The molecules

TL;DR (5 bullets a chief of staff would read)

• The pool is smaller and the problem narrower than advertised. COCONUT Oct-2026 = 738,794 records but only 479,690 distinct metric keys (35 % are stereo/tautomer "spellings"); NPAtlas adds 138; enveda-180 (the only timsTOF stratum of train) is Enamine synthetics with 0.028 % COCONUT overlap. Inside the test mass window a typical NP faces a median 35 same-formula skeletons — ranking positional isomers, not stereo or formula, is the bottleneck.
• The grader is almost, not quite, the stored inchikey14. Train has zero stereo and zero label noise; RDKit canonicalisation changes 1.6 % of keys but merges 1,541 groups of train labels (58,954 spectra) and splits 11 skeletons. Re-key train labels, pool and submission with the exact grader function, pinned to rdkit 2026.03.3, offline.
• NP classes have a "timbre", but only some sections play. Sugar losses (162/146/132/176) and a few alkaloid formant ions (tropane 124.1121, desosamine 158.1176) raise the odds of the right substructure 20–370×; terpenoid/steroid water ladders, coumarins, lignans and immonium ions carry no usable signal (LR+ ≤ 9). Use rules as promote-only priors; absence is never evidence (86 % of hexoside [M+H]+ spectra show no 162 loss).
• The timbre was learned on the wrong instrument. Train's 1.16 M timsTOF rows hold 27 (+) and 21 (−) O-hexoside spectra and zero strict oleananes; glycoside/flavonoid/saponin behaviour on timsTOF is met for the first time in the hidden test.
• About half the hidden 400 is reachable in COCONUT; the med-chem analog slice (15–25 %) is unreachable by COCONUT policy (F and sulfonamide entries deleted at import). Gated one-step tailoring enumeration is worth +0.01…+0.04 MRR; ungated it costs score.

New concrete facts (numbered; file + source count; confidence)

1. COCONUT Oct-2026: 738,794 rows / 728,390 full InChIKey / 479,690 IK14 / 97,757 formulas; 17.2 % of records have no stereo layer; 14,320 IK14 have ≥5 stereo variants (max 72). Union with train = 729,389 IK14; NPAtlas 3.0 adds 138 and is CC BY-NC. — np_chemical_space_*.md, 25+ sources. High.
2. In-window density: median 27 COCONUT IK14 at ±5 ppm (p90 193); 22 at ±1 mDa vs 52 at ±20 ppm; sesqui/diterpene formulas 600–1,300. Positional-isomer view (244–459 Da): median 35, mean 76, p90 198 same-formula skeletons; 58 % have ≥25. — both files. High.
3. Train census (4 columns, 382 s): 277,566 SMILES → 275,810 IK14; 0 stereo, 0 label noise, 0 multi-fragment; 1,612 IK14 groups with >1 drawing (1,477 tautomer, 118 charge, 17 isotope). Canonicalize() changes 1.61 % of keys (enveda-180 0.5 %, riken 7.7 %, msdial 7.0 %); merges 1,541 label groups; splits 11. — beyond_2d_*.md, 60-pair table in .json. High.
4. Tautomer merging: 3-hydroxyflavone keto/enol, curcumin, tetracycline, usnic acid, anthrone/anthranol, naphthoquinone, imine/enamine merge only via RDKit; pyridones, purines, tropolones already merge in standard InChI; chalcone/flavanone is not a tautomer pair. Bit-identical on 2026.03.3 and 2026.03.6. High.
5. Enharmonic redundancy (COCONUT): steroids 54 %, macrolides 59 %, glycosides 47 %; C21H20O11 449→142 keys, rutin 495→157, C30H48O3 890→385, quercetin C15H10O7 61→59. High.
6. Fragmentation likelihood ratios (142,728 stratified train spectra, ±5 mDa, ≥5 %): hexose NL 162.0528 LR+ 46; deoxyhexose 40; pentose 22; glucuronide 176.0321 ~270; C-glycoside 120.0423 66–76; tropane 124.1121 42; desosamine 158.1176 369; indole 144.0808 9.7; aryl-OMe CH3• loss 35 (−); water ladders ≤ 2.6. 55-row rules table + 19 SMARTS reranking rules. — class_specific_*.md/.json, ~40 sources. Medium-high (SMARTS-dependent).
7. Chords do not add: a sugar halves the flavonoid RDA rate (0.13→0.08); 162-loss sensitivity 0.33 (400–500 Da) → 0.02 (800–1,200 Da); >50 eV 0.036; M−162 is base peak in only 9 % (+) of hexosides. The 14 % overall rate is a RIKEN artefact (64 % of rows, 7 % firing; GNPS/MassBank 29–38 %). High (computed).
8. Published per-class accuracy: CASMI 2022 best class 69 % vs best 2D structure 38 %; MassSpecGym MIST hit@1 14.6 %, no per-class table; ViT flavonoid-glycoside position classifier 84.7 % (F1 8-C 0.72). Negative: no tool publishes per-NPClassifier-pathway structure accuracy. High that it is absent.
9. NPClassifier 7/70/672 classes, beats ClassyFire on 47/62; LOTUS 39 % terpenoids (46 % of plant structures); NPAtlas bacteria 34 % PK / 27 % peptides; 10.8 % of NPAtlas pathway-hybrid; terpenoids median 8 stereocentres. COCONUT alkaloid pathway mean NP-likeness 0.07 (bulk-import dilution). — np_classes_*.md. High.
10. Tailoring-delta density: 95.7 % of COCONUT IK14 have a one-step relative (+CH2 84.1 %, −H2 80.2 %, +O 79.8 %, +C2H2O 78.0 %, +C5H8 68.9 %, +hexose 33.7 %); 247/250 enveda-np-examples have one. High.
11. Enveda: timsTOF Pro/Pro 2, 2D-LC mass-targeted isolation; analog patents on withaferin A, isoandrographolide, nootkatone with F/heteroaryl decoration; PRISM has no paper; enveda-180 = Enamine HLL-200 at 20/40/60 eV; test chemistry 0.4 % F vs enveda-180 23 % F. — enveda_platform_*.md, 18 primary items + 24 patent families. Medium (hidden-set model, §6).

Forced analogies — what the cross-domain lens produced

• Timbre = overtone series (class_specific): partially paid off. Neutral losses act as pitch-invariant overtone ratios, diagnostic ions as formants. Yielded a calibrated log-LR prior and the rule that LR weights must be precursor-mass-dependent. Breaks on intensity additivity (sugars mute the aglycone), on the silent note (the lost sugar is never heard in positive mode), on charge-site dependence. Duds: water/C5H8 ladders (terpenoids are the mute family), the 152.0106 radical RDA ion (LR+ 1–2), immonium "timbre" for peptides (use residue spacing instead).
• Enharmonic spelling vs semitone (beyond_2d): paid off. One slot per metric key, never two spellings; semitone neighbours = median 35 per hit; MS2 hears semitones only by intensity ratios, and the test's multi-CE merge erases the energy-breakdown dimension. Breaks where the grader mis-hears spellings itself (11 split skeletons).
• Transposition (chemical_space): mostly paid off (95.7 % one-step density); breaks on regiochemistry combinatorics (6 OH → 6 isomers per delta, 10–100× dilution).
• Variations on a theme (enveda): paid off for shift-tolerant analog propagation (+0.19 MRR in a public write-up); breaks for med-chem decorations that move the protonation site and are excluded from COCONUT.
• Instrument families (np_classes): partial — ladders audible for peptides/polyketides/glycosides, silent for terpenoids; hybrids are two sections on one chord (multi-label DAG, never softmax).

Contradictions within my subdomain

• CANOPUS "99.7 %" vs CASMI 2022 best class 69 % on real unknowns: the former is a sparsity artefact (MCC for 782 of 2,497 classes). Trust 69 %.
• Hexose-loss sensitivity 14 % overall vs 65 % (apigenin-7-O-glucoside) vs 29–38 % (GNPS/MassBank): a library/CE artefact; "typical intensity" is a library property.
• Chalcone/flavanone: np_classes treated them as near-tautomeric; beyond_2d proved two keys. The latter is correct.

Gotchas for the modeler

1. Train labels are not the metric: relabel with K(s) before CV grouping or overlap statistics are off by ~2 %.
2. Fingerprints on canonical tautomers differ from those on database drawings (Δ4-3-keto steroids 41 %, naphthoquinones 41 %, ansa-macrolides 79 % re-drawn). One convention for train and candidates.
3. Hydroxycinnamoyl esters mimic sugars at 15–21 mDa (caffeoyl 162.0317, coumaroyl 146.0368, feruloyl 176.0473): tolerance ≤10 mDa, check the complementary acylium ion.
4. Shared formants: 120.0808 (Phe immonium = necine), 138.0913 (scopolamine = heliotrine), 137.0597 (benzylisoquinoline = guaiacyl).
5. Double counting: a fingerprint model already encodes the 162→hexose bit; adding the rule over-sharpens (q5: combiners stop helping at ρ ≈ 0.6). Apply rules mainly to enumerated/de-novo candidates.
6. The timsTOF stratum of train is the wrong filler set (Enamine synthetics); instrument-matched NP validation exists only via the 250 enveda-np-examples and 48 hexoside spectra.
7. Canonicalisation cost 17.5 ms mean, >0.4 s p99, 4,847 CPU-s for train; never inside the scoring loop.
8. COCONUT's 1,761-entry C6H5N3O bucket and alkaloid NP-likeness 0.07 are curation artefacts; demote by NP-likeness, do not filter.
9. Even-electron fragmenters under-generate CH3• losses (LR+ 35, negative mode); simulator rerankers are biased against polymethoxylated flavonoids, isoquinolines, lignans.

Questions for other queens

• q6_candidate_space_retrieval: emit per-candidate NPClassifier pathway, is_glycoside, NP-likeness and the metric key with the window, so the reranker needs no SMARTS pass; adopt K(s), not raw IK14.
• q1_instrument_physics: expected firing rate of a 162 loss for a 450-Da flavonoid glucoside on a timsTOF multi-CE merge vs a single 40 eV Orbitrap spectrum? We have 27 spectra to calibrate on.
• q5_methods_lineage: does any forward model (ICEBERG, FraGNNet, GrAFF-MS) reproduce the 0.33→0.02 mass-dependence of sugar-loss sensitivity? If not, forward-model rerankers over-penalise large glycosides.
• q2_data_forensics: confirm the 11 split skeletons (beyond_2d .json) include no enveda-np-examples structure.
• q8_serendipity: can the "crib" recalibration estimate per-class LR shifts on timsTOF, not just ppm/CE?

Layman sentences

1. Nature's molecules come in families, and like instruments in an orchestra some families (sugar-bearing molecules, a few alkaloids) have an unmistakable sound in the mass spectrometer while the biggest family, the terpenes, is nearly mute.
2. The scorer is tone-deaf to mirror images and some proton shuffles, so two answers that differ only that way count as one — never spend two of your 25 guesses on them.
3. The hard part is not the family or the formula but the dozens of same-formula cousins differing only in where one group sits — and the instrument hears that difference only as loudness, not pitch.

Gaps I could not fill

• No published per-NPClassifier-pathway structure accuracy exists; per-class difficulty is inferred from pool density and must be measured on our own stratified hold-out.
• Demarque 2016, Cuyckens & Claeys 2004, CANOPUS/MIST bodies paywalled; CANOPUS per-class MCC from snapshots (low confidence). paperclip not installed here.
• LOTUS not recomputed (Zenodo 403, SPARQL timeout); its COCONUT footprint (146,067 IK14) used instead.
• THEOBROMA stereo-family distribution and TIMS-Bench MSV000101619 compound list unverified.
• Strict oleanane SMARTS matched nothing in train, so triterpene saponin rules have no empirical LR.
• Hidden-set composition (four sub-populations) is a reasoned model, not a measurement.