Articles

Writing about how this archive was built, what went wrong, and what we learned checking it. They cover the argument, the build, the measurements, how the honesty is enforced, and how the work actually gets done. Start with the first if you are new.

A specialist told us machine transcription of Yiddish handwriting is not reliable. She was right, and about more than she claimed. But the two errors that actually reversed the meaning of a letter did not come from the AI at all.

The technical companion. Seven approaches that did not work, the instruments that lied to us about their own results, and the numbers from reading every Yiddish letter four times over.

The technical account: how the archive is put together, the division of labour between the machines and the people, what the honesty checks actually check, and what we learned when they caught something.

How the archive enforces its own honesty. Nobody left in the family reads Yiddish, so the checks could not be a matter of care — they had to be built in, on both sides: the software declines to guess, and the people work to explicit contracts.

The argument about AI in family history is conducted as a contest, and the contest framing stops describing anything once you do the work. Four things this project found, an honest account of where each came from, and the two questions that both get called judgement — only one of which a machine can answer.

People keep asking for the code, which is the wrong request — almost none of it would help you. What transfers is the posture: three functions, a small vocabulary, six commitments, and a file you can hand to an agent before any design decisions get made. Deliberately incomplete, and used exactly once.

We ran the same 75 Yiddish letters through four different handwriting-recognition models to see what each one could and could not do. One model that was not built for handwriting returned fluent, confident letters that did not appear anywhere on the page — and we watched what it cost to find that out.