Skip to content

Open-work ledger

Updated 2026-10-14 for v1.1.4. This file lists unresolved work only. Completed implementation details belong in the CHANGELOG; measurements belong in the validation report and benchmark index.

Active follow-up

Issues #734 and #754 are closed. Their reaction, SMARTS, autoconf, and fuzz corpora remain recurring regression gates. Issues #739, #769, #779, and #785 are also closed; implementation and verification records remain linked from the changelog and validation pages.

  • v1.1.x artifact packet: the v1.1.4 expectation retains the 4,072-row exact HBD/QED stress lane. CI continues to compare A6 with the verified v1.1.3 wheel until the exact v1.1.4 PyPI artifact exists. Run the full hosted Linux/macOS/Windows published-wheel workflow before replacing the current broad v1.0.42 package packet.
  • Equivalent-work performance: close prepared-operation deficits without moving work into parsing; add second-host and peak-RSS records. The v1.1.4 follow-up restores prepared QED and reduces the reported parse-inclusive TPSA/Morgan regressions to statistical ties on one host; cross-platform artifact evidence remains open.
  • Compatibility CI: keep the v1.1.2 downstream exhaustive-Chirality compile fixture in every release candidate. Treat public enum variant additions as semver-significant even when the new syntax is additive.
  • Remaining unusual-input boundaries: square-planar tags do not receive tetrahedral CIP labels; wildcard ligands return typed oracle_unstable; one isotope fixture intentionally follows Sequence Rule 2 before remote Rule 5. Broader Cleanup parity remains an independent workflow task.
  • File interchange: expand the initial V3000/MOL2/CML/CDXML semantic gate to stratified corpora before broad Open Babel claims.
  • CDK/Indigo comparison: the new source workflow packet covers bounded reaction, parent-selection, and labelled R-group lanes with full mismatch accounting and 21 alternating timing blocks. Expand beyond 15/39/16 rows, add prepared-object and memory lanes plus a second host, then rerun exact published artifacts. CDK still has only the earlier smoke gate.
  • R-group contract: maintain the bounded terminal-wildcard subset across Rust, Python, and WASM; broader multi-core and enumeration behavior remains outside the contract.
  • A6 residuals: RDKit 2026.09.1 topology bounds match within 1e-12 across all 768,664 frozen-corpus cells. The conformer-tail candidate fixes a real gradient early-exit bug, aligns relevant-ring traversal, and exposes an optional selective rescue profile. Once-only v3 results are 94/100 within 1 kcal/mol at essentially tied but 0.27% slower median time; rescue is 95/100 and 2.1% slower. A6 plus holdouts v1/v2/v3 are now development data. A local MMFF distance cache funds a bounded one-candidate/100-iteration retry profile that was faster than RDKit in all three development blocks and improved the within-1-kcal/mol count to 261/265. Candidate 43619387 was frozen before an outcome-blind v4 set was selected and evaluated once. v4 reached 95/100 within 1 kcal/mol and 98/100 within 2 kcal/mol with no failures, but CheMatic was faster in only two of three blocks and the median advantage was 0.38%. Treat this as a quality improvement and speed tie. v4 is now inspected data. A separate final-gradient-gated rescue reaches 261/265 within 1 kcal/mol on the 265-row development set and was faster in all three alternating blocks (1.0179x median), but that set was used to fix the threshold. Keep atom typing, convergence, stereochemistry, clashes, same-coordinate energy, basin diversity, minimization tails, time, and memory separate; add 21-block, second-host, peak-memory, and exact public-artifact evidence before any broad claim.
  • Independent workflow corpora: expand beyond the current bounded lanes with multi-reactant reactions, multi-core R-group cases, and real MOL2/CJSON files. Keep refusals and semantic-loss reports in the denominator.

Evidence refreshes

These are recurring gates, not separate feature requests:

  • Published chemistry: rerun the reaction, SMARTS, CIP, fingerprint, and 3D gates from exact PyPI/npm/crates.io artifacts on their declared platforms.
  • Performance: compare equal outputs and equal work in alternating repeated blocks; split parsing, perception, prepared reuse, and memory.
  • New RDKit releases: pin distributed artifacts and hashes. Keep source, package, and browser lanes separate.
  • Representation limits: retain strict/reporting paths for lossy MOL, V3000, CDXML, Markush/polymer, Standard InChI, and canonical identity.

Dependency classes

  • local-open: implementation, regression tests, documentation, and source gates.
  • toolchain-open: cross-platform wheels, npm/WASM builds, and browser runners.
  • external-open: new comparator packages, upstream review, and independent adjudication.
  • data-sealed: one-time evaluation only; inspected rows never become sealed again.
  • historical: retained provenance that does not establish current behavior.

Accuracy packages

Package State
A0 Evaluation contract Complete
A1 Perception and descriptors Active
A2 Stereo and identity Active
A3 Fingerprints and retrieval Maintenance
A4 Workflows and interchange Active
A5 Independent adjudication External
A6 3D and force fields Experimental

Completion rule

For every gate, record the support domain, versions, corpus, options, failure policy, and counts for success, failure, refusal, and skipped inputs. A refusal is not a match, a local pass is not a published-package result, and registry availability is not chemistry validation.