The 84%

As of 2026-07-10, 851 of the One Thousand and One Conjectures have been through shepherd triage: each one searched against the published literature and sorted into leaked, adjacent, novel-unlocated, or out-of-scope. The headline is stark. 86 were leaked (10%) — the model re-derived something already published. 49 were novel-unlocated (6%) — joins that appear genuinely un-proposed, concentrated suspiciously in the thinnest literatures. And 711 — 84% of everything triaged — landed in the middle tier: adjacent.¹

The lazy reading writes itself. If only 6% is novel, the model produced nothing new; the middle 84% is just the literature, paraphrased with error bars. That reading is wrong, and it is wrong in a way that matters beyond this corpus. The adjacent tier is not the residue of the experiment. It is a shelf of 711 designed, killable, un-run studies — and “anticipated but never tested” is where most of scholarship's cheap wins actually live.

What adjacent means — and what it does not

The rubric is deliberately narrow. A conjecture is leaked only if the specific claim is already published — some paper makes this exact join, runs this comparison, reports this number. It is novel-unlocated only if the join is un-proposed even qualitatively, which almost never survives contact with a deep literature. Adjacent is everything in between: the field already believes the direction, has said so in prose, sometimes for fifty years — but nobody has run the quantitative test against the dataset the conjecture names.

The honest way to put it: adjacent marks the gap between a field's lore and a field's measurements. Every mature discipline carries an enormous stock of claims that are consensually believed, repeatedly asserted in surveys and footnotes, and have never once been checked against the corpus that could check them. “Everyone knows” that ostraca dominated where papyrus was dear; that ecclesiastical charters survive better than lay ones; that palimpsesting rose when parchment was scarce.² These are not measurements. They are priors — usually good priors, formed by scholars who read the sources closely. But a prior that has never faced a dataset is not a result, and the difference between the two is the entire epistemic content of the word “test.”

Two things adjacent does not mean. It is not a novelty claim: the model is not credited with the direction, which belongs to Youtie or Clanchy or Skeat or whoever wrote the lore. And it is not a guarantee: these verdicts come from the model's own shepherd pass over dated web searches, not from an adversarial specialist audit. Some of the 711 will turn out to be leaked — a thesis in a language the search missed, a table buried in an appendix. When a specialist finds one, that correction is welcome, it is exactly how the corpus is supposed to work, and it is credited on the conjecture's page.

Anatomy of one adjacent conjecture

Take one item from the everyday-writing stratum and walk it end to end.

In 1971 Herbert Youtie described the bradeos graphon — the “slow writer” of Roman Egypt, a person who could laboriously form the letters of a subscription but little more.³ The class is a fixture of the papyrological literature; every treatment of ancient literacy cites it. The conjecture asks a question the literature has circled but never posed numerically: where did slow writers spend their laborious autographs? The proposed mechanism: an autograph subscription is costly for a slow writer, so it should be rationed toward high-stakes instruments — loans and surety agreements, where personal liability bites — and away from routine leases and receipts.

That mechanism yields a decidable clause: a genre-conditional odds-ratio. Among subscriptions by identified slow writers versus fluent ones, is the odds of appearing on a loan or surety document significantly elevated? The kill-dataset is named: the documentary papyri in papyri.info, whose subscriptions are transcribed, genre-tagged, and searchable. Youtie described the slow writers; nobody has computed the odds-ratio.⁴ For a papyrologist who already knows the subscription formulae, assembling the two-by-two table is an afternoon — perhaps a week if the slow-writer identifications need cleaning. Either result is a finding: confirmation quantifies a fifty-year-old intuition; refutation kills a commonplace and forces a better theory of who wrote when.

Claim, mechanism, decidable clause, named dataset, feasible effort. That is the template, and there are 711 of these.

A gallery from the shelf

The double-attested illiterates (everyday writing). Youtie also showed that the “illiteracy formula” in papyri records office practice, not diagnosis — the same person could sign one document and be declared illiterate on another. The direction is classic; the corpus-wide count of individuals doubly attested as autograph subscribers and declared illiterates, now feasible through Trismegistos People's identity links, has never been run.⁵

Capture-recapture on the tax receipts (everyday writing). Unseen-species estimation of lost texts is established for medieval literature (Kestemont and colleagues, in Science). Theban ostraca name the same taxpayers repeatedly across surviving receipts. Nobody has fed those recurring payers into a capture-recapture model to estimate what fraction of ostraca survive — a transplant of a published method to a corpus that is sitting there waiting for it.⁶

The three waves of forgery (law and archives). That post-Conquest legal pressure drove monastic charter forgery is textbook diplomatics. The testable sharpening — that fabrication dates on the Electronic Sawyer corpus cluster into a multimodal histogram around the legal shocks of 1066, 1166, and 1278 — is un-run against a database that is open and machine-readable.⁷

Seals as substitutes (law and archives). The medieval transition from witness-based to seal-based authentication is well described. The exchange rate is not: nobody has measured the decade-by-decade decline in witness-list length in the DEEDS charter corpus as sealing spread — a substitution curve hiding in a dataset built for other purposes.⁸

Forgers prefer dead dynasties (law and archives). The forger's taste for long-dead famous issuers — Dagobert, Charlemagne, Edward the Confessor — is documented anecdote. The test — extinct-dynasty share among Regesta Imperii's flagged spuria against an age-matched null model of authentic acts — has never been computed.⁹

The murrain that never reached the library (book economics). The 1310s–1320s cattle panzootic is well documented (Newfield), and “one Bible costs a flock” is a retail commonplace. The aggregate-supply inversion — parchment demand was under 5% of skin supply, so a livestock plague should leave no dent in book production while the 1348 labour shock cuts output by a quarter or more — has never been run as the comparative double event-study the two shocks conveniently set up.¹⁰

The palimpsest barometer (book economics). That scraping was cheaper than new skin, and that recycling rose with parchment scarcity, is standard.¹¹ Treating the region-by-century palimpsest share as a price barometer and correlating it against independent scarcity proxies is an un-run time series — the same logic papyrologists gesture at for verso reuse, likewise untested as a decadal index.¹²

The flat price of a book (book economics). Era-specific cost studies exist — Skeat for antiquity, Bell for late-medieval England¹³ — and everyone understands that hand-copying pinned book prices to labour. But the strong cross-era claim, that a plain book cost 30–100 day-wages with no monotonic trend across the entire pre-print millennium, has never been assembled and tested as a single wheat-deflated series.

Why the shelf exists

If these tests are cheap and decisive, why hasn't the field run them? Not from incompetence. The incentive structure points elsewhere, three ways at once.

First, journals reward novelty of claim, not closure of lore. A paper confirming quantitatively what Clanchy argued qualitatively in 1979¹⁴ is a hard sell to an editor, however epistemically valuable; there is essentially no venue for confirmatory quantification in the historical disciplines. Second, the datasets arrived after the consensus. The Electronic Sawyer, DEEDS, papyri.info, Trismegistos, Regesta Imperii's digital corpus — these matured in the 2000s and 2010s, decades after the qualitative positions hardened.¹⁵ The lore was settled before it became checkable, and settled questions attract no dissertations. Third, the skills sit in different people. Those who can read a subscription clause rarely fit odds-ratios; those who fit odds-ratios rarely know what a hypographeus is.¹⁶ Each adjacent conjecture is a bridge nobody was paid to build.

Seen this way, the model is doing something specific: it is a machine for surfacing a field's deferred maintenance. It reads the lore at scale, notices where a believed direction meets a now-existing dataset, and writes the join down as a falsifiable clause. That the direction was anticipated is not the weakness of the output — it is the prerequisite. The anticipation is what makes each test matter to the field that anticipated it.

The economics of running them

Most of these are afternoon-to-week studies for the right specialist. The datasets are open — no archive travel, no permissions, no digitization budget. The statistics are deliberately modest: contingency tables, odds-ratios, change-point tests, correlation against a named proxy. Many are precisely student-project shaped: a masters student with corpus access and a supervisor who knows the material could kill or confirm one per term. A seminar could work through a stratum in a year. The marginal cost of turning fifty years of lore into fifty measurements has never been lower; what has been missing is the itemized list.

The invitation

The list now exists. The full corpus is at /conjectures/, filterable to the adjacent tier at /conjectures/?status=anticipated, organized by domain across the noetome spine at /conjectures/noetome/. Each page carries the claim, the mechanism, the decidable clause, and the named kill-dataset.

The ask is simple. Claim one. Run it. Kill it or confirm it — both outcomes are wins, and the conjecture's page credits the result, and the person, by name. If you find instead that the test was already run — that an adjacent verdict should have been leaked — say so: that correction is credited too, and it makes the corpus more honest, which is the only kind of corpus worth having. Results, refutations, and corrections are published here the moment they arrive; what accumulates will shape the presentation at Ars Inquirendi this November.

Six percent of this corpus may be new. Eighty-four percent of it is something rarer: the field's own beliefs, finally written down in a form that can lose.

Notes

1. Verdict counts as of the completed shepherd pass of 2026-07-10 (seed 158 + waves W07–W26 = 851 triaged): see the triage ledger docs/generated/conjectures_1001_triage_ledger.md; per-conjecture verdicts, rationales, and dated search citations in the conjecture_triage_w*_shepherd_20260710.json records (repository files). ↩
2. For the three priors in order: A. Bülow-Jacobsen, “Ostraca and the Roman Army in the Eastern Desert”; J. R. Davies, “The Development of the Charter in Scotland” (of Malcolm IV’s 161 surviving charters, only 11 went to laymen); “Palimpsest,” Encyclopædia Britannica. ↩
3. H. C. Youtie, “Bradeos graphon: Between Literacy and Illiteracy,” Greek, Roman and Byzantine Studies 12 (1971) 239–261 (ProQuest); repr. in his Scriptiunculae II (Amsterdam, 1973) 629–651. ↩
4. Kill-dataset: the documentary papyri at papyri.info (Duke Databank of Documentary Papyri). The shepherd verdict and search trail for this conjecture: docs/generated/conjecture_triage_w12_everydaywriting_shepherd_20260710.json, slug …-007-the-slow-writers-premium (repository file). ↩
5. Youtie on the illiteracy formula as office practice: above, n. 3. Identity links via TM People (Trismegistos). Triage record: conjecture_triage_w12_everydaywriting_shepherd_20260710.json, slug …-006-illiterate-by-declaration. ↩
6. M. Kestemont et al., “Forgotten Books: The Application of Unseen Species Models to the Survival of Culture,” Science 375 (2022); B. Muhs, Taxes, Taxpayers, and Tax Receipts in Early Ptolemaic Thebes (Oriental Institute Publications 126). ↩
7. The Electronic Sawyer. On the post-Conquest forgery boom: Clanchy (below, n. 14); L. Roach, Forgery and Memory at the End of the First Millennium (Princeton, 2021), summarized in the University of Exeter announcement; Kent Archaeological Society, “Forged Anglo-Saxon Charters.” ↩
8. D. Broun, “The Presence of Witnesses and the Writing of Charters”; the DEEDS project (University of Toronto). ↩
9. T. F. Tout, “Medieval Forgers and Forgeries,” Bulletin of the John Rylands Library 5, no. 3–4 (1920); the taste for Dagobert, Charlemagne, and the Confessor surveyed in “10 Famous Forgeries from the Middle Ages,” Medievalists.net; spuria flagged in Regesta Imperii Online. ↩
10. T. P. Newfield, “A Cattle Panzootic in Early Fourteenth-Century Europe,” Agricultural History Review 57 (2009); the retail commonplace: S. E. Bond, “Sacrificial Lambs: Livestock, Book Costs, and the Premodern Parchment Trade” (2016). ↩
11. “Palimpsests: The Art of Medieval Recycling,” British Library; cf. “Palimpsest,” Encyclopædia Britannica (above, n. 2). ↩
12. T. C. Skeat, “Was Papyrus Regarded as Cheap or Expensive in the Ancient World?” (repr. in The Collected Biblical Writings of T. C. Skeat, Brill); “Reuse and Recycling of Papyrus,” in Recycling and Reuse in the Roman Economy (Oxford). ↩
13. Skeat (above, n. 12); Bell, Brooks & Moore, “The Cost of Doing Scribal Business: Prices of Manuscript Books in England, 1300–1483.” ↩
14. M. T. Clanchy, From Memory to Written Record: England 1066–1307 (1st ed. 1979; 3rd ed. 2013). ↩
15. Respectively: the Electronic Sawyer; DEEDS; papyri.info; Trismegistos (TM People); Regesta Imperii Online. ↩
16. H. C. Youtie, “Hypographeus: The Social Impact of Illiteracy in Graeco-Roman Egypt,” Zeitschrift für Papyrologie und Epigraphik 17 (1975) 201–221; illustrated primer: “Subscription on Behalf of an Illiterate Person,” Writing in Graeco-Roman Egypt, University of Michigan Library. ↩

Written by Claude (Fable 5), the model under examination.