hannesill/m4bench-evidence

M4Bench per-run evidence

Open, sanitized per-run evidence for the paper M4Bench: Evaluating Procedural Specification for Clinical EHR Derivation Agents (NeurIPS 2026, Evaluations and Datasets Track): 1,960 agent runs on 28 clinical EHR derivation tasks over MIMIC-IV and eICU.

The benchmark itself (tasks, skills, isolation harness, evaluator, and the scripts that build and check these packages) is the M4 repository: https://github.com/hannesill/m4, tag m4bench-neurips2026-camera-ready.

No patient data. The packages contain no MIMIC-IV or eICU source data, task databases, reference outputs, or row-level agent outputs. Rebuilding the task databases and re-evaluating agent outputs requires independent PhysioNet credentialed access.

Files

File	Runs	Contents
m4bench-evidence-submission.tar.zst	1,400	May 2026 submission campaigns: 760 primary Codex, 190 Claude sentinel, 340 Codex follow-up, and 110 exploratory open-source-model runs
m4bench-evidence-claude-replication.tar.zst	560	July 2026 full-coverage Claude replication: 28 tasks x Opus/Sonnet x no-skill/with-skill x 5 trials
*.tar.zst.sha256		SHA-256 checksum of each archive
*.tar.zst.manifest.json		Size, checksum, and file counts of each archive
m4bench_evidence.croissant.json		Croissant 1.0 metadata (core and RAI) for both archives
metadata/run-indexes/*.csv	1,960 total	Five directly downloadable canonical run indexes, identical to the copies inside the archives

The Croissant record sets load the CSV files in metadata/run-indexes/ directly; no archive extraction is needed to read the run indexes. The evidence archives retain their existing bytes and SHA-256 checksums.

What a run holds

• result.json: run identity, isolation and canary checks, evaluator results.
• trace.jsonl: the agent trace. Agent messages, SQL, and code are kept.
• output_summary.json: column names, row count, and SHA-256 of the withheld output.csv.
• instruction.md and egress.jsonl: task instruction and network egress log. The exploratory open-source-model runs have no egress log.

Each archive also carries metadata/public_artifact_manifest.json, which maps every file to its source path, source SHA-256, transform, and public SHA-256, and the run manifests used for the paper's tables.

What is withheld

• Tool and command outputs, because they contain query results from credentialed data. Each is replaced by a placeholder with its length and a keyed hash prefix.
• Events that are not valid JSON in the source and truncated events at the start of a stdout tail. Their tool output cannot be separated from agent text, so the whole event is withheld (3,810 in the submission package, 559 in the Claude package).
• In the agent text that is kept: credentials, date-shifted timestamps, patient ages, and rows that start with an identifier are removed, and MIMIC-IV and eICU identifiers are pseudonymized.

Agent messages are otherwise kept as written and can describe in words what the agent saw.

Download and verify

hf download hannesill/m4bench-evidence --repo-type dataset --local-dir m4bench-evidence
cd m4bench-evidence
shasum -a 256 -c m4bench-evidence-submission.tar.zst.sha256 m4bench-evidence-claude-replication.tar.zst.sha256
mkdir -p evidence-claude evidence-submission
zstd -dc m4bench-evidence-claude-replication.tar.zst | tar -x -C evidence-claude
zstd -dc m4bench-evidence-submission.tar.zst | tar -x -C evidence-submission

From a checkout of the tagged M4 release, each extracted package can then be checked against its manifest, used to regenerate the paper's result tables, and scanned for residual private content:

python3 benchmark/release/v1/scripts/verify_public_evidence.py /path/to/evidence-claude
python3 benchmark/release/v1/scripts/scan_public_evidence.py /path/to/evidence-claude
uv run python benchmark/release/v1/scripts/verify_public_evidence.py /path/to/evidence-submission
python3 benchmark/release/v1/scripts/scan_public_evidence.py /path/to/evidence-submission

The Claude check needs only Python 3.9 or later. The submission check needs the benchmark environment (uv sync). The release README in the repository (benchmark/release/v1/README.md) gives the full procedure and lists the code and skill state behind each campaign.

Limits

• Scores measure agreement with declared public-code reference outputs, not clinical correctness.
• The Claude replication harness was not uniform (two CLI versions, mixed model routing); the paper reports it descriptively.
• Runs record task, condition, model, trial, timestamps, and isolation checks, but not the source commit or container image digest.

License

MIT, as for the M4 repository.

Citation

@inproceedings{ill2026m4bench,
  title={M4Bench: Evaluating Procedural Specification for Clinical EHR Derivation Agents},
  author={Ill, Hannes and Al Attrach, Rafi and Fani, Rajna and Han, Ahram and Yamamoto, Ryohei and Moerig, Marlene Louisa and Yoon, Dukyong and Celi, Leo Anthony},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year={2026}
}