yomitoki

Fast, explainable, route-free molecular synthesizability diagnostics.

yomitoki is a fast, explainable, route-free molecular synthesizability diagnostics library built on chematic.

Instead of returning only a synthetic accessibility score, yomitoki reads molecular structure and explains why a molecule appears easy or difficult to synthesize. It identifies the structural evidence behind the assessment and reports how confident that judgment is.

The name comes from the Japanese word yomitoki（読み解き）, meaning to carefully examine something and uncover its meaning — that is the library's whole job: not to change the molecule, but to read it and explain what it finds.

yomitoki does not merely estimate synthesizability; it exposes the evidence and reasoning behind the estimate.

Status: 0.2.0-alpha.1 — first release of the v0.2 accuracy-redesign generation, built on 0.1.1. Two real scoring-behavior changes so far: ring_topology's ring-family aggregation moved from a plain sum to an L2 norm, and size_topology gained a heteroatom-count burden term (weight frozen at 0.03 after a dedicated calibration attempt). Both are development-set-only findings (MPScore) — TS1/TS2/TS3 and a new holdout haven't been re-evaluated yet, which is why this is still an alpha; see CHANGELOG.md for the full evidence behind each change. All six planned components are implemented, including fragment_precedent (renamed from fragment_rarity in round 18, since it argues difficulty both up and down, not just up) — opt-in: no fragment-frequency corpus ships with yomitoki itself (AGENTS.md §5.4 forbids embedding one directly as a huge binary), so it stays inactive unless you build one (tools/build-fragment-corpus/) and configure it (AnalysisConfig.fragment_model). fragment_precedent is an explanatory reference-corpus signal, not a direct synthetic-difficulty term — round 20 found the corpus-relative signal too corpus-sensitive to trust as a scoring input (two honestly-labeled synthesis-focused corpora disagreed with each other on its direction 34.6% of the time over 500 probe molecules; plain pyridine swung between LikelyAccessible and HighlyChallenging purely from which corpus was configured), so round 21 removed it from overall.difficulty entirely (option C). If you configure a corpus: fragment_precedent is still computed and still reported (SynthesizabilityReport.fragment_precedent) as explanatory evidence, but it can no longer change overall.difficulty, dominant_penalties, or dominant_supports — configuring a corpus (or switching which one) never changes the score, only the evidence available alongside it. See Limitations for the full before/after data and rules.rs's "Fragment precedent" section for the complete round 16–21 history. See CHANGELOG.md for the full version history. See docs/architecture.md for the current scope and what's still missing.

Where it sits

chematic    Molecular representation and cheminformatics
    |
yomitoki    Read and explain molecular synthesizability
    |
renkin      Plan retrosynthetic routes

chematic · renkin

yomitoki never runs route search — that boundary is permanent, not a v0.1 scoping choice. See "What it does not do" below.

What it does

Parses a molecule (via chematic) and returns a structured SynthesizabilityReport, not a single number.
Breaks the assessment down into independent components (ring topology, size/topology, stereochemical burden, functional-group liabilities, input quality/applicability, and fragment rarity — the last is opt-in, see "Limitations" below).
Separates score (synthesizability/difficulty), confidence (how reliable the judgment is), and applicability (whether the molecule is even in the model's domain) into distinct fields — a hard-to-make molecule is not automatically a low-confidence one.
Emits machine-readable finding codes with structured evidence, not just prose.
Never runs retrosynthesis search. yomitoki evaluates a molecule on its own; it does not plan a route to make it.
Ships a yomitoki CLI for single-molecule and batch (.sdf/SMILES-file) analysis — see Command-line interface below.
analyze_batch(&[Molecule], &AnalysisConfig) -> Vec<Result<...>> gives library callers the same input-order-preserving batch entry point, without going through the CLI or a file format.

What it does not do

Retrosynthesis planning, reaction template application, precursor generation, or route ranking — that's RENKIN's job.
Molecule parsing, ring perception, aromaticity, or stereochemistry assignment — that's chematic's job; yomitoki only consumes it.
Toxicity, SDS/hazard classification, yield prediction, or cost prediction.
Full periodic-table or organometallic/polymer coverage in v0.1.

Quick start

use yomitoki::{analyze_smiles, AnalysisConfig};

let config = AnalysisConfig::default();
let report = analyze_smiles("C1CC2CCC1C2", &config)?; // norbornane

println!("{:?}", report.overall.verdict);
println!("difficulty = {}", report.overall.difficulty.value());
println!("confidence = {}", report.overall.confidence.value());

for finding in &report.findings {
    println!("{:?}: {}", finding.code, finding.explanation);
}

Run the full example:

cargo run --example basic

Command-line interface

yomitoki analyze "C1CC2CCC1C2" --format json
yomitoki analyze --input molecules.sdf --format jsonl --output reports.jsonl

yomitoki analyze "<SMILES>" [--format human|json|jsonl] [--fragment-corpus <dir>] — analyze one molecule from an argument.
yomitoki analyze --input <file> [--format human|json|jsonl] [--output <file>] [--fragment-corpus <dir>] — batch mode. <file> may be a .sdf file or a SMILES-per-line file (optionally with a whitespace-separated name column, the standard .smi convention).
Batch mode preserves input order and never stops on one record's failure — a failed record becomes an error entry (JSON "error" field, or an ERROR: block in human format), not a skipped one. The process exits non-zero only after every record has been attempted, if any failed.
jsonl output uses the same {"input", "report"|"error"} wrapper shape in both single-molecule and batch mode, so a downstream line-by-line parser sees one schema regardless of which invocation form produced it.
Exit codes: 0 success, 1 a molecule failed to parse/analyze (single mode) or at least one batch record failed, 2 a usage error (bad arguments).
--fragment-corpus <dir> loads a tools/build-fragment-corpus output directory and enables fragment_precedent for the run (loaded once, before any molecule is analyzed). Without it, reports have fragment_precedent: null, same as before this flag existed — no corpus ships with yomitoki itself, see Limitations below.

Report shape

Actual output of cargo run --example basic, current as of this component set. Norbornane (C1CC2CCC1C2) exercises ring topology, and its bridged ring produces a simplification suggestion:

Verdict: ModeratelyAccessible
Synthesizability: 0.66
Confidence: 1.00
Dominant penalties:
1. Bridged ring system spanning 7 atoms — bridgehead connectivity typically increases synthetic difficulty.
Simplification suggestions (heuristic, not a guarantee):
1. ReplaceBridgedRingWithMonocyclicAnalog: Bridgehead connectivity in this ring system is a direct driver of the ring_topology contribution to difficulty. A monocyclic (or less-fused) analog, if the target application allows one, would remove this specific burden — this is a structural heuristic, not a guarantee the replacement is chemically equivalent or that synthesis actually becomes easier.

A stereocenter-dense fragment (CC(O)C(N)C(C)C(O)C(N)C) exercises stereochemical_burden and functional_group_liability (difficulty) plus applicability's independent confidence penalty for unspecified stereochemistry — note that difficulty and confidence move separately, not together:

Verdict: ModeratelyAccessible
Synthesizability: 0.69
Confidence: 0.85
Dominant penalties:
1. 5 tetrahedral stereocenter(s) (specified or unspecified) requiring synthetic control.
2. Stereocenter density 0.42 is above the 0.25 threshold — stereocenters are concentrated in a compact region, leaving little room for staged, orthogonal control.
3. Reactive/unstable functional group detected: primary amine (Brenk et al. 2008 structural alert).
Simplification suggestions (heuristic, not a guarantee):
1. ReduceStereocenterDensity: Stereocenters are concentrated in a compact region, leaving little room for staged, orthogonal stereocontrol. Reducing the number of stereocenters, or spreading them further apart in the structure, would lower this contribution to difficulty — this is a structural heuristic, not a guarantee.

An epoxide (C1CO1) exercises functional_group_liability on its own — it wraps chematic's Brenk et al. (2008) structural-alert set directly:

Verdict: LikelyAccessible
Synthesizability: 0.87
Confidence: 1.00
Dominant penalties:
1. Reactive/unstable functional group detected: epoxide (Brenk et al. 2008 structural alert).

A 9-membered ring (C1CCCCCCCC1) exercises the macrocycle branch of ring_topology, and its own simplification suggestion:

Verdict: LikelyAccessible
Synthesizability: 0.75
Confidence: 1.00
Dominant penalties:
1. Macrocyclic ring of 9 atoms (at or above the 9-atom macrocycle threshold).
Simplification suggestions (heuristic, not a guarantee):
1. SimplifyMacrocyclicClosure: Macrocyclic ring closure is a direct driver of the ring_topology contribution to difficulty (large-ring closures often need high-dilution or specialized macrocyclization methods). A smaller ring or acyclic analog, if chemically acceptable, would remove this burden — this is a structural heuristic, not a guarantee.

Pentaerythritol tetraacetate (CC(=O)OCC(COC(C)=O)(COC(C)=O)COC(C)=O) has 4 distinct, non-adjacent ester environments and exercises functional_group_liability's "dense functionalization" signal (chematic::chem::identify_functional_groups, Ertl 2017 clustering) on top of its Brenk alerts:

Verdict: LikelyAccessible
Synthesizability: 0.81
Confidence: 1.00
Dominant penalties:
1. 4 distinct functional-group environments (Ertl 2017 clustering), above the 3 threshold — multiple independent reactive/functional regions can compete for reagent selectivity and complicate protecting-group strategy.
2. Reactive/unstable functional group detected: ketone alpha (Brenk et al. 2008 structural alert).
3. Reactive/unstable functional group detected: acetal ketal (Brenk et al. 2008 structural alert).

Alaninate (C[C@@H](N)C(=O)[O-], deprotonated alanine) has a specified stereocenter and a negatively charged atom. Through chematic 0.12, the negative charge triggered a real upstream overflow bug (chematic#267) — yomitoki worked around it by skipping stereo analysis entirely for any negatively charged atom, at the cost of a real confidence penalty and a silently-zeroed stereochemical_burden for every such molecule. chematic 0.13.0 fixed the bug directly (verified: alaninate now returns the identical result its neutral-acid form would); the workaround was removed accordingly, and this molecule now gets full, correct stereo analysis like any other:

Verdict: LikelyAccessible
Synthesizability: 0.86
Confidence: 1.00
Dominant penalties:
1. 1 tetrahedral stereocenter(s) (specified or unspecified) requiring synthetic control.
2. Reactive/unstable functional group detected: primary amine (Brenk et al. 2008 structural alert).
3. Reactive/unstable functional group detected: acetal ketal (Brenk et al. 2008 structural alert).

Configuring a fragment corpus (see below) only changes the fragment_precedent evidence field's contents — since round 21 (option C), overall.difficulty/ overall.synthesizability are computed from ring_topology/size_topology/ stereochemical_burden/functional_group_liability alone and are identical with or without a corpus configured.

Every report also carries a Provenance block (schema version, yomitoki version, chematic version, ruleset version, fragment-corpus model version, config hash) so results are comparable across versions — see docs/architecture.md.

Component status (v0.1)

Component	Status
`input_quality` / applicability	implemented
`ring_topology`	implemented
`size_topology`	implemented
`stereochemical_burden`	implemented (tetrahedral centers only — see Limitations)
`functional_group_liability`	implemented (reactive/unstable groups + dense functionalization — see Limitations)

fragment_precedent is implemented and opt-in (None unless AnalysisConfig.fragment_model has a corpus configured), but is not a ComponentScores field — since round 21 it doesn't contribute to overall.difficulty, so it lives in its own top-level SynthesizabilityReport.fragment_precedent: Option<FragmentPrecedentEvidence> field: a corpus-relative percentile signal reported as explanatory evidence only, see Limitations.

Components stay None in ComponentScores/fragment_precedent when genuinely not evaluated, never as fabricated zero scores.

suggestions: Vec<SimplificationSuggestion> is populated for 3 of its 6 possible codes (ReplaceBridgedRingWithMonocyclicAnalog, SimplifyMacrocyclicClosure, ReduceStereocenterDensity) — see Limitations. IncreaseFragmentPrecedent is retired (unreachable since round 21 — kept in the enum for schema stability, never emitted, since "this would lower this contribution to difficulty" is no longer a true statement). Every suggestion is diagnostic-only, heuristic, and never claims certainty (expected_effect is always MayReduceDifficulty, never LikelyReducesDifficulty).

How this differs from existing tools

SAscore returns fragment-frequency and complexity penalties as a single number. yomitoki returns component-wise diagnostics, confidence, applicability, evidence, and simplification suggestions.
SYBA is an easy/hard classifier. yomitoki is a diagnostic and explanation tool.
SCScore is a learned synthetic-complexity score. yomitoki decomposes transparent, chemically-named factors instead.
RAscore approximates retrosynthesis success. yomitoki is route-free and explains the structural reasons behind its assessment.
BR-SAScore retrains SAscore's fragment table on USPTO reaction/eMolecules building-block data and reports a single reaction/building-block-informed score. yomitoki returns five independent structural components plus applicability, confidence, and machine-readable findings/simplification-suggestion codes — a structured diagnostic report, not a single number, and requires no reaction corpus to produce overall.difficulty at all (corpus choice is contractually guaranteed not to change it — see External benchmark below for how the two actually compare on accuracy).
AiZynthFinder, ASKCOS, RENKIN are route planners. yomitoki never generates a route.

Comparison with SAscore

A minimum in-process comparison against chematic::chem::sa_score (Ertl & Schuffenhauer 2009) — the completion criterion in AGENTS.md §27. Not a calibration or accuracy claim: the two scores aren't fit against each other and measure different things (SAscore: fragment frequency + complexity penalty; yomitoki: structural burden, decomposed by component). Scales run in opposite directions and aren't rescaled onto a shared axis here — SAscore is 1 (easy) to 10 (hard); yomitoki's difficulty is 0.0 to 1.0.

Real output of cargo run --example sa_score_comparison:

molecule                                sa_score      yomitoki_diff  verdict
ethanol                                     3.45               0.01  LikelyAccessible
benzene                                     2.40               0.10  LikelyAccessible
norbornane (bridged)                        8.20               0.34  ModeratelyAccessible
stereocenter-dense fragment                 8.32               0.31  ModeratelyAccessible
epoxide                                     6.23               0.13  LikelyAccessible
aspirin                                     4.67               0.27  ModeratelyAccessible
paracetamol                                 4.56               0.24  LikelyAccessible
caffeine (fused heterocycle)                4.94               0.29  ModeratelyAccessible
acyl halide                                 6.62               0.09  LikelyAccessible
cyclopropane (strained)                     3.94               0.13  LikelyAccessible
nitrile                                     4.97               0.04  LikelyAccessible
alanine (specified stereocenter)            3.55               0.14  LikelyAccessible
bridged ring + several stereocenters       10.00               0.72  Challenging
spiro ring system                           5.52               0.20  LikelyAccessible

The interesting rows are where the two diverge, not where they agree — a divergence isn't automatically a yomitoki bug. Acyl halide is the sharpest case: SAscore rates it 6.62 (fragment-uncommon), yomitoki rates it 0.09 (LikelyAccessible) — a small, cheap, extremely common acylating reagent with essentially no structural burden by yomitoki's own model. Aspirin (4.67 vs. 0.27) is the same shape as the Brenk-validity gap already described in Limitations. Where they broadly agree (caffeine, spiro, bridged-plus-stereo), that's not evidence either one is "correct" — neither has been validated against real synthesis outcomes yet.

yomitoki_diff for "stereocenter-dense fragment" and "bridged ring + several stereocenters" moved (0.27→0.31, 0.69→0.72) with the chematic 0.13.0 upgrade — a separate stereocenter-counting bug fix from the negatively-charged-atom fix below (an implicit-hydrogen rank -0 sentinel could collide with a real atom's normalized rank 0, silently undercounting stereocenters at certain positions). Both numbers are now higher, correctly, not lower — the old counts were an undercount, not an overcount.

External benchmark (v0.1.0)

yomitoki v0.1.0's frozen default was measured against BR-SAScore's own TS1/TS2/TS3 test sets, alongside SAscore and BR-SAScore itself, on the exact same molecules. Full methodology, per-molecule results, and honest limitations: docs/benchmark.md.

Stated plainly, not rounded up: yomitoki is competitive with BR-SAScore on TS1 (ROC-AUC 0.952 vs. 0.983; balanced accuracy and MCC at yomitoki's own threshold actually exceed SAscore's), has no discriminative power on TS2 (ROC-AUC 0.476 — chance level, diagnosed as a genuine structural finding: TS2's easy/hard classes are ring/size/stereo/functional-group homogeneous under yomitoki's model), and is weaker than both competitors on TS3 (0.673 vs. 0.839 / 0.905). The confidence-based selective-prediction evaluation this benchmark set out to validate as a differentiator did not confirm that story — on TS1, higher-confidence predictions were measurably less accurate than lower-confidence ones, traced to overall.confidence acting as a proxy for dataset provenance (stereo-tag completeness) rather than for prediction correctness. All of this is reported in docs/benchmark.md because it's true, not because it's flattering — the same document states what would need to change before these numbers improve, and yomitoki's per-round development process treats these results as confirmatory, not as something to retune against.

TS2's chance-level result was checked against a second, unrelated dataset — and it reproduced. benchmarks/synthesizability/DEVELOPMENT_SET.md validates the frozen baseline against MPScore, a published expert-chemist-labeled dataset (three chemists' independent easy/difficult ratings, methodologically unrelated to TS1/2/3's retrosynthesis-planner-derived labels, ~0.03% molecule overlap): ROC-AUC 0.513 on the full set, still inside the 95% CI for chance. Two independent ground-truth sources — one algorithmic, one human — now agree that yomitoki's four structural components miss a large share of what makes a molecule hard to synthesize in practice (72.6% false negatives on that set). That document also runs a controlled, unlabeled ablation panel isolating which structural axes the components respond to (and where they saturate or respond backwards), and records four evidence-based design-change candidates without implementing any of them yet — development-only, kept deliberately separate from the TS1/2/3 confirmatory numbers above.

Limitations

All six planned components are implemented (see table above), but fragment_precedent never contributes to overall.difficulty/ overall.synthesizability — since round 21 (option C), configuring a corpus (AnalysisConfig.fragment_model; no corpus ships with yomitoki itself, AGENTS.md §5.4) changes what fragment_precedent reports, not what overall.difficulty measures. overall.difficulty always reflects exactly ring_topology/size_topology/stereochemical_burden/ functional_group_liability, corpus configured or not.
size_topology's rotatable-bond term over-penalizes simple, commercially available long unbranched chains (many rotatable bonds, essentially no synthetic difficulty) — e.g. dodecane's overall.difficulty is 0.068, aspirin's (tripping several Brenk alerts in functional_group_liability) is 0.273, ModeratelyAccessible, despite being one of the most trivially synthesizable molecules there is. Through round 20, fragment_precedent corrected this once a corpus was configured (dodecane → 0.000, aspirin → 0.095); round 21 removed that correction (see the next point for why), so this over-penalization is now an acknowledged, currently-unaddressed limitation regardless of corpus configuration. fragment_precedent still reports (as explanatory evidence, not a score adjustment) that such fragments are strongly precedented, so a report reader can see the mismatch even though the score no longer reflects it.
fragment_precedent is an explanatory reference-corpus signal, not a direct synthetic-difficulty term — this is the public contract, not just a caveat. It used to feed overall.difficulty (rounds 17–20); round 20's cross-corpus validation found that unsafe: configured against ChEMBL vs. two different real, honestly-labeled synthesis-focused corpora (Open Reaction Database, SynRXN), several structurally -legitimate molecules (caffeine, norbornane, spiro/stereocenter-dense systems) scored harder under one corpus and not under another, with no way to predict which in advance — worst case, plain pyridine (no plausible synthetic-difficulty story at all) swung between LikelyAccessible and HighlyChallenging purely from which corpus was configured, driven entirely by this component's own uncapped penalty term. Round 21 (option C) removed fragment_precedent from overall.difficulty entirely — the signal is still computed exactly the same way and still fully reported (SynthesizabilityReport.fragment_precedent), but it can no longer change a score, a verdict, dominant_penalties, or dominant_supports. Verified end-to-end: overall.difficulty is now bit-for-bit identical regardless of which corpus (ChEMBL/ORD/SynRXN/none) is configured, for every molecule tested. See rules.rs's "Fragment precedent" section for the full round 16–21 history and reasoning.
stereochemical_burden only covers tetrahedral stereocenter count and density. Investigated and still not implemented, each for a different reason (see docs/architecture.md for the full evidence):
- E/Z double-bond stereo — chematic can assign E/Z directly from SMILES //\ bond markers (no 2D coordinates needed, correcting an earlier, inaccurate note here), but only for bonds the input SMILES already marked. There's no detector for a stereogenic-but-unspecified double bond the way tetrahedral centers have one, so a specified-only count would measure how carefully the SMILES was written, not how many E/Z centers exist — the same class of problem atropisomerism was rejected for below, arriving a different way.
- Atropisomerism — chematic's detect_atropisomers was tested directly and disqualified: it rates unsubstituted biphenyl as an atropisomer when written c1ccccc1-c2ccccc2 but not when written c1ccccc1c2ccccc2 (same molecule), and rates para-substituted biphenyl identically to genuinely hindered ortho-substituted biphenyl. Wrapping it would violate yomitoki's own atom-order/representation invariance guarantee.
- Contiguous stereocenter runs, quaternary-carbon adjacency — both need an atom-level list of stereocenter candidates (specified and unspecified), which chematic doesn't expose (only aggregate counts). Building that inside yomitoki would make this crate the owner of stereocenter perception instead of a consumer of a validated chematic primitive — a line every implemented component has stayed on the other side of so far.
- Meso compound detection — needs graph automorphism / topological symmetry classes; chematic has this internally (chematic-smiles::canonical_automorphism) but doesn't expose it.
functional_group_liability covers reactive/unstable functional groups (chematic's Brenk et al. 2008 structural-alert set directly) and dense functionalization (distinct functional-group cluster count, via chematic's Ertl 2017 identify_functional_groups). Mutually incompatible functional-group combinations and protecting-group pressure are not implemented — unlike the two liabilities above, neither has a citable, validated primitive to build on (chematic exposes none), and hand-curating either would be exactly the "chemically weak rules, over-generalized" AGENTS.md warns against. Chemoselectivity burden, polyfunctional symmetry breaking, and difficult oxidation-state combinations are also not implemented — chematic exposes no oxidation-state API to build the last one on. Brenk's set was validated as a med-chem screening-library desirability filter, not a synthetic-difficulty signal, so several of its alerts fire on common, cheaply-precedented groups — aspirin, for example, trips four Brenk alerts and lands at ModeratelyAccessible despite being trivially synthesizable, the same score regardless of corpus configuration since round 21. This is a known gap with the same shape as the rotatable-bond one above — through round 20, fragment_precedent corrected it once a corpus was configured (aspirin 0.273 → 0.095); round 21 removed that correction (see the fragment_precedent contract point above) rather than continue tuning a signal round 20 found too corpus-sensitive to trust for scoring. Dense functionalization has its own known gap: it counts topologically disconnected functional-group clusters, so a single densely interconnected polyfunctional system (e.g. glucose's ring of hydroxyls, or a fused β-lactam) collapses to one cluster — identical in count to a molecule with a single, ordinary functional group.
No fragment corpus ships with yomitoki (AGENTS.md §5.4), so out of the box, novel/rare substructures are not detected — fragment_precedent stays None until you build one (tools/build-fragment-corpus) and configure it. No decision has been made yet about shipping one by default (the yomitoki-core/yomitoki-models/yomitoki-data split §5.4 sketches, or a feature-flagged external file).
Simplification suggestions cover 3 of SuggestionCode's 6 variants (bridged ring, macrocycle, stereocenter density) — the other 3 are unreachable, each for a different reason: IncreaseFragmentPrecedent was retired in round 21 (it can no longer be truthfully described as reducing difficulty); quaternary-carbon adjacency isn't computed anywhere; and brenk_matches_detailed unions atoms per pattern rather than per occurrence (so "remove one of several similar reactive groups" can't identify which occurrence to point at). Every suggestion's confidence is a flat, named constant (0.5), not per-suggestion-code, since none are calibrated against real synthesis outcomes.
ApplicabilityReport.domain_distance is always None until a calibration corpus exists (Phase 2+).
Coverage is limited to a curated organic-element subset — no attempt at full periodic-table or organometallic support.
Scores and thresholds are rule-based, not fit to any labeled dataset. An external benchmark now exists (see External benchmark above) with mixed results — competitive with BR-SAScore on TS1, chance-level on TS2, weaker than both competitors on TS3 — and no weight/threshold has been changed in response to it (see that section's test-set-integrity rule).

Reproducibility

Given the same input, AnalysisConfig, and yomitoki/chematic/ruleset versions, analyze/analyze_smiles always return the same report — no randomness is used in the core evaluation path.

License

Licensed under either of Apache License, Version 2.0 or MIT license at your option.

Citation

No paper or citable release exists yet.

Roadmap

Remaining planned work: a shipped fragment corpus (fragment_precedent itself is implemented but opt-in — see Limitations), calibration against SAscore/RAscore/route-search outcomes (a minimum comparison against SAscore, not calibration, now exists — see above), and eventually Python bindings.

yomitoki 0.2.0-alpha.1