Inspiration

  • Before a treatment reaches a clinical guideline, a trained team spends months finding every trial, reading them all, extracting the numbers, and pooling them statistically. Evidence synthesis is some of the highest-stakes, most labor-intensive knowledge work that exists — and it's a natural question whether AI agents can do it.
  • Published meta-analyses document every decision the experts made, so each stage — screening, extraction, risk-of-bias judgment, meta-analysis — can be re-posed as an independent agentic task over the same papers and graded deterministically against what the experts actually decided.
  • This is exactly the shape a verifiable-reward RL environment needs: files in, JSON out, deterministic verifiers, no judge model touching the reward — and structurally long-horizon, since one instance stages more full-text papers than fit in any context window.

What it does

  • synthesis-bench measures whether an AI agent can do medical evidence synthesis: given the full-text papers behind a published meta-analysis, extract arm-level results (means, SDs, Ns, events) and judge risk of bias, graded by deterministic code against the published expert decisions.
  • The open data version covers 13 reviews / 40 papers / 194 extraction facts plus 10 reviews / 35 papers / 178 risk-of-bias domain judgments, built on the public MetaPsy databases — with every gold value pinned to upstream release SHAs.
  • Grading is recall-only and forgiveness-asymmetric by design: honest omission costs little, fabricated numbers cost ~10× more — the failure mode that actually matters in medicine.
  • Frontier models score roughly 27–50% on extraction. Months of expert work, and the frontier is at half — that's a benchmark with headroom.

How we built it

  • One continuous six-day Codex session: 154 asks from me, 7,221 commands, 881 patches, 385 subagents spawned by Codex. I drove strategy (task fit, grading philosophy, scope decisions); Codex drove execution.
  • Codex built the acquisition pipeline that resolved DOIs, downloaded PDFs, and OCR'd 753 papers with zero failures — after self-diagnosing that its first pass used the wrong downloader and re-ordering the routes itself.
  • Gold data came from the researchers themselves, and 194/230 rows were verified to exist in the context supplied to the AI agent.

Challenges we ran into

  • Authoring defensible gold was the whole fight: the frozen instruction demanded exhaustive extraction while the gold held one hand-picked fact per paper and the grader punished correct extras — a real contract contradiction we resolved by exhaustive row adjudication plus flipping the grader to recall-only.
  • "Model failures" kept turning out to be curation bugs: the panic review's recall-zero traced to gold that was internally inconsistent with the review context every model correctly followed. Four broken reviews were excluded; the benchmark got more honest, not easier.
  • Scope contamination in long agent sessions is real: a contaminated continuation silently discarded six human-approved facts and rewrote durable goal files. Codex caught it, removed the false approvals, and restored 34 frozen provenance files byte-for-byte.
  • LLM auditors hallucinate confidently — one "corrected" a right answer (PSWQPSQW); only a deterministic source comparison caught it. Every LLM output in the pipeline now passes through a deterministic gate.
  • Due to copyright restrictions, publishing meant rebuilding both repos into fresh clean histories with leak scans and ownership guards — and correctly diagnosing a flaky pre-push test as a race rather than bypassing the gate.

Accomplishments that we're proud of

  • A fully reproducible public benchmark: anyone can clone, rebuild gold byte-for-byte from pinned public upstream data, and watch the oracle prove every instance grades exactly 1.0 — no credentials, no trust required.
  • Deterministic, degenerate-case-proof graders — no judge model anywhere in the reward path.
  • Real results with a real story: 117-cell Luna/Terra/Sol sweeps showing the frontier at 27–50%, and that more reasoning effort doesn't reliably buy more reward — one variant peaks at low effort.

What we learned

  • Put determinism everywhere you can and use the model only where you must: the winning architecture was "deterministic script + workflow for the nondeterministic LM parts," after we caught ourselves generating invariant fields with a model.
  • Never trust a single LLM pass on ground truth. Draft→audit helps, but only a deterministic comparison against source data reliably catches confident hallucinations.
  • When a benchmark scores zero, suspect the gold before the model. Our worst "failures" were curation inconsistencies, and fixing them was the most valuable error analysis we did.
  • Human-drives-strategy / agent-drives-execution genuinely works at six-day scale, if the human owns scope decisions and the agent owns catching its own drift.

What's next for Synthesis Bench

  • Evangelize Synthesis Bench and the importance of high quality evidence synthesis for medicine and public health
  • Extend to the remaining pipeline stages — full-text screening and the meta-analysis itself — so an agent can be measured end-to-end, from search results to pooled effect.
  • Benchmark more models =)

Built With

Share this project:

Updates