Where the data comes from
Every published record has an evidence table. Use it to distinguish a checked scientific claim from a general reference, an unresolved detail or a catalogue decision.
Download the source-origin tables
Release 2026-09-16-1deea030a998, published 2026-09-16. The table contains 45,747 rows across 1,669 records. A statement citing two sources occupies two rows; this is not a count of independent findings.
Each row stores the stable record and field, original value, source URL and DOI, version, evidence location, retrieval date, artifact hash and its scope, review method and review date. Empty cells mean that information was not recorded. Hashes identify inspected bytes; they are not a guarantee that a scientific claim is correct.
Use record_id to select a model, benchmark or result. Use field_path to find its original field, and source_id to join the source record. The JSONL value_json preserves types, nulls and exact strings. CSV cells that could be spreadsheet formulas are escaped with a leading apostrophe; use JSONL for lossless machine processing.
Read the review scope first
| Scope | Rows | Meaning |
|---|---|---|
| Individual claim | 11,556 | A specific statement has its own citation and location. Check the review status: an unreported field or conflicting claim is not a verified value. |
| Context-only reference | 8,630 | The source is attached to the record, but this individual field has not been separately checked. This includes unresolved evaluation conditions and inherited discovery metadata. |
| Source metadata | 11,216 | Bibliographic and retrieval metadata describes an inspected resource. A source’s review status does not verify everything it contains. |
| Catalogue metadata | 14,345 | Rewire’s identifiers, classifications, review history and preserved migration fields. These are editorial or administrative records, not experimental observations. |
Profile explanations were checked against primary papers, supplements, official implementations and documentation through automated source review. This is not human sign-off. A source-checked score confirms the reported transcription; it is not an independent rerun. Existing rewire runs keep their original run history and have not been rerun for this release.
A locator shared by several references is marked shared_claim_locator. It identifies the evidence for the statement as a whole, not a separately verified location in every cited source. Source disagreements and superseded records remain visible.
Coverage and remaining gaps
All 226 model records and 170 benchmark/task records have explanatory profiles. Their 4,391 structured facts include 3,494 source-checked facts. Missingness and applicability are counted separately:
- unreported: 647
- unextracted: 0
- unavailable: 0
- inapplicable: 250
These counts cover profile facts only. Dataset, baseline and evaluation records still contain fields needing extraction or review, including exact checkpoints, splits, adaptation and scoring details. The evidence tables expose those gaps rather than treating a linked paper as proof. Unknown conditions prevent automatic comparisons.
The catalogue is a dated collection, not a complete census of biological models. Popularity, citations and repository activity do not measure biological performance. Results remain attached to their actual configuration and protocol; we do not combine incompatible scores into a universal leaderboard.