Codex Science v0.4.0
Highlights
- Adds read-only Crossref and PubMed metadata connectors plus keyed OpenAlex support.
- Preserves sanitized provider snapshots and imports only explicitly selected metadata records.
- Adds an eight-task transparent scientific-agent benchmark with immutable outputs, deterministic grading, optional blinded human scoring, and same-suite comparison.
- Integrates connector snapshots and evaluation runs into project validation and research packets.
- Hardens credential routing, redirect handling, release validation, tests, and Node 24 GitHub Actions.
Verification
- 19 offline unit and integration tests.
- CI passes on Python 3.10 and 3.12.
- Live read-only Crossref and PubMed smoke tests passed.
- OpenAlex support is implemented but requires a user-owned
OPENALEX_API_KEY; no live OpenAlex run is bundled.
The bundled benchmark supports regression and exploratory same-task comparison. It is not proof of overall scientific intelligence or complete Claude Science parity.