Name-screening L1 · OFAC SDN (SAMPLE book, 69 rows)

■ SCOPE

This note reports the result of the public run of pack kyc.sanctions-adjudication.v1 under pre-run commitment No. 2. The commitment was signed and timestamped on 16 October 2026 at 14:33 UTC, before the first public run started.

■ RESULT

Outcome: works.

No gold hit row was answered clear in any pass. Each pass met every threshold set in the commitment.

All three passes reached the same outcome. Where passes share the worst outcome, the kit reports the first of them, so the graded pass recorded in the receipt is pass 1. Pass 3 had the most hold answers and the lowest hit precision of the three.

Statistical bound. With 0 missed sanctions in 25 gold hit rows, the one-sided 95% upper bound (Clopper-Pearson) on the miss rate is 11.3%. The bound applies to this book only, not to the task in general. The three passes cover the same rows, are not pooled, and do not enlarge the sample.

■ ERRORS ON GOLD HOLD ROWS

A gold hold row is one on which the correct answer is to refer the case to a human reviewer. Of the 14 such rows, the candidate decided 6 itself in every pass: it answered clear on S17, S22, S54 and S56, and hit on S55 and S57. In pass 3 it also answered clear on S20. Under the rubric these errors are not fatal; the two hit answers reduce hit precision. In an L1 screening process each of these cases would be expected to go to a human reviewer.

Consistency. The three passes gave the same answer on 64 of 69 rows.

■ COMPARISON WITH THE BASELINE

The baseline is a deterministic fuzzy matcher with no language model (baseline.ofac-style-fuzzy.v1), run on the same rows and list. Its outcome was failed: it answered clear on the gold hit rows S12 and S42. Its decision rows matched, byte for byte, the expected result recorded in section 9 of the commitment before the run. The candidate answered hold on both rows in every pass.

■ CONDUCT OF THE RUN

The first public model attempt did not complete. The API key entered for that attempt contained a character that is not permitted in an HTTP header, and the HTTP client of the runner refused each request before sending it. No row of the book reached the model. The kit marked the attempt as eligible for a rerun, which section 8 of the commitment permits for an attempt of this kind. The second attempt completed and is the run reported above.

The signed result manifest lists every attempt, baseline and model, together with the index, isolation and configuration checks made before and after each attempt. The receipts publish the model's answers as decision rows only. The model's reasoning and response text are kept in the run logs, which are not published; each receipt carries the SHA-256 of its log, and each decision row the SHA-256 of the response.

■ LIMITATIONS

• This is one run on one SAMPLE book of 69 rows with internal review only. It is not an audit, a certification or a readiness rating, and it supports no conclusion about other lists, books, models, settings or production screening.
• Because no pass produced a missed sanction, the book gives limited information about the candidate's miss rate beyond the bound stated above.
• The model is a service of its provider and is not deterministic. A new run of the same configuration may give different decision rows.
• Re-scoring confirms that the published figures follow from the published decision rows. It does not confirm that the model produced those rows.

■ STATUS OF THIS RESULT

No further run will be made under commitment No. 2. Any further run of this pack, including on a revised book, will be made under a new commitment with a new number.

■ VERIFICATION

The commitment, the kit, the book, all receipts and the signed result manifest are published in Role-Exposure/kyc-sanctions-adjudication. The command node run.mjs --rescore <receipt> --manifest RESULT-MANIFEST.json re-grades a receipt offline, without keys or network, and reproduces the figures above. VERIFY.md describes the signature and timestamp checks.