[abrichr]

Follows #1795. Fara-7B's first half (57 rows, P01–A06, stack 9820) ran before the verifier's token fix, so most of its live checks failed with HTTP 401. This PR replays them on the frozen server state, the same way #1795 replays claude and openai. It runs no new agent.

Pre-registered first. PREREGISTRATION-FARA.md and SHA256SUMS-FARA were committed (22429e7) before any Fara verdict was computed. Every pinned file is unchanged in the results commit and on current main.

Column	Sensitivity	Specificity
live	15/15	1/2 (the miss is HTTP 401)
main, operations, operations-v2	15/15 [Wilson 0.80, 1.00]	2/2 [0.34, 1.00]

• Fidelity gate: 4/4. These are the only Fara rows the live verifier judged.
• Few truthful successes: Fara said DONE on 17 of 57 runs and was wrong on 15. This half therefore tells us little about specificity.
• Unlabelled flags: 5 runs with no DONE claim are flagged because Fara wrote after the check. They count in neither metric.

Changes

• Inputs: inputs/frozen-9820.jsonl.gz and inputs/rows-fara.jsonl.gz, from Write leaderboard: computer-use agents in Medplum's web app, pre-registered #1737 at 01854c3; each row keeps its source path and sha256. Fara's times are merged into inputs/times.json.
• rescore.py: a fara: 9820 stack, plus CHECKED_BEFORE, which skips Fara rows checked at or after 2026-11-04 11:55 UTC (those have live verdicts). The default agents are still claude and openai.
• score_rescore.py: --agent; the default is still the pre-registered pair.
• Results: results/*-fara.json, plus a section in RESULTS.md and the inventory entries. The claude and openai results are not re-run.

Public-data check: the frozen rows and dumps hold only synthetic bench-project records, example.org and .invalid emails, and no tokens or secrets.

[abrichr]

Independent adversarial review: APPROVE, no blockers

A separate agent with no part in writing this change reviewed head e59fbd0 read-only.

Claims checked

• The pre-registration commit is the parent of the results commit, and the results commit changes no file pinned in SHA256SUMS-FARA; shasum -c passes.
• All 57 rows match their source blobs at 01854c3: sha256 and every field. None are missing or extra. frozen-9820.jsonl.gz is byte-identical to that commit's copy.
• The 69 fara keys equal file-times.json, and all 408 existing keys are unchanged.
• The default agents and scoring are unchanged for claude and openai. The UTC cutoff is correct and skips none of the 57 rows.
• RESULTS.md, README.md and the PR body match score-fara.json. A re-run of --contracts operations reproduces all 57 rows.
• No secrets, real-looking PHI or private names are in the data.

Minor items and responses

• Live 15/15 comes from 401 failures, not detection. Fixed in 319b4d8: the table and text now say so.
• Commit dates one second apart. The rebase onto main after Verifier specificity: operation semantics as contract data, and the leaderboard re-scored #1795's squash merge rewrote the committer dates. The author dates (67 s apart) and the original pre-rebase commits (40f03d9, then 2d3c50c) keep the order.
• Cutoff time. The 11:55 UTC start of the second half comes from the leaderboard harness's operator. The cutoff only matters for rows after it, and none of the 57 rows is later than 11:44:52 UTC.
• "Median steps 52" is unchecked. It is the steps field of score_rescore.py's outcome vector, printed in its output.