renkin 0.22.0 → 0.23.0

package/README.md

@@ -352,7 +352,7 @@ for the full acceptance criteria and licensing split.
 | **Partner-free retrieval hints** | `renkin-forward hints --reactants <SMILES>...` — no partner input at all: reports matched template slots, missing-partner SMARTS, and bond deltas for patent/database search, never a concrete product — see the [Forward Retrieval Hints guide](docs/guides/forward-retrieval-hints.md). `predict` / `enumerate` / `hints` compared: [table](docs/guides/forward-retrieval-hints.md#predict--enumerate--hints-at-a-glance) |
 | **Forward validation** | `renkin-forward validate` verifies each step by applying templates forward; accepts `--route-json` or stdin |
 | **Ring-context safety guard** | `--ring-context-policy conservative --ring-context-sidecar <path>` — opt-in match-level filter that rejects an extracted template's ring-opening/closing disconnection when its historical training data never observed that bond as ring-forming/-breaking; default is `disabled` (unchanged legacy behavior) — see [Issue #72](https://github.com/kent-tokyo/renkin/issues/72) |
-| **LightGBM candidate reranker** (experimental, Issue #101) | `--reranker-model model.txt --reranker-freq-table frequency_table.json` — opt-in, ordering-only: re-ranks same-step candidates using a frozen LightGBM model, expressed as a rank-derived bonus on the same scale as the template-frequency bonus. Never changes which candidates are generated, only their search order. Omitting either flag (the default) reproduces legacy ordering byte-for-byte; a bad model/table path falls back to legacy ordering with a stderr warning rather than failing the run. Pure-Rust model reader, no C/C++ dependency. Pending its runtime paired-gate result before this row loses the "experimental" tag. |
+| **LightGBM candidate reranker** (Issue #101; CLI shipped v0.22.0, Python surface + batteries-included distribution shipped v0.23.0) | `--reranker-model model.txt --reranker-freq-table frequency_table.json` (CLI) or `reranker_model_path`/`reranker_freq_table_path` (Python `find_routes()`) — opt-in, ordering-only: re-ranks same-step candidates using a frozen LightGBM model, expressed as a rank-derived bonus on the same scale as the template-frequency bonus. Never changes which candidates are generated, only their search order. Omitting either flag/param (the default) reproduces legacy ordering byte-for-byte; a bad model/table path falls back to legacy ordering with a stderr warning rather than failing the run. Pure-Rust model reader, no C/C++ dependency. Paired 100-target route-search gate: `route_to_configured_stock` 16→20 (+4/-0). The trained `model.txt` is not bundled into any published package (its USPTO-50k training data's license is undocumented upstream — see `docs/guides/open-source-retrosynthesis-comparison.md`'s "Known gaps"); fetch it (alongside a re-verified `frequency_table.json`, which is otherwise already committed/bundled) with `python3 scripts/fetch_reranker_model.py` — downloads from GitHub Release assets already attached to the v0.22.0 release, verifies each with a double SHA-256 check. |
 | **Plausibility report** | `renkin-bench --plausibility` — forward-validates best routes and reports composite plausibility score |
 | **PaRoutes benchmark** | `renkin-bench --input-format paroutes` for multi-step ground-truth evaluation with `depth_delta` and `route_diversity` |
 | **Atom balance check** | `renkin-bench` flags steps where `target_MW > Σ precursor_MW` (CompleteRXN reference) |
@@ -623,7 +623,8 @@ renkin/                          ← Cargo workspace root
 - [x] `apply_retro`/`run_reactants` performance regression resolved — `chematic` moved from a narrow git-pinned fix to the published `0.8.0` release (upstream automorphism-orbit-pruned canonicalization, [chematic#193](https://github.com/kent-tokyo/chematic/pull/193)); on a fixed 30-target gate, measured in one session against current master: total elapsed **34.7%** faster, p95 **33.8%** faster, and the single worst-case target **42.2%** faster (confirmed via repeated isolated measurement, not a one-off run). Zero correctness change (`apply_retro` call counts identical across versions)
 - [x] `renkin-forward` CLI hardening — versioned `ForwardPredictionReport`, deterministic candidate IDs/merge/provenance, reactant-order-independent matching (up to 3 reactants), strict CLI/route-JSON validation
 - [x] RETROSPECT-inspired offline candidate-reranking foundation — candidate proposal/selection separation, feature schema v1, manifest v2, leakage-safe train/val/test splitting, 7 deterministic baseline arms + trained-ranker arm, paired bootstrap + offline gate tooling ([#59](https://github.com/kent-tokyo/renkin/pull/59))
-- [x] Offline candidate reranker trained and gated against real data ([#101](https://github.com/kent-tokyo/renkin/issues/101) Task 35) — LambdaMART model trained on real USPTO-50k labels, passed its VAL screening gate (top1 +11.7pp, MRR +11.3pp, top10 +9.3pp, bootstrap-CI-confirmed), frozen, and passed a formal 4,903-target TEST evaluation against the frozen model exactly once (top1 +12.7pp, MRR +11.9pp, top10 +9.1pp — consistent magnitude with VAL, no overfitting signal). **Offline candidate-ranking gate only — not yet wired into route search**, see "In progress" below
+- [x] LightGBM candidate reranker, trained/gated offline and wired into route search ([#101](https://github.com/kent-tokyo/renkin/issues/101) Task 35, CLI shipped v0.22.0) — LambdaMART model trained on real USPTO-50k labels, passed its VAL screening gate (top1 +11.7pp, MRR +11.3pp, top10 +9.3pp, bootstrap-CI-confirmed) and a formal 4,903-target TEST evaluation against the frozen model exactly once (top1 +12.7pp, MRR +11.9pp, top10 +9.1pp — consistent magnitude with VAL, no overfitting signal), then wired into `find_routes` as an ordering-only rank bonus and confirmed with a paired 100-target route-search gate: `route_to_configured_stock` 16→20/100 (+4/-0). See the Key Features table above
+- [x] Reranker made actually usable: Python exposure (`find_routes()`'s `reranker_model_path`/`reranker_freq_table_path`) and batteries-included model distribution (`scripts/fetch_reranker_model.py`, SHA-256-verified fetch from the v0.22.0 GitHub Release's canonical assets) ([#101](https://github.com/kent-tokyo/renkin/issues/101), shipped v0.23.0) — v0.22.0 proved the reranker works; v0.23.0 is the usability/distribution unlock, not a new accuracy claim
 - [x] Stable `template_id` (`rule:<name>` / `smirks-sha256:<hex>`) + `--template-metadata` evidence sidecar + `renkin template ids` ([#41](https://github.com/kent-tokyo/renkin/issues/41) phase 1)
 - [x] Substrate-specific `examples` (`schema_version: 2`) — per-step exact-substrate vs. same-template-different-substrate resolution, surfaced in `--format explain` and as `match_kind` in JSON ([#41](https://github.com/kent-tokyo/renkin/issues/41) phase 2)
 - [x] Deterministic ORD (Open Reaction Database) evidence import — offline `renkin evidence match` exact-set batch template matcher + `scripts/ord_evidence_audit.py` audit/converter into `schema_version: 2` sidecars; no network access, no fuzzy matching, ambiguous/unprovenanced records excluded and counted in an audit report rather than guessed at ([#41](https://github.com/kent-tokyo/renkin/issues/41) phase 3A)
@@ -636,7 +637,7 @@ renkin/                          ← Cargo workspace root
 
 ### In progress
 
-- [ ] Runtime integration of the trained candidate reranker ([#101](https://github.com/kent-tokyo/renkin/issues/101) Task 35, see the Key Features table above) — implementation and tests complete; a paired 100-target route-search gate (reranker ON vs OFF, same commit/corpus/stock/templates/budget) is the remaining step before this is reported PASS/FAIL. No release/tag either way until then
+- [ ] Candidate-generation coverage gap — 33.0% (1,618/4,903) of the formal TEST corpus has zero positive candidates in-pool, a ceiling reranking cannot fix by construction; template-diversity-scaling and higher-level-template research directions identified, not yet started
 - [ ] Template retrieval index (element bitmask + bond-center prefilter) for the 50k template set
 - [ ] Calibrated route confidence (map `success_probability` to empirical solve rate)
 

package/package.json

@@ -5,7 +5,7 @@
     "kent-tokyo <kent-tokyo@users.noreply.github.com>"
   ],
   "description": "Ultra-fast retrosynthesis engine for computer-aided synthesis planning (CASP) — pure Rust, WASM-ready, Python bindings via PyO3",
-  "version": "0.22.0",
+  "version": "0.23.0",
   "license": "MIT",
   "repository": {
     "type": "git",