Reaction Integrity Lab v1.0.0

Stable, reproducible research-product release for auditing reaction-condition benchmark integrity.\n\nHighlights:\n- Deterministic full-dataset identity, scaffold, source-category, and time audit.\n- Canonical product identity overlap: 3,784/65,444 (5.782%).\n- Nonempty valid Murcko-scaffold overlap: 51,617/63,852 (80.8385%).\n- Prespecified 1,000-product similarity audit: 60.5% have maximum Morgan similarity >= 0.70 (Wilson 95% CI 57.44–63.48).\n- Interactive threshold explorer and machine-readable v1 audit artifacts.\n- Four classical baseline figures reproduced within 0.46 percentage points of the cited paper.\n\nEvidence boundary: the upstream repository and condition archive do not provide the exact neural checkpoints or prediction files needed for independent reproduction of the published neural-model figures. Those figures remain cited reference results. v1.0.0 denotes stable audit-product maturity, not neural-checkpoint reproduction.

Reaction Integrity Lab v0.2.0

Four-cell baseline reproduction

This release establishes the first complete computational result for Reaction Integrity Lab.

• Reproduces all four ORDerly frequency-informed top-3 combined-condition baselines.
• Local values: 51.57%, 52.22%, 19.55%, and 20.24%.
• Every cell is within 0.46 percentage points of the final peer-reviewed Table 3 values.
• Corrects older reference cells retained in the upstream repository README and records the correction before model inspection.
• Adds complete Figshare v3 provenance, a deterministic reproduction script, machine-readable results, tests, and an updated interactive laboratory.

Evidence boundary

The frequency baselines and exact split-integrity audit are independently reproduced. The 67%, 68%, 35%, and 36% neural-model cells remain published reference results pending an exact archived model environment. Chemical-similarity, patent-family, and temporal audits also remain future work.

Validation

Python 3.10, 3.12, and 3.13 CI, research-registry validation, CodeQL, Ruff, strict mypy, 15 tests with 100% coverage, and package build are green.

Reaction Integrity Lab v0.1.0

Reaction Integrity Lab v0.1.0

First public research-product release of the transparent ORDerly reproduction and reaction-benchmark integrity audit.

Completed evidence

• froze the primary paper, official code, benchmark identities, licenses, checksums, configurations, and published endpoints
• independently verified both released v4 condition files against their Figshare MD5 checksums
• audited 625,697 training and 65,445 test rows (691,142 total)
• found zero exact train/test collisions on the declared reactant/product key
• found zero exact cross-split full-record duplicates
• published a tested Python audit package, machine-readable evidence, and accessible interactive laboratory

Evidence boundary

The 31/44%, 33/47%, 4/21%, and 5/24% top-3 cells remain published reference values. Independent four-cell model reproduction, seed variation, and chemistry-similarity/provenance-aware leakage analyses remain pending. Exact identity separation is not evidence of prospective wet-lab performance.

Verification

Python 3.10–3.13 CI, strict typing, lint/format checks, 100% test coverage, package build, CodeQL, and GitHub Pages are green.