own showcase held to the scrutiny this page asks of everything else.

The archive of loss, inherited

There is a sharper reason to point this instrument at the pre-print world in particular. The model’s training corpus is downstream of the archive: it read what survived; of what survived, what was edited; of what was edited, what was digitized; of what was digitized, what was discussed in the languages scholarship favours. Every unevenness in that chain — the unedited majority, the catalogued-but-unread, the traditions rich in manuscripts and poor in editions — is inherited by the noetome as an unevenness of knowledge. The model does not hold a map of the pre-print world; it holds a map of the archive’s map, with the losses compounded at each remove.

That makes this corpus a measuring instrument for something its readers already study. Pose the same kind of conjecture about Latin school authors and about Ethiopic hymnody, about the Cairo Geniza and about Dunhuang, and the model should fail differently: reinventing the literature where the literature is dense and digitized, erring confidently where the record it learned from runs thin, falling quiet where the archive itself is silent. Novelty, reinvention and error rates by region and by domain — computed from the authoritative tier only — are a stated research object of this corpus: a differential probe of how the archive’s survival bias reproduces itself inside the machines now trained on it. Where the model looks most original may be exactly where the record is most lost, and telling those two apart is the same problem, at one remove, as estimating the archive of loss itself.

Track record

Conjectures here do not get to rest at “interesting.” The strongest shepherd-tier items carry predictions registered before the data is touched, with explicit kill-conditions, and the resolutions are public whichever way they fall: so far 6 supported and 3 killed against real data. The kills stay on the page — novel and correct are different axes, and the noetome’s guesses have to earn their place. A problem book that recorded only its successes would be the Nights’ copyists all over again.

Where this sits

This page is one exhibit in a larger cartography. It doubles as a small public preview of Inferpedia, an encyclopedia of the missing — a mapping of the known, the inferred and the not-yet-known across the substrates that hold knowledge, of which a model’s weights are now one — and it is the working exhibit behind arsinq.com, where the gathering it serves keeps its own house.

An honest note on the instrument

Read all of the above with its limits in view. This corpus probes a single model — Claude (Fable 5) — at scale: one instrument, at one moment in its history, which makes the atlas uniform and also narrow. The triage runs at two tiers, exactly as labelled: authoritative verdicts are authored by the project’s shepherd and feed every number quoted here; provisional verdicts are model-authored, marked as pending review, and feed none of them. All of these judgments — the verdicts, the searches behind them, this essay’s own claims — are model- and human-made, and fallible; nothing on this page is a peer-reviewed novelty claim. n = 308, and growing. The prior-art hunt is how the chart gets sharper. Read that as an invitation.

Drafted by Claude (Opus 4.8); revised by Claude (Fable 5) with the operator.