Technical reference
Start with the scientific guide to interpret predictions, or the reproduction guide to run the retained examples. These references explain the inputs, dependencies and contracts behind those guides.
Scientific inputs and attribution
- Source attribution and licence evidence — upstream code, released artifacts, byte identities and limits of known provenance.
- Structure-reading and SMILES fixtures — atom order, units, conformers, independent expectations and scientific references.
Dependency decisions
- Inference dependencies and source decisions
- RDKit dependencies and native-reader qualification
- Provider runtime dependencies
- Offline checker dependencies
Dependency reviews are dated evidence. Their recorded advisory results do not establish that the same versions have no subsequently reported vulnerabilities.
Reusable provider package
- Package requirements and runtime contracts
- Normative, adversarial, contract and integration tests
- Pinned API contract fixtures and release provenance
Repository checks
Run and maintain repository checks — choose a focused command, understand its coverage, and resolve findings while preserving the underlying requirement.
- Container and build checks: source crosswalk
- Code-quality checks: port provenance and rationale
- Scientific checks: port provenance and rationale
Documentation and retained data
The maintained prose lives in these HTML pages. The short root README points here; there are no parallel Markdown guides or document-generation step. Design records retain dated decisions and evidence separately from current instructions. Literal record text keeps its original line anchors and commit-pinned citations, including paths used at the time. Page metadata identifies each record's previous source and revision.
The non-HTML text exceptions are the root README.md, the machine-read
source manifest containers/inference/image-inputs.txt, and the scientific
fixture tests/fixtures/small_chloromethanes.txt. JSON manifests, lockfiles,
configuration examples, source comments and legal notices remain in their native
formats. The pinned upstream submodule is unchanged.
Frozen scientific inputs and expected-output files retain their original bytes, including historical documentation paths. The current fixture explanation is Structure-reading and SMILES fixtures; moving documentation does not authorize changing an independent numeric oracle.
Run make test/repository/documentation for the offline page/link checks.