# Review provenance — September 30, 2026

Target: Datamol 0.13.0 / RDKit 2026.03.6, Python 3.13 on macOS. This skill uses a
local Python API; it specifies no hosted molecular-service REST endpoints. S3/GCS/HTTPS
examples are provider-dependent paths, not verified live access or write operations.

## Primary sources checked

- [PyPI release metadata and artifacts](https://pypi.org/project/datamol/0.13.0/):
  release September 9, 2026, Python >=3.11. Runtime includes cloud/viz/Excel/SELFIES dependencies.
- [Official release](https://github.com/datamol-io/datamol/releases/tag/0.13.0) and
  [upgrade guide](https://docs.datamol.io/stable/migration.html): supported stack,
  heterocycles descriptor-key spelling, symmetry-aware RMS, binary/JSON interchange.
- [Release source tree](https://github.com/datamol-io/datamol/tree/0.13.0/datamol):
  checked installed wheel source against signatures/behavior rather than relying on stale
  prose/type hints (notably reaction return shapes and MMPA fragment strings).
- [Molecule API](https://docs.datamol.io/stable/api/datamol.mol.html),
  [conversions](https://docs.datamol.io/stable/api/datamol.convert.html),
  [fingerprints](https://docs.datamol.io/stable/api/datamol.fp.html),
  [similarity](https://docs.datamol.io/stable/api/datamol.similarity.html), and
  [clustering](https://docs.datamol.io/stable/api/datamol.cluster.html).
- [I/O](https://docs.datamol.io/stable/api/datamol.io.html),
  [descriptors](https://docs.datamol.io/stable/api/datamol.descriptors.html),
  [visualization](https://docs.datamol.io/stable/api/datamol.viz.html), and
  [conformers](https://docs.datamol.io/stable/api/datamol.conformers.html).
- [Scaffolds](https://docs.datamol.io/stable/api/datamol.scaffold.html),
  [fragments](https://docs.datamol.io/stable/api/datamol.fragment.html),
  [reactions](https://docs.datamol.io/stable/api/datamol.reactions.html), and
  [bundled data](https://docs.datamol.io/stable/api/datamol.data.html).

## Runtime scope

The repository's Datamol suite executes the local reference examples and checks parsed
failure retention, source-ID/selection correspondence, binary and JSON round trips,
descriptor definitions, fingerprint and distance shapes, selection/cluster return contracts,
Murcko frameworks, labelled BRICS fragments, MMPA pairs, fuzzy-scaffold DataFrames,
reaction product groups, and the full small bundled solubility example. It verifies
SDF/SMI/MOL/MOL2/PDB/CSV/Excel/Parquet/JSON paths, genuine PNG/SVG encodings, finite conformer
coordinates/energies, RMS cutoff, SASA, and conformer-ID/energy correspondence after selection.
Two-worker batches were also exercised. Parallel descriptor batching requires an explicit
positive `batch_size` on this stack; the default None raises in joblib.

The toy regression is an API check, not a validated benchmark. Finite conformer energy
is not proof of minimization convergence or population weighting. Notebook widget construction
was exercised; interactive browser rendering was not. Remote authentication, cloud writes,
Windows FreeSASA, other Python/RDKit combinations and large-library scaling were not tested.
Datamol explicitly disables its SASA wrapper on Windows. Source files and original molecular
records remain necessary because a parsed table cannot recover text from invalid source records.