Methodology

Comparable only when the protocol says so.

BioBench Atlas records provenance and evaluation conditions before it records scores. This prevents superficially similar numbers from becoming a misleading leaderboard.

Inclusion boundary

Production exports accept benchmark-creator sources, official model-provider materials, peer-reviewed papers, and stable versioned preprints from arXiv, bioRxiv, medRxiv, or ChemRxiv. Independent papers use independent_reproduction. Ordinary blogs, secondary leaderboards, news summaries, and unstable or unverifiable claims remain excluded.

Five entities, five meanings

  1. Benchmark defines the scientific task, access, counts, taxonomy, and runnable status.
  2. Work identifies the versioned paper, system card, report, or official page making a claim.
  3. BenchmarkUse states how a Work used a Benchmark: creation, evaluation, training, fine-tuning, validation, model selection, or external result summary. It can remain partial when settings are missing.
  4. EvaluationRun captures one coherent, sufficiently reported scope and protocol within a work.
  5. Model preserves the exact reported model identity without fuzzy merging.
Run-splitting rule: a different subset, prompt, tool, budget, grader, repeat count, or reasoning mode is a different run—even if the work and model are the same.

Scope is explicit

full is accepted only when the realized n matches the benchmark version total. subset requires a subset ID, filter, and realized count. track refers to a formal benchmark track. When the source does not let us determine these, scope is unknown; we do not infer full coverage.

Missing means missing

Unreported values use null plus reporting_status: not_reported. Zero is never used as a missing-value code. The LifeSciBench binding example demonstrates the difference: binding is explicitly in scope, but no standalone count is reported.

Scientific Task is a separate axis

Domain describes scientific context, Capability describes what the evaluated system must do, and Modality describes the input. Scientific Task identifies the actual scientific problem—such as protein monomer structure prediction, protein sequence design, PPI, protein-ligand pose or affinity prediction, promoter detection, RNA design, cell-type annotation, or end-to-end computational analysis.

Mappings use only a creator taxonomy, formal track, or locatable official artifact. We record the most specific supported task and never duplicate its parent on the same record. Mixed suites may remain partial. Parent task pages roll up their leaf mappings, while every count claim retains its original unit and basis.

Application is not task identity: FLIP fitness prediction is not relabeled as protein sequence generation. LifeSciBench ligand binding remains broad because its source does not establish pose versus affinity. Zero-coverage task terms are shown as registry gaps.

Metrics and comparability

Metrics retain their source label, unit, direction, range, aggregation, threshold, and tolerance. Result rows only reference a metric declared by their run. Charts compare models only inside an explicit comparability_group, which must agree on benchmark version, scope, metric, and key protocol settings.

A labeled improvement is stored as a delta metric with an explicit baseline model. It is never displayed as an absolute score. Unlabeled bar heights and line positions are not digitized.

No universal life-science score. We do not normalize across benchmarks or publish a cross-benchmark “best model” ranking.

Evidence

Each critical field points to an official work plus a locator—page, section, figure, or table—and lists the JSON paths it supports. Remote papers and webpages are linked rather than mirrored unless redistribution permission is explicit.

v1.1 field-level audit

Each benchmark family moves independently from legacy to audited or audited-with-caveats. Audited records carry structured version history, permanent resource and evidence IDs, typed locators, and resolvable field paths. A provisional or conflicted value stays visible beside an explicit warning and remains machine-readable in JSON and CSV.

Current audit boundary: 14 of the 15 launch families have completed field-level review. VirBench's detailed audit was intentionally deferred by maintainer decision, so its verified v1 record remains marked legacy. CI requires every legacy record to have a matching machine-readable exception in registry metadata.
Conservative comparison rule: a provisional or conflicted total cannot establish scope: full. Provisional or conflicted result rows remain available for inspection and download but do not enter comparison charts.

Release and reproducibility

Verified YAML is validated, normalized once, and emitted as website data, JSON, CSV, and schemas. Monthly release snapshots are maintainer-triggered and immutable. Weekly link checks open issues but never mutate facts.

Local paper intake

Every week, high-precision candidates are discovered through Europe PMC, Crossref, and arXiv. At most ten issues are opened using fixed protein (4), DNA/RNA (2), small-molecule (2), and omics/cell (2) quotas. Exact benchmark aliases, creator citations, official repositories, and explicit bio/chem benchmark language are accepted as discovery signals. Candidate issues never enter the Registry.

Discovery never launches extraction. The repository owner explicitly selects an issue or paper URL before the local Codex workflow runs. Only legal open full text or a submitter-authorized source is read. Sources are capped at 45 MiB and 150 pages, fingerprinted with SHA256, kept in a gitignored local temporary directory, deleted during cleanup, and never committed or redistributed.

Independent local double-pass verification

The extractor and verifier run as two new ephemeral codex exec sessions under read-only sandboxes. They use the locally authenticated Codex CLI rather than a repository API key. Neither pass may use the network, apps, MCP servers, or outside knowledge. The extractor emits Pydantic-schema claims rather than YAML; the verifier independently relocates every page, table, figure, and number. Only claims supported at high confidence by both passes can reach the deterministic Registry generator.

Precision gate: an unlabeled plot never becomes a number; uncertain benchmark versions, model identities, or subset sizes remain a partial BenchmarkUse. A refusal, parse failure, source conflict, or stale local golden receipt pauses production and requests human review.

Local Codex opens a Ready PR under the owner's account but never merges it. A paper-intake PR passes paper-owner-gate only after wang422003 comments /approve-paper-intake <full-current-head-sha>. A later push changes the SHA and invalidates the old approval. Regular CI never invokes Codex or reads an OpenAI API key.

Explore papersContribution protocolInspect normalized JSON