Verified, or flagged: never guessed.
Every figure either proves out against a primary source, or comes back with the exact item to pull. Checking the work stops being the work.
It can't invent a number.
It doesn't write the figure. It pulls it, and carries the document, the form and the period it came from — as a live link to the regulator's own site.
What the chip carries
- Document
- aci-20260228.htm
- Filer
- Albertsons Companies, Inc.
- Form · period
- 10-K · FY2025
- Host
- www.sec.gov
It commits only when its passes agree.
It reads the value more than once and writes nothing when the passes disagree. A slot can never hold two answers, so a disagreement reaches you flagged instead of papered over.
The passes agree, so the value is written once — with its source attached.
The passes disagree, so no value is written — the row is handed to a person instead.
One slot, one writer. Only the agreement step can write to it, so nothing overwrites a committed figure later.
It fails closed, and labels the rest.
Every line is proven, quoted, flagged, or absent — and it names the exact item to pull next. Nothing is silently estimated.
What we measured.
64 claims we had rendered as verified, re-fetched, re-parsed and recomputed independently — without looking at the original answer.
Measured silent-error rate: 0 errors in 64 verified claims, 95% confidence interval upper bound 5.7% — on value, period, and computation, checked against primary sources.
What the screening matcher finds.
A missed designation looks exactly like a clean result, so we measured it against the lists' own alternate spellings. Every screen states these figures in its own limitations, including the class where they are worst.
Measured 2026-07-25 against 20,746 designations on the OFAC and UN lists. Ground truth is the government record: each list publishes its own alternate spellings, so holding one out and asking whether the designation is still reached needs no human labelling. Reproduce it with scripts/measure_matcher.py.
See it run on a company you pick.
Bring a name. We'll run the brief live and open every source with you.
Book a demo