Clause. Teaching AI which number on a datasheet is the right one.
- Type
- AI extraction service
- Year
- 2026
- Role
- Solo builder
- Outcome
- 161 of 162 fields right, score predicted before the run, $0 spent.
Domain rules, not parsing.
Datasheets print two plausible numbers for the same question. Knowing which one is right is the real work.
Measured, not guessed.
Hand-written ground truth, measured run-to-run variance, and a prediction committed before the baseline.
Built, tested, removed.
Citations, retrieval and a verification pass were each measured, then cut when the numbers didn't justify them.
Production basics.
Auth, rate limits, request IDs and a 15-check API test suite, all on a free tier.