An AI in the archive, in public

The reading

The dig follows an expedition: a topic mapped to searches in the Library of Congress's Chronicling America newspapers (1770–1963), the Internet Archive's public-domain books and journals, and the Biodiversity Heritage Library. Each page's machine-read text (OCR) goes to claude-haiku-5-5, which writes the reading notes you see on the stage and flags passages worth a closer look, with an exact quote.

The checks (code, not the AI)

A flagged quote must appear on the page, word for word or with only OCR noise; otherwise the flag is thrown away, whatever the AI said. The quote's position on the scan comes from the archive's own word coordinates, which is how the stage can light it up. Everything we display passes a blocklist: old newspapers contain slurs and cruel material, and none of it is shown in our text (the scans are shown as they are).

The verdict

Strong flags go to claude-sonnet-5-5, which reads the passage in context, rejects ads, fiction and anything unsafe, and writes the headline, the explanation, the year and place, and how forgotten it seems. Then it searches the web (Claude's web search tool, a few searches at most) to ask whether the event is documented today, and Wikipedia gives a free second opinion. That sets the label: POSSIBLY UNKNOWN (no modern record found), ALREADY KNOWN (with the modern source linked) or VERIFIED QUOTE (the quote is on the scan; novelty not established).

Investigations

When a find scores high, the AI opens a numbered case and works it like an archivist: it searches other newspapers from the same week and region, reads the most promising pages (they appear on the stage), notes what confirms or contradicts the story, and closes the case with a timeline and a conclusion. It has hard limits: 12 tool turns, 8 pages, $0.75 per case and 5 cases a day. Every quote in a case file is checked against its page.

Ask the Archive

Holders can ask one question a day. A small model screens it (history only, no abuse, no personal data); the AI then digs with the same tools and answers with citations to the scans. Citations whose quotes aren't on their page are removed. Answers are public on /asks. Limits: 20 questions a day in total, $0.25 each.

The Dig Gazette

Every evening at 20:00 UTC, claude-opus-5-5 writes a front page in the style of an old broadsheet from the day's finds and cases. It can only refer to real finds; their verified quotes and scans are printed next to its stories.

Who pays: the DIG FUND

The coin's creator fees are shared through pump.fun's fee sharing: a share goes to a public wallet, the DIG FUND. The team tops up the Anthropic API credit from that wallet and lists every top-up, with the SOL transaction that paid for it, on the ledger. The fuel tank on the homepage is simply: SOL received × 100% (the AI share) × the SOL price, minus what the AI has spent.

The reading speeds up and slows down with the fuel: with more fuel, more pages an hour (between one every 10 s and one every 300 s), and never past the hard daily cap of $10.00. There is a kill switch. Visitors never trigger AI calls: one shared reading loop is streamed to everyone, so a thousand viewers cost what one costs.

Every token on the ledger

Each API call's cost is computed from the API's own usage report (input, output, cache reads and writes, web searches) and the published prices, and stored per page, per find, per case, per answer and per edition. It's all public on the AI ledger. Backlog digs use the Message Batches API at half price, and the reader's long instructions are cached so each page costs less.

What it isn't

The AI can be wrong, which is why every find links to its scan. "Possibly unknown" means our searches found nothing, not that nobody knows. The archives are public domain; THE DIG is not affiliated with or endorsed by the Library of Congress, the Biodiversity Heritage Library or the Internet Archive. $DIG is a memecoin; the fees fund the reading, nothing more is promised.