Project Lighthouse — shining AI on the pages history never read

Famous scientists left behind huge piles of handwritten pages that are digitized but essentially unread. This is an open, shared effort to read them at scale with AI, look for anything the record missed — unfinished threads, finished results nobody credited, missed connections — and then try very hard to kill every candidate before calling it a find. Anyone can contribute pages, translations, readings, reviews or expert verdicts; the data and the ledger are shared so no page is read twice and no dead end is rediscovered.

Status (2026-11-09): about 12,000 pages deep-read so far (Gauss, Riemann, Hilbert, Dedekind, Grothendieck). One lead stands: L-019, Riemann's shock-wave sheets (7/10, awaiting a palaeographer). Everything else checked so far was history already known under another name. The dead candidates are in the ledger too, with the sources that killed them — that is half the value of the repo.

The one that held

Riemann's published 1860 paper on air waves is famous for getting the energy part of the shock conditions wrong (Rankine 1870 and Hugoniot 1887 fixed it). Two sets of unpublished sheets in his Nachlass at SUB Göttingen (Cod. Ms. B. Riemann 26 fol. 70–71; Cod. Ms. B. Riemann 19 fol. 3v) contain, in his own 1856–60 notation, the energy-conserving shock relations, Prandtl's relation (1906), the Prandtl–Meyer function with a sketch (1908), oblique-shock relations with a wedge sketch, the characteristics method for steady supersonic flow (1929) and a vorticity–entropy relation (Crocco 1937). Date bracket: 1856 – July 1866. The algebra is verified by computer (verify/); a blind AI handwriting comparison with Riemann/Weber/Weierstraß controls assigns the pages to Riemann. Not yet confirmed by a human palaeographer. Full record: ledger/LEADS.md, entry L-019; the request to experts: docs/L019_expert_request_draft.md.

Scope

Any author, any field, any language, any century. Manuscripts, notebooks, letters, or printed works nobody reads any more (most of Euler is printed and still unread). The only conditions: public domain, digitized, and on a host whose terms permit fetching. Add a row to queue/QUEUE.md and a folder under data/<author>/.

How it works

1. Transcribe every page (image → verbatim text, formulas in LaTeX, struck text marked). src/transcribe.py (IIIF image URLs) or src/read_source.py (scan PDFs → page JPEGs → pages).
2. Deep-read every maths page with its neighbours: reconstruct the argument, name the modern equivalent and who is credited, score 0–10 for how surprising it would be to a specialist. src/deep.py, src/read_source.py.
3. Candidate protocol (all four, in order, before any size is stated):
   1. a human reads the page against the scan;
   2. the maths is re-derived by code (verify/);
   3. a literature search on the meaning — the author's own publications first, then older sources (the model's "first credited to" dates run late);
   4. an independent, unanchored review: a second agent/person given only the facts, never the conclusion.
4. Ledger (ledger/LEADS.md): every candidate with its verdict, including duplicates and errors.
5. False-positive guard (docs/FALSE_POSITIVES.md): the twelve standard ways a "find" turns out to be nothing, and the check for each; its checklist is required in every candidate file.

Cost: about $0.25 per page (transcription + deep read, Claude Opus 5.5). Bring your own API key (~/.config/hunt/anthropic_key, never in the repo).

Labels

Every contribution carries a label that says what was done, not what anyone believes:

label	meaning	who can set it
transcribed	page JSON exists	anyone
read	deep read exists	anyone
self-checked	finder read the scan, ran the maths, searched the literature	the finder
reviewed	the four checks were redone independently of the finder — by another contributor, or by an unanchored agent given only the facts — and a verdict with sources is written; the file says which	not the finder's own reading
expert	a named specialist signed off on the scan	that specialist

Votes sort what gets reviewed first. They never set a label.

Contribute

Fastest way in: docs/HOT_PAGES.md — pages the model rated surprising that no human has checked. Read the scan, kill it or open a candidate.

See CONTRIBUTING.md. Six ways: share transcriptions or translations you already have (any tool, any model, by hand — no analysis required); run the scripts on a new source; fix transcriptions by hand; review a candidate; give an expert verdict. data/<author>/STATUS.tsv shows which bundles are only transcribed and still need deep reads. The scarce input is not tokens — it is people who can read the page and know the literature.

Priority and citation

Everything here is public the moment it is pushed. For any finding, the date of the first commit or release containing its ledger entry is its public date; a GitHub release archived on Zenodo gives it a DOI and an independent timestamp. Cite the repository via CITATION.cff. Transcriptions are CC BY: use them, but say where they came from.

Sources and rights

Scans belong to the libraries (SUB Göttingen / GDZ for Gauss, Riemann, Hilbert, Dedekind — public domain; Université de Montpellier for Grothendieck). Scans are linked, never hosted. Transcriptions and analyses in this repo are CC BY 4.0; code is MIT (see LICENSE). Grothendieck exception: per his 2010 declaration this repo carries only analyses and short quotations of his unpublished notes, not full transcriptions.

Data

data/<author>/transcripts/<bundle>/cNNN.json and data/<author>/deep/<bundle>_cNNN.json; formats in data/FORMAT.md. Gauss ~6,300 files, Riemann ~5,800, Hilbert ~1,200, Dedekind ~175, Grothendieck deep reads ~5,000 (transcripts withheld).

Next targets

Euler's 866 published works (public domain, scanned at the Euler Archive, mostly Latin, mostly untranslated and unread) — queue in queue/. Ramanujan's notebooks (public domain; fully proved by Berndt and Andrews–Berndt, so only missed connections are plausible).