Expert synthesis

What the whole hive concluded, written for the person who starts building tomorrow. Below it: the sixty numbers that drive decisions, each with the file and source it came from, and a glossary of every term the narrative defines.

Scope: everything below is sourced from the two king logs, the eight queen digests, the Emperor's tick-1 log and the train EDA; worker files were opened only for research/q7_compute_execution/ninety_day_plan_milestones_risks_and_kill_criteria.md (plan dates), research/q4_challenge_intelligence/casmi_history_2012_2022_winners_methods_lessons.md (2022 figures) and one line of research/q3_natural_product_chemistry/enveda_platform_isolation_profiling_what_analogs_means.md. Where queens or kings disagree, both numbers are given. Confidence tags are the queens' own.

1. The problem, formally

Inputs. Per hidden molecule_id: 1–16 spectra (median 3), each a raw-integer-count centroid peak list (ms2_mzs, ms2_normalized_intensities with base peak = 1.0, base_peak_intensity in counts), precursor_mz, adduct (7 strings in the visible file; q1/q4 count 10 in official texts — unresolved), ionization_mode, instrument_type = timsTOF always, collision_energy_ev as a list ([20], [40], [60] or [20,40,60] = merged). No MS1, no isotope pattern, no retention time, no ion mobility. Test precursor range in the visible file is 245–460 m/z, but that is the Enamine HLL-200 mass band of the decoy (99.67 % of enveda-180 monomers), not an organiser filter; 42 % of enveda-np-examples lie outside it (q2, high).

Outputs. Up to 25 semicolon-separated SMILES per molecule_id, best first.

Metric. MRR@25 = mean over molecules of 1/rank of the first correct guess, 0 if none in 25. Correctness: MolFromSmiles → rdMolStandardize.TautomerEnumerator().Canonicalize → MolToInchiKey[:14], RDKit 2026.03.3 pinned via the metric/rdkit-2026-3-3-wheel dataset; a heavy-atom-composition gate precedes canonicalisation; comma-separated multiple acceptable truths are supported; no salt stripping ([Na+].CC(=O)[O-] ≠ acetic acid). >25 tokens, NaN or a duplicate molecule_id rejects the whole file; invalid or duplicate SMILES silently burn slots; missing ids score 0 (q4 high; q2 medium — official notebook reconstructed from three write-ups, 2026.03.3 ≡ 2026.03.6 on all local tests).

What the metric cannot see. Stereochemistry (all of it), tautomers that RDKit merges (3-hydroxyflavone keto/enol, curcumin, tetracycline, usnic acid, anthrone/anthranol, naphthoquinone, imine/enamine — but not chalcone/flavanone, two keys), charge and isotope labels. Consequences measured by q3/q2: canonicalisation changes 1.61 % of train keys, merges 1,541 train label groups (58,954 spectra), splits 11 skeletons, changes 12–17 % of NP keys and 2.3 % of COCONUT keys (one team lost 0.105 on that subset), is stereo-dependent for 0.45 % and non-idempotent for 0.3 %. 35 % of COCONUT rows are stereo/tautomer "spellings" of another row (steroids 54 %, macrolides 59 %, glycosides 47 %). Canonicalize() hits maxTransforms=1000 on polyphenols (quercetin 363 tautomers / 179 ms, rutin 357 ms) with no warning, but raising the cap to 20,000 changed zero of 8 polyphenol IK14s; the residual exposure is de-novo SMILES in odd tautomeric forms (q7 high). Rule: one slot per canonical key; canonicalise unique structures once, offline, never in the scoring loop (17.5 ms mean, >0.4 s p99).

Noise on the metric. Per-molecule reciprocal rank at mean ≈ 0.4 has SD ≈ 0.45 → SE ≈ 0.022 on 400 molecules, ≈ 0.04 on the ~130-molecule public subset (public fraction ≈ 33 % from one kernel, low confidence). Byte-identical reruns moved up to 0.105 (q4). Random-ranker baseline = H₂₅/n = 3.816/n (q7).

2. What is actually measured

Instrument chain (reconstructed from data; nothing published). Bruker timsTOF Pro 2, ESI, PASEF; quadrupole isolation 2 Th below m/z 700 and 3 Th above (Meier 2018); CID at fixed lab-frame 20/40/60 eV — not a 1/K0 ramp (q1, high; 186,392 train (IK14, adduct) groups have all four patterns). "Merged" rows are the per-m/z raw-count sum of the three scans, re-centroided (reconstructed on 595/600 quartets; effective weights ≈ 0.5/0.4/0.1; merged median 258 peaks vs 88–132 singles); 26.6 % of enveda-180 and 24.6 % of test rows are merged; 68.5 % of test molecules have one. enveda-np-examples instead use ";"-separated lists and 35/50/80 eV (q2).

Where information is created. Precursor survival for [M+H]+ at 250–450 Da: ~11–13 % of TIC at 20 eV, absent in 63–72 % of spectra at 40 eV and 99 % at 60 eV. Same-molecule 20 vs 60 eV sqrt-cosine is only 0.285 — the three energies are close to independent views, and the merge erases the energy-breakdown dimension (q1, q3).

Where information is destroyed, and the artefacts that replace it. (i) Isolation-window leakage: precursor peak in 45.9 % of timsTOF spectra, M+1 in 38.7 % (GNPS-Orbitrap 6.1 %), M+2 in 21.7 % (0.7 %); M+1/M p50 = 0.18 → near-full transmission inside 2 Th. (ii) 13C satellites at +1.0034 Da on 71.1 % of peaks ≥ 5 % (control 3.0 %; GNPS-Orbitrap 14.0 %). (iii) Raw integer centroids, floor exactly 9 counts, no cap (p95 823 peaks): 91.8 % of peaks < 1 % of base, 75.1 % < 0.1 %, 39.8 % < 0.01 % (GNPS-Orbitrap 60/16/5 %). (iv) Peak count tracks abundance: Spearman(bpi, n_peaks) = +0.71; weakest tercile median 30 peaks, strongest 332. (v) Low-mass fragment bias: C5–C9 fragments read +0.3–0.6 mDa (+4 to +6.5 ppm) high, gone by m/z 180; sub-m/z-40 peaks are non-recurring noise (mechanism unknown). (vi) 2.44 % of peaks lie above the precursor, all ≤ +2.12 Da in enveda-180, ≤ +2.00 Da in np-examples (q1/q2 — which trim the hidden set uses is open). (vii) Water loss in 17.7 % of enveda-180 and 29.4 % of np-examples [M+H]+ spectra vs 10–15 % in public libraries.

Collision-energy currency. Train's collision_energy_ev for Orbitrap rows is NCE·m/z/500·f(z) (f = 1/0.9/0.85; 99.8 % reproducible). On 2,421 same-molecule timsTOF/Orbitrap pairs, timsTOF 20 eV ↔ NCE 20 (sqrt-cosine 0.82), 40 ↔ 45 (0.70), 60 ↔ 60 (0.61), merged ↔ NCE 30–45; i.e. the parquet's eV over-states NCE ~1.4× and NCE-conditioned models should be driven at ~0.7× the formula value (q1, medium — own measurement). QTOF→Orbitrap shift in the literature is ~4–4.7 eV; no published small-molecule timsTOF-vs-Orbitrap MS2 study exists.

Mass accuracy. Recomputed (signed, electron mass included) timsTOF precursor error: enveda-180 |ppm| ≤ 5.13 (99.98 % within ±5, 98.1 % within ±4), median 0.98 ppm, with a +1.9 ppm positive-mode bias of unknown cause; np-examples [M+H]+ p90 3.38 / max 8.7 ppm, [M−H]− max 21.1 ppm. Errors are correlated within a run. ±5 vs ±10 ppm windows were LB-neutral in one write-up (0.302/0.303). Fragment tolerance: max(10 ppm, 5 mDa) plus a −0.4 mDa correction below m/z 150; a bare symmetric ±5 ppm sub-formula window rejects about half of true C5–C9 fragments.

Noise model to score against (q1 GT worker). SIRIUS-style Gaussian ppm + Pareto intensity noise, extended with CE mixing weights, a per-spectrum ppm offset, 13C satellite pairing and a low-mass cut-off; specified as a log-odds scorer to A/B against entropy similarity. The five-step "telluric mask": delete m/z > precursor − 0.02; merge +1.00335 ± 0.005 satellites with 0.02–0.5 ratio into the parent; cut < 20 counts; subtract 0.4 mDa below m/z 150; optional frequency weighting (sign to be tested). matchms default intensity_from=0.001 deletes 75 % of a timsTOF spectrum; MassSpecGym's ">300 peaks" rule discards ~25 % of enveda-180.

Adducts. Five of ten test adduct strings are exactly isobaric with [M'+H]+/[M'−H]− of another formula; Na−H vs C2−2H differ by 2.4 mDa; build the window around the molecule-level consensus neutral mass across its 1–16 spectra, never per spectrum; admit H↔Na-swapped (Δ 21.981944 Da) and NH4/NH3-ambiguous neutral masses at low prior in slots ~20–25 (q1/q6). Adduct intensity ratios are kinetic; no equilibrium constant exists (Saha analogy died).

3. The data, forensically

Libraries (research/_meta/train_eda.md): 2,539,608 spectra / 275,810 IK14 / 51,291 formulas / 82 instrument labels / 121 adduct strings. timsTOF = 1,154,969 rows = enveda-180 1,153,785 (182,941 structures; Enamine HLL-200 on timsTOF Pro 2; Zenodo 10.5281/zenodo.20436851, 29 May 2026; 1,158,293 spectra / 184,330 compounds at source, train has 4,508 fewer) + enveda-np-examples 1,184 (250 structures; Kaggle page says 1,151). Public libraries: pluskal_ms2 527,581 / 46,821; riken 347,171 / 15,892; gnps 220,849 / 45,750; massbank 101,727 / 9,180; mona 92,416 / 11,681; spectraverse 50,933 / 9,631; msdial 40,765 / 9,127; drug_plus 2,545 / 2,539; masaryk 652 / 416. Median peaks per spectrum range from 9 (riken, massbank, masaryk) to 142 (np-examples). 64.6 % of train precursors fall in 245–460 m/z.

Biases (library barcodes). Intensity floors: MassBank 1/999, MS-DIAL 0.01, RIKEN 0.01/0.005, pluskal ~7e-5, timsTOF none; base_peak_intensity null in exactly the 7 public libraries; GNPS/drug_plus/masaryk 100 % null CE; seven of 11 libraries ship a computed precursor m/z (pluskal 99.3 %, msdial 96.5 %, drug_plus 93 %, massbank 76 %, mona 69 %, spectraverse 60 % equal theory). The shipped precursor_error_ppm is |error| without electron mass, computed before a formate relabel: 52,460 rows show ~2,400 ppm; r = −0.02 with the true signed error; only np-examples is signed and correct. 62,834 rows lie beyond 20 ppm uncleaned (mostly pluskal negative-mode formate rows). Polarity contradicts adduct sign in 6,087 rows. A corrupt precursor_mz of 2,430,333 exists. 1,612 IK14 groups have >1 SMILES (1,477 tautomer, 118 charge, 17 isotope); 0 stereo, 0 label noise, 0 multi-fragment SMILES.

Adduct mix. enveda-180: 51 % [M+H]+, 16 % [2M+Na]+, 8 % [2M+H]+ (29 % dimers overall; 282,074 rows); test: 79 % [M+H]+, 16 % [M−H]−, no dimers. Dimers dissociate to monomers at 20 eV, so all-adduct survival statistics are wrong; 1,942 two-energy merges (median 5–9 peaks) are junk. If the hidden set was thinned like the decoy, negative mode is ~1.3× under-represented.

Decoy test. 1,213/1,213 visible rows join uniquely to enveda-180 on (precursor_mz, num_peaks, base_peak, adduct); 400 ids → 400 IK14; 395 exist only in enveda-180. Structure selection is random (within the 95 % band of 2,000 random draws); row selection is not: 0/664 dimer rows kept, 1,213/1,818 monomer rows kept (0.71 positive / 0.53 negative; keep-rate rises 0.59→0.71 with base-peak quintile). Decoy chemistry (0/400 in COCONUT, NP_Score −1.47, 36 % halogenated) says nothing about the hidden set. Public kernels detect commit-vs-rerun via SHA256 of test.parquet.

Domain shift. The timsTOF stratum of train is a chemistry decoy: 51/182,941 enveda-180 structures in COCONUT (0.028 %); test chemistry 0.4 % F vs enveda-180 23 % F. Train ∩ COCONUT = 26,111 IK14 (9.47 %); the 9 public libraries hold 95,157 IK14 of which 26,068 are COCONUT — the entire "NP with any train spectrum" universe, none on timsTOF. Train's whole timsTOF stratum contains 27 (+) / 21 (−) O-hexoside spectra and zero strict oleananes; glycoside and saponin behaviour on timsTOF is met for the first time in the hidden set. Only 2,538 structures are measured on both timsTOF and another instrument ("P-inst" set). Library-search sensitivity cross-instrument is 88.0/91.5 % vs 99.5/99.7 % self; 95.1–98.4 % within 20–50 eV vs < 81 % at QqTOF 5–10 eV.

Leakage. All 250 np-examples structures are in other libraries (RIKEN 246, GNPS 233; 0 exclusive) — hold them out by IK14 across every library plus analogs at MCES < 10 / Tanimoto ≥ 0.7. A per-spectrum split puts ~94 % of structures on both sides. Random train splits predicted the opposite ranker order to the LB; decoy-shaped folds gave 0.456 local → 0.112 LB and 0.445 → 0.063. MassSpecGym audit (arXiv 2606.19624): problems in ≥ 17/26 papers; 82.41 → 6.70 % hit@1 when a canonicalisation shortcut was removed; spectrum-blind PubChem order 49.98 % hit@1; formula leak 30×; MIST R@1 67.75 % (simulator saw test) vs 53.76 % clean; padding-mask bug +21 points. Public FPNet weights: 0.94 Class-2 MRR on seen structures vs 0.477 held-out. The organisers co-authored that audit — assume the answer key defeats canonicalisation and PubChem-order shortcuts. Always report a spectrum-blind baseline.

Fold spec that exists (q2): heldout_A = 1,977 structures / 13,556 timsTOF spectra (250 np-examples + 1,727 NP-like enveda-180); heldout_B = 293 strict database NPs. q7 owns the truth-absent hold-out (~1,000 scaffold- and IK14-disjoint NPs, analogs removed, truth deleted from the pool for a random half, test adduct × n-CE mix); the 250-structure three-mask backtest is the primary selector (the only local test that predicted LB order, three teams).

4. Candidate coverage — the ceiling

MRR@25 ≈ c·r·e. Back-solved at the LB-0.417 tier (q6, medium): c (truth in pool) ≈ 0.55–0.61, r (ranked ≤ 25 | in pool) ≈ 0.90, e (mean 1/rank | ranked) ≈ 0.76; headroom c +0.27, e +0.13, r +0.05. Within-hit MRR across ten write-ups is 0.75–0.95; the within-list hit curve fits k^−2.07 (64 % of hits at rank 1, 92 % in top 5); slots 1–5 carry ≥ 93 % of expected value.

Class mixture of the hidden set (two write-ups, open): Class 1 (reference spectrum exists) 16 % at channel MRR 0.917; Class 2 (structure in some DB, no spectrum) 45 % at 0.465; Class 3 (in no database) 39 % at 0 → single-pool ceiling ≈ 0.36 — versus 16/27/55, which moves the retrieval ceiling from ≈ 0.61 to ≈ 0.45 (q4).

The map: COCONUT Oct-2026 738,794 rows → 728,390 full InChIKeys → 479,690 IK14 (q6 and q3 independently); LOTUS v11 +254; NPAtlas 3.0 +138 (CC BY-NC); THEOBROMA 29-source union 486,215; SuperNatural/NPASS/KNApSAcK ~0. 28.5 % of COCONUT has NP-likeness ≤ 0; 38,771 revoked rows and the 1,761-entry C6H5N3O bucket must be dropped. PubChem: 124,717,116 CIDs; a ±10 ppm window there is ~10⁴ candidates. ±10 ppm in IK14-dedup COCONUT: median 36, mean 87, p90 204 (300–400 Da median 64; 400–500 Da 49; 500–600 Da 20); stereo-expanded median 107; ±5 ppm median 27, p90 193. Same-formula skeletons in 244–459 Da: median 35, mean 76, p90 198; 58 % face ≥ 25. Adding train IK14s raises the window median 36 → ~56 with ~zero coverage gain for novel NPs.

Pool-size evidence: MetFrag 338 → 2,508 candidates moved median rank 8 → 31.5; CSI:FingerID 68.5 → 34.4 % top-1 (biodatabase → PubChem); MetFrag fragmentation-only 22 → 6 %; LB: train+COCONUT 0.112 → train+PubChem 0.079 (−30 %; 68 % of top-1 slots stolen); Enamine tier −0.013; +62,744 ChEBI/LIPID MAPS +0.000; gated PubChem top-50 (library confidence < 0.7) +0.011; ungated one-step analog enumeration +0.000. Counter-evidence: CASMI 2016 (PMC5368104) found performance worst at 200–1,000 candidates — confounded by formula, so stratify any pool experiment by formula (open). Prior-density ratio COCONUT : PubChem-only ≈ 1,500–2,000× → a PubChem candidate needs ~7 nats of spectral evidence. Insertion break-evens from Stone's theorem: slot 1 needs P ≥ 0.2, slot 5 ≥ 0.006; appending at 11–25 costs 0.0001–0.0003 MRR per slot.

Ways to raise c: one-step analog enumeration (15 operators) reaches 82.8 % of np-examples but only 38.7 % of random COCONUT (upper bound ~2× optimistic for the hidden set), at ~28× dilution (window pool ~1,015 vs 35); five operators (dehydroxylation, sp3/aromatic hydroxylation, O-methylation, O-hexosylation, O-acetylation) give most of the reach at ~1/5 of the dilution; 95.7 % of COCONUT keys have a one-step tailoring relative; 247/250 np-examples do. Expected gated gain +0.03–0.05 (q6) / +0.01–0.04 (q3), not +0.2. The med-chem analog slice of the hidden set (15–25 %, q3's model) is unreachable by COCONUT policy (F and sulfonamide entries deleted at import). De novo is a Class-3 slot filler only (§5). Formula-first hard filtering is a net loss: the true formula already covers 62–90 % of in-window COCONUT structures (median 2–4 formulas per window), MS2-only formula top-1 is 50–77 % on honest benchmarks; break-even P(formula) ≈ 0.75–0.9 → soft multiplicative feature, hard-gate only on mass and adduct.

The single most load-bearing unmeasured number is π_in — np-examples ∩ COCONUT under RDKit 2026.03.3 (open: 249/250 computed on standard keys vs "8 of 250" in one write-up, almost certainly a keying error; ~50 CPU-h recomputation queued for Explorer).

5. Methods, ranked by evidence for THIS test

Ranked by evidence of transfer to novel NPs on timsTOF, not by paper headline. "Comparable data" = structure-disjoint NP retrieval or the CASMI-26 public LB.

# Method What it needs Best reported number on comparable data Failure mode Cost
1 Spectrum → substructure fingerprint (Morgan r2+r3+MACCS / FPNet) + COCONUT ±10 ppm pool + Tanimoto, LightGBM/HGB LambdaRank reranker, RRF of 2–3 engines, library lock at lib_max ≥ 0.9 train spectra, pool, GPU-hours (25 M-param transformer ≈ 2 min/epoch on H200, ~$10 for 50 epochs) every ≥ 0.37 public kernel; LB 0.42–0.48; single MLP + pool 0.11–0.25; four fingerprint variants within 0.007 LB isomer blindness (hesperidin/neohesperidin indistinguishable; glycoside-position classifier tops at 84.7 %); learns library barcodes and dimer mix; recall ceiling c ≈ 0.55 low–medium
2 Library search (modified cosine / entropy similarity, RRF across a molecule's spectra) train spectra only LB 0.098 / 0.063 alone; Class-1 channel MRR 0.917 useless for the ~84 % without a reference spectrum; cross-instrument sensitivity 88.0/91.5 %; CE mismatch −8–12 points CPU minutes
3 Candidate-set softmax cross-entropy over the real mass-matched pool (MSAlign 2026) as the training objective same as #1 plus pool at train time 47.5 % vs 19.5 % R@1 for the same encoder vs in-batch InfoNCE; fusion weights on dev MRR → 58.2 % unaudited preprint; hard negatives give no gain in two papers; nobody optimises MRR directly low (loss change)
4 Jointly fitted stacker over calibrated log-LR features (LightGBM lambdarank), RRF k=60 as the no-training control, CombMNZ bake-off ≥ 2 engines with an admission gate CASMI 2016 six-method consensus Top-10 166/208 vs 131/208; six-fragmenter fusion top-1 82 vs 70; RRF +4–5 % over best (Cormack 2009); simulated +0.03–0.06 rank-averaging loses to the better engine at ρ ≥ 0.3; independence combiners stop helping at ρ ≈ 0.6; weaker engine halves the consensus gain (+0.038 → +0.018); no MS paper with a LambdaMART ablation could be opened low
5 Gated pool extension: PubChem top-50 when library confidence < 0.7; five-operator one-step analog enumeration when top-1 posterior is weak; posterior-sorted seating calibrated posterior (isotonic), absence detector (AUROC ≥ 0.8, ECE ≤ 0.05) +0.011 (gated PubChem, LB); predicted +0.01–0.03 over quota seating, +0.03–0.05 over single pool (untested) ungated: +0.000 (analogs), −30 % (PubChem); detector degrades under shift; "in-DB?" gate showed no signal (42.8 % vs 59.4 % base rate) low–medium
6 Telluric mask + energy-response map R(fragment m/z, precursor m/z, CE) fitted on 2,421 dual-instrument pairs parquet only specified, not measured; keep if MRR@25 +0.02 on the three-mask backtest and base-peak-tercile curve flattens R is molecule-dependent (CID is kinetic); fragmenters are the weak link CPU < 1 h
7 In-silico fragmenters / forward models (MetFrag, CFM-ID 4, ICEBERG 2, GLACIER) as an orthogonal channel candidates, wall-clock budget MetFrag 24/208 top-1 (CASMI 2016); on 87 never-seen challenges MetFrag+CFM 13 ≈ CSI:IOKR 16 ≈ CSI:FID 13 none models timsTOF CE (CFM-ID fixed 10/20/40 eV; MassFormer NCE only); MS-FINDER ignores intensities; even-electron fragmenters under-generate CH3• losses; fixed budgets starve (18 % coverage at 300 s) MetFrag 0.02–0.1 CPU-s/cand; CFM-ID 5–11 CPU-s/mol; ICEBERG 2 46.4, GLACIER 368.7 spectra/s on 2×A5000
8 Class-specific neutral-loss / diagnostic-ion log-LR priors (19 SMARTS rules; hexose LR+ 46, deoxyhexose 40, pentose 22, glucuronide ~270, tropane 124.1121 LR+ 42, desosamine 158.1176 LR+ 369, aryl-OMe CH3• 35 (−)) SMARTS pass over candidates measured on 142,728 stratified train spectra absence is never evidence (86 % of hexosides show no 162 loss; sensitivity 0.33 → 0.02 with mass); double counting with fingerprint bits (ρ ≈ 0.6) → ≤ +0.01 when a fingerprint model is present; terpenoid ladders LR+ ≤ 2.6; hydroxycinnamoyl esters mimic sugars at 15–21 mDa negligible
9 Formula posterior as a soft feature (MIST-CF / BUDDY / FIDDLE / SIRIUS-style tree) MS2 only (no MS1) MIST-CF 50.5 % top-1 (MassSpecGym MCES), 74–77 % NPLIB1; FIDDLE 90.6 % ≤ 800 Da, 64.2 % > 800 Da; SIRIUS < 10 % error below m/z 300, > 50 % toward 800 hard filter displaces the whole posterior; MIST hit@1 halves (51.45 → 26.64 %) on predicted formulas low
10 MIST (MIT, open) weights, offline packaging 14.64 % hit@1 MassSpecGym (random 0.37 %); 9.57 % on formula-restricted sets padding-mask bug; CE-string parsing → NaN logits for 1,212/2,499 queries in one run medium
11 DreaMS (116 M params, MIT, HF weights) weights (≤ 700 M GNPS spectra pretraining) audited retrieval 6.7 % hit@1 tail filler only in public kernels medium
12 Popularity / reference-count priors, multi-database attestation PubChem/ChemSpider counts MetFrag top-1 30 → 336 → 420 of 473; 143/208 CASMI 2016 at rank 1 from counts alone; LB +0.004…+0.011 in one lineage, monotonically negative in another inverted on a never-published test; rankers learn it implicitly via training-spectrum counts — audit score-vs-train-count correlation; use only on the library-hit class, gated zero
13 De novo generation (DiffMS, FRIGID, DualLGD, MS-GPT, MSNovelist) 1–3 GB weights, 0.93–7 s/spectrum, true formula 0.00 % exact without formula on the structure-disjoint split; with true formula + [M+H]+: DiffMS 2.3 %, FRIGID 18.3 %, DualLGD 23.9 %, MS-GPT 23.9/28.7 % top-1/10 SELFIES 100 % valid, 0 % exact; DiffMS drops 2.30 → 1.86 % on predicted formula; all numbers [M+H]+ only; licences (3DMolMS NC-SA, FRIGID unresolved) high
14 CSI:FingerID / SIRIUS / CANOPUS licensed web service CASMI 2016 70 → 36 top-1 with training overlap removed; 68.5 → 34.4 % bio → PubChem; CANOPUS "99.7 %" is a sparsity artefact, CASMI 2022 best class 69 % unusable offline in a Kaggle rerun n/a
15 PRISM (Enveda/Microsoft, 1.2 B spectra) — relative numbers only (+23.3 % retrieval hits vs unnamed baseline) no weights, code or preprint; nothing to build on n/a

Intensity handling (q8): every optimised exponent since 1994 is 0.25–0.6; Li 2021's entropy-adaptive rule switches the minor-peak boost off above spectral entropy 3; no published ablation removes the base peak; Hart 2024 up-weights common m/z. Cap to top-K ≈ 100–150 peaks and check MRR is flat across base-peak terciles. Strip > precursor − 1.0 Da (the surviving precursor is the base peak in 25–29 % of merged [M+H]+ and inflates cosine for every right-mass candidate).

6. Cross-domain transfers worth testing

From the two king logs; each with its one-line test and kill rule.

  1. Telluric / cosmic-ray mask (astronomy, q1). Run the public FPNet/HGB pipeline on the three-mask np-examples backtest with and without the five-step mask; keep if MRR@25 rises > 0.02 and the base-peak-tercile curve flattens.
  2. Instrument response function (astronomy, q1). Fit R(fragment m/z, precursor m/z, CE) on the 2,421 pairs; scaffold-split MRR@25 on a COCONUT ±10 ppm pool with ICEBERG scores ± R; kill if gain < 0.01.
  3. Custody log of a number (forensics, q2). Recompute signed ppm with electron mass for all 2.54 M rows, re-filter at 20 ppm; the 62,834 "beyond 20 ppm" rows should shrink to true outliers; check whether the +1.9 ppm bias vanishes under one m/z convention.
  4. Standard of proof → posterior-sorted seating (law, q4). Same candidate scores, block seating (13–15 primary then hedges) vs posterior-sorted; keep if ≥ +0.01 on the np-examples backtest; the first alternative-hypothesis candidate belongs at seat 2 unless P(primary) ≥ 0.90.
  5. Daubert known error rate (law, q4). Re-submit a byte-identical notebook; if variance approaches the observed 0.105, all model selection is local; never promote on an unpaired LB delta < 0.04.
  6. Timbre / overtone ratios (music, q3). Add the 19 SMARTS log-LR rules to the reranker; expect ≤ +0.01 with a fingerprint model present, more on enumerated/de-novo candidates; kill if glycoside MRR falls.
  7. Enharmonic spelling (music, q3) — already decisive. Count duplicate canonical keys in any public top submission; each is a dead seat; one slot per metric key.
  8. Stone's optimal search (search-and-rescue, q6). Union of {library, COCONUT ranker, gated 5-operator analogs, gated PubChem top-50} ordered by isotonic-calibrated posterior vs fixed 15/7/3 quota vs COCONUT alone, on the three-mask set + a masked-COCONUT set (MCES ≥ 10 from train); report (c, r, e) per class; kill if posterior ordering is not sign-consistent in ≥ 70 % of class × adduct cells.
  9. Mock-witness test (eyewitness line-ups, q7). Ranker on precursor m/z + formula + adduct only, no MS2; if it beats H₂₅/n the pool is biased and local scores harvest line-up bias. Cheapest experiment in the hive (CPU-hours).
  10. Morelli's minor details (art attribution, q8). Intensity ablation on an isomer-rich hold-out: base peak removed / peaks < 10 % only / sqrt / log / entropy-adaptive; readout = isomer-swap rate and MRR@25; stratify by spectral entropy to locate the Li 2021 boundary; < 1 GPU-day.
  11. Mosteller–Wallace function words (stylometry, q8). z-score each candidate's explained common-loss profile (H2O, CO, CH3) against the corpus norm as one LightGBM feature; keep only if the paired-bootstrap CI excludes 0.
  12. Known-plaintext cribs (cryptanalysis, q8). Fit ppm offset, CE response and score temperature on the ~40–150 confident library hits in the hidden set, apply to the rest; kill if the gain is confined to the library-like tail.
  13. Open-set assignment (eDNA taxonomy, q8). Absence detector (AUROC ≥ 0.8, ECE ≤ 0.05) routes slots 11–25 to a gated analog/PubChem generator; negative control: routing must not lower MRR on the truth-present half.
  14. Class-conditional likelihood ratios (differential diagnosis, q5). Per-bit naive-Bayes rescoring conditioned on NPClassifier class; success ≥ +0.02 Class-2 MRR in glycosides.
  15. P-inst (q5/q1, not a lens but the highest-value measurement). Any learned method on the 2,538 dual-instrument structures, with and without the mask; if the timsTOF penalty collapses, no timsTOF fine-tuning is needed.

Duds kept honestly: sky subtraction (recurring m/z bins are chemistry), Saha equilibrium for adducts, burden shifts / double jeopardy, decoy halogen fraction as a hidden-set hint, Enigma-style deterministic elimination (rules explain 78–85 % of ions → would discard 15–22 % of truths), sequential cascade (killed three ways: simulation, ROC literature, q8), double-blind administration (only pre-registration survived), beam-search curse (moot), hard "diagnosis of exclusion" switching (no in-DB signal).

7. The 90-day plan

Calendar: 68 days from 2026-10-07 to 2026-12-14 23:59 UTC; team-merge deadline 2026-12-07; 5 submissions/day, 2 final picks; ~9 h rerun ceiling, internet off (one GitHub mirror — medium); Kaggle image is Python 3.13 with no RDKit (the rdkit==2026.03.3 cp313 wheel, 37,187,145 B, must be vendored as a dataset); to_pandas() on the list columns costs 25.7 GB in a 29–32 GB session — use pyarrow or a flattened float16 top-100 layout (0.80 GB). Kaggle GPU quota 30 h/week → ~6–10 scored GPU submissions per week. Explorer: public gpu partition 4 h default / 8 h max, 104 pending vs 52 running; gyorilab 4×H200 (30-day MaxTime) fully occupied by 26 d / 11 d / 11 d jobs; /scratch 91 % full; proxy http://10.99.0.130:3128; scp not rsync; CRLF breaks sbatch. Modal A100-40GB $2.10/GPU-h (not the brief's $1.40; q7's plan file quotes A100-80GB ~$2.50/h, H100 ~$3.95/h). All parquet computation on this box serialised; the ~50 CPU-h RDKit canonicalisation job runs on Explorer.

Milestones (q7, week by week): W1 Oct 07–11 data + pool downloads + pinned RDKit in all three environments; W2 Oct 12–18 minimum viable library-search submission by Mon 10-12, scored by Wed 10-14, metric.py reproduces MRR@25, structure- disjoint split by IK14; W3 Oct 19–25 unified pool (train ∪ COCONUT ∪ LOTUS ∪ NPAtlas, canonical-key dedup, adduct table), recall report at ±5/10/20 ppm, target recall@25 ≥ 0.60; W4 Oct 26–Nov 01 spectrum→fingerprint model on Explorer; W5 Nov 02–08 LambdaRank reranker + paired-bootstrap harness; W6 Nov 09–15 stress test (1,600 spectra / 450 molecules ≤ 6 h on Kaggle GPU, OOM guard, schema-robust loader); W7 Nov 16–22 second orthogonal scorer + culprit-absent gate; W8 Nov 23–29 ablations, per-adduct/per-CE behaviour, slot-waste rate = 0; W9 Nov 30–Dec 06 code freeze 12-04, both final candidates scored once without error; W10 Dec 07–14 final two selected by 12-11 23:59 UTC, code packaged for open-sourcing. The Emperor's recommendation inserts two experiments before any model work: the mock-witness test and the np-examples three-mask backtest harness.

Kill criteria (dated 23:59 UTC): K1 10-14 no non-error LB score > 0 → modelling stops; K2 10-25 pool recall@25 on local val < 0.50 → week 4 reassigned to the pool; K3 11-01 any de novo branch < 0.15 local MRR → dropped; K4 11-08 no new architecture may start; K5 11-15 stress rerun > 6 h or OOM → halve per-molecule budgets; K6 Modal spend ≥ $120 by 11-30 → Explorer only; K7 11-22 ensemble < +0.010 over best single with CI excluding 0 → ship the single; K8 12-04 freeze; K9 12-06 12:00 UTC team merged or proceed solo; K10 12-11 final two selected; K11 standing: any claimed gain < 0.010 or CI containing 0 is a null result; K12 from 10-20: local gain with LB drop > 0.04 → investigate val/hidden mismatch. Final-two rule (fixed 10-20): Pick A = highest local structure-disjoint MRR within 0.05 LB of the best eligible; Pick B = highest public LB, or a different scorer family within 0.02 local of A; ties break to shorter rerun. Pre-register in hive/PREREG.md by 2026-10-20.

Validation gates for any confidence-driven decision (q7/q8): monotone CAC curve, ECE ≤ 0.05, AUROC(absence) ≥ 0.8; freeze first-pass scores before any LB steering; paired bootstrap over molecules; accept a gain only if sign-consistent in ≥ 70 % of class × adduct × n-CE × mass-tercile cells standardised to the hidden case-mix.

Human task (not an agent): a logged-in read of Kaggle Rules §2.6, Overview, discussion 741404, the public/private fraction, notebook RAM/GPU/9-h limits, copied verbatim into research/_meta/kaggle_verbatim.md.

8. Contradictions still open

Carried forward unresolved from the Emperor's tick-1 list, plus those the kings and queens flagged that the Emperor did not list. None is silently resolved here.

  1. Hidden spectra count: ~1,500 (Kaggle mirrors) vs ~2,500 (press release). 2,500/400 = 6.25 = enveda-180 all-adduct mean; 1,500/400 = 3.75 = monomer-only thinned. Decides whether dimers and the thinning operator apply.
  2. Test adducts: 7 (brief, q2, visible file) vs 10 (q1, q4 official texts). [M−H2O+H]+ occurs in np-examples.
  3. Mass window: ±10 ppm (every ≥ 0.37 kernel; q1 "LB-neutral") vs asymmetric −4/+5 ppm with ±10 fallback (q2). Untested pool-shrinking lever.
  4. Mechanism of the +1.9 ppm positive-mode precursor bias (and the +0.4 mDa low-mass fragment bias). One hypothesis: electron-mass omission in Enveda's own m/z calculation — unverified.
  5. np-examples ∩ COCONUT: 249/250 (computed, standard keys) vs 8/250 (one write-up). Almost certainly keying; not yet recomputed under RDKit 2026.03.3. Every c-estimate and π_in hangs on it.
  6. Above-precursor trim: +2.12 Da (enveda-180) vs +2.00 Da (np-examples) — which pipeline produced the hidden set. Both queens agree: trim identically in train and test.
  7. Pool-size penalty: monotone (MetFrag, CSI:FingerID, Laudicina) vs non-monotone, worst at 200–1,000 (CASMI 2016). Accepted resolution: stratify by formula; not yet done.
  8. Class mixture 16/45/39 vs 16/27/55 (two write-ups) → retrieval ceiling ≈ 0.61 vs ≈ 0.45.
  9. Popularity priors: +0.004…+0.011 (one lineage) vs monotonically negative (another); 2016 vs 2017 precedents disagree. Working rule: library-hit class only, gated.
  10. Consensus magnitude: CASMI 2016 +35 Top-10 hits vs sibling evidence that a weaker engine halves the gain and consensus is a catastrophe detector, not a fine ranker. Admission gate binds.
  11. Analog reachability: 82.8 % (np-examples) vs 38.7 % (random COCONUT) vs 95.7 % (q3, formula-level). Different definitions; hidden-set value unmeasured.
  12. Morelli vs Li 2021: do merged multi-CE timsTOF spectra sit above or below spectral entropy 3? q8 asked q1; no answer yet.
  13. Public-LB fraction ≈ 33 % (one kernel) vs unknown. Sets every kill threshold's SE.
  14. LB snapshot discrepancies between queens (q4: 2,703 teams / max 0.481; q7: 2,673 / 0.480) — different pull times on 2026-10-07, not a conflict, but the dashboard should timestamp.
  15. Compute pricing: brief $1.40/GPU-h vs q7 digest $2.10 (A100-40GB) vs q7 plan file ~$2.50 (A100-80GB).
  16. np-examples spectra: 1,184 (file) vs 1,151 (Kaggle page).
  17. Hidden mass range "157–1,159 Da" — recommended by one worker, not found verbatim by another; keep pools wide.
  18. CANOPUS 99.7 % vs CASMI 2022 class 69 % — resolved by both queens in favour of 69 % (sparsity artefact); listed so the dashboard never quotes 99.7 %.
  19. Chalcone/flavanone — resolved: two keys, not tautomers (q3 beyond_2d over np_classes).

9. Sources

Only identifiers that appear in the files read for this synthesis; worker files hold the full lists.

  • Kaggle competition page: https://www.kaggle.com/competitions/enveda-CASMI26-molecule-id-mass-spectra (JS-rendered; facts via Kaggle API pull 2026-10-07 13:59 UTC and GitHub write-ups).
  • enveda-180 (Enamine HLL-200 on timsTOF Pro 2): Zenodo 10.5281/zenodo.20436851; MassIVE MSV000102122.
  • MassSpecGym leakage audit (June 2026): arXiv 2606.19624. Split-inflation / padding-mask findings: arXiv 2610.02249.
  • Schymanski et al., CASMI 2016 automated methods, J Cheminform 2017;9:22, DOI 10.1186/s13321-017-0207-1 (PMC5368104).
  • Schymanski & Neumann, "CASMI: And the Winner is...", Metabolites 2013;3:412–439, DOI 10.3390/metabo3020412 (PMC3901266); "The CASMI: Challenges and Solutions", Metabolites 2013;3:517–538, DOI 10.3390/metabo3030517.
  • Fiehn Lab CASMI 2022 results: https://fiehnlab.ucdavis.edu/casmi/casmi-2022-results
  • C&EN 2026-09-17 (Fiehn and Allen quotes): DOI 10.47287/cen.612552.newsarticle.
  • Enveda Series E press release 2026-09-23: https://enveda.com/news/enveda-raises-311-million/
  • MS2Mol / EnvedaDark: ChemRxiv 10.26434/chemrxiv-2023-vsmpx. MS2Prop: bioRxiv 10.1101/2022.10.09.511482.
  • Dührkop et al. 2015 (CSI:FingerID): PMC4611636. MetFrag fragmentation-only: PMC4732001.
  • Meier et al. 2018 (PASEF; isolation widths): DOI 10.1074/mcp.TIR118.000900 (PMC6283298).
  • Kind & Fiehn 2007, Seven Golden Rules: DOI 10.1186/1471-2105-8-105.
  • Cormack, Clarke & Buettcher 2009, Reciprocal rank fusion, SIGIR 2009.
  • Huber et al. 2021 (Spec2Vec): DOI 10.1371/journal.pcbi.1008724.
  • Isomer blindness (hesperidin/neohesperidin): PMC2722651; glycoside-position classifier: PMC11644359.
  • Library-search CE/instrument sensitivity: PMC6359582, PMC9400100.
  • Dark metabolome / in-source fragmentation: DOI 10.1002/ansa.70012 (PMC12077755); DOI 10.1021/jacsau.5c01063 (PMC12728610).
  • leesirleesirlee-star CASMI-2026 README (9 h ceiling, ~18 s/molecule, recall@25 ~0.545): https://github.com/leesirleesirlee-star/Enveda_CASMI-2026_Molecule-ID-From-Mass-Spectra
  • Cited in digests without an identifier (see the named worker file): Li 2021 entropy similarity; Hart 2024; Stein & Scott 1994; Wixted & Wells 2017; Balding & Donnelly 1996; Murali 2018 IDTAXA; Conneau 2018 CSLS; Liu 2020 energy score; MSAlign 2026; FRIGID, DualLGD, MS-GPT, DiffMS; MIST-CF, BUDDY, FIDDLE; Laudicina, gnshx, kabir0774, nghianguyen7171, billylu24 and ibnmarzuk write-ups.

Decision-relevant numbers

60 facts. Confidence is the hive's own grading, audited by the predator pass.

FactValueConfidenceWhere
train.parquet spectra2539608 spectrahigh_meta/train_eda.md
train.parquet distinct InChIKey14 structures275810 structureshigh_meta/train_eda.md
timsTOF rows in train (all Enveda's own: enveda-180 1,153,785 + np-examples 1,184)1154969 spectrahigh_meta/train_eda.md
enveda-180 structures in COCONUT (of 182,941)51 structureshighsourceq2_data_forensics/_QUEEN_DIGEST.md
enveda-np-examples structures (the only NPs measured on the test instrument)250 structureshighq2_data_forensics/enveda180_and_np_examples_anatomy.md
np-examples structures exclusive to that library (RIKEN has 246, GNPS 233)0 structureshighq5_methods_lineage/failure_modes_generalisation_instrument_shift_evaluation_pitfalls.md
visible test rows that are verbatim enveda-180 train rows (of 1,213)1213 rowshighq2_data_forensics/test_set_decoy_hidden_set_inference_and_domain_shift.md
hidden test molecules400 moleculeshighsourceBRIEF.md
hidden test spectra — Kaggle mirrors ~1,500 vs press release ~2,500 (OPEN)1500 spectra (vs 2,500)lowq2_data_forensics/test_set_decoy_hidden_set_inference_and_domain_shift.md
test adduct strings — visible file 7 vs official texts 10 (OPEN)7 adducts (vs 10)low_meta/king_measure_KINGLOG.md
fixed collision energies on the hidden instrument20/40/60 eVhighq1_instrument_physics/cid_fragmentation_physics_collision_energy.md
test rows that are merged (raw-count sum of 20/40/60 eV scans)24.6 %highq2_data_forensics/adducts_polarity_collision_energy_merging_statistics.md
test molecules with at least one merged spectrum68.5 %highq2_data_forensics/adducts_polarity_collision_energy_merging_statistics.md
timsTOF 20 eV ↔ Orbitrap NCE 20 sqrt-cosine on 2,421 same-molecule pairs (parquet eV over-states NCE ~1.4×)0.82 cosinemediumq1_instrument_physics/cid_fragmentation_physics_collision_energy.md
timsTOF peak-list intensity floor (raw counts, no cap)9 countshighq1_instrument_physics/noise_artifacts_isotopes_insource_chimeric.md
timsTOF peaks below 1 % of base peak91.8 %highq1_instrument_physics/noise_artifacts_isotopes_insource_chimeric.md
Spearman correlation of base-peak intensity with peak count (peak count encodes abundance)0.71 rhohighq1_instrument_physics/noise_artifacts_isotopes_insource_chimeric.md
timsTOF spectra containing the precursor peak (M+1 38.7 %, M+2 21.7 %)45.9 %highq1_instrument_physics/noise_artifacts_isotopes_insource_chimeric.md
positive-mode precursor bias on enveda-180 (cause unknown; OPEN)1.9 ppmhighq2_data_forensics/adducts_polarity_collision_energy_merging_statistics.md
train rows whose shipped precursor_error_ppm shows a spurious ~2,400 ppm (formate relabel artefact)52460 rowshighq2_data_forensics/adducts_polarity_collision_energy_merging_statistics.md
enveda-180 rows that are dimer adducts ([2M+Na]+ 16 %, [2M+H]+ 8 %); test has none29 %highq5_methods_lineage/failure_modes_generalisation_instrument_shift_evaluation_pitfalls.md
np-examples structures outside the decoy's 245–460 m/z band (band = Enamine mass range, not an organiser filter)42 %highq2_data_forensics/test_set_decoy_hidden_set_inference_and_domain_shift.md
COCONUT Oct-2026 distinct InChIKey14 (from 738,794 rows)479690 structureshighq6_candidate_space_retrieval/candidate_databases_pubchem_coconut_lotus_npatlas_chembl_coverage_math.md
median candidates in a ±10 ppm window of IK14-dedup COCONUT (mean 87, p90 204, max 1,761)36 candidateshighq6_candidate_space_retrieval/candidate_databases_pubchem_coconut_lotus_npatlas_chembl_coverage_math.md
median same-formula skeletons per NP in 244–459 Da (58 % face ≥ 25)35 skeletonshighq3_natural_product_chemistry/np_chemical_space_coconut_lotus_npatlas_sizes_overlap.md
np-examples ∩ COCONUT on standard keys (vs '8 of 250' in one write-up; OPEN until recomputed under RDKit 2026.03.3)249 of 250mediumq2_data_forensics/_QUEEN_DIGEST.md
train structures measured on both timsTOF and another instrument (P-inst calibration set)2538 structureshighq5_methods_lineage/failure_modes_generalisation_instrument_shift_evaluation_pitfalls.md
train label groups merged by RDKit tautomer canonicalisation (58,954 spectra; 1.61 % of keys change; 11 skeletons split)1541 groupshighq3_natural_product_chemistry/beyond_2d_tautomers_stereo_isomers_what_the_metric_cannot_see.md
COCONUT rows that are stereo/tautomer spellings of another row35 %highq3_natural_product_chemistry/np_chemical_space_coconut_lotus_npatlas_sizes_overlap.md
public LB rank-1 score, 2026-10-07 13:59 UTC (2,703 teams; rank 10/100/1000 = 0.441/0.425/0.357; median 0.328)0.481 MRR@25highsourceq4_challenge_intelligence/public_leaderboard_and_open_writeups_oct2026.md
teams on 16 exact clone scores (32.7 % of 2,703)885 teamshighq4_challenge_intelligence/public_leaderboard_and_open_writeups_oct2026.md
teams inside the 0.42–0.43 public-LB band228 teamshighq7_compute_execution/_QUEEN_DIGEST.md
derived medal lines gold / silver / bronze0.436 / 0.425 / 0.421 MRR@25medium-highq4_challenge_intelligence/_QUEEN_DIGEST.md
standard error of MRR@25 on the ~130-molecule public subset (0.022 on 400)0.04 MRR@25highq7_compute_execution/ninety_day_plan_milestones_risks_and_kill_criteria.md
public-LB fraction of the hidden set (one kernel, unattributed; OPEN)33 %lowq4_challenge_intelligence/_QUEEN_DIGEST.md
largest public-LB move from a byte-identical notebook rerun0.105 MRR@25medium-highq4_challenge_intelligence/public_leaderboard_and_open_writeups_oct2026.md
back-solved pool coverage c at the LB-0.417 tier (r ≈ 0.90, e ≈ 0.76)0.55–0.61 fractionmediumq6_candidate_space_retrieval/coverage_ceiling_analysis_recall_at_25_and_how_to_raise_it.md
hidden molecules estimated to be in no database (Class 3); alternative write-up says 55 % (OPEN)39 %mediumq6_candidate_space_retrieval/coverage_ceiling_analysis_recall_at_25_and_how_to_raise_it.md
library search alone on the public LB (two write-ups: 0.098, 0.063)0.06–0.10 MRR@25highq4_challenge_intelligence/public_leaderboard_and_open_writeups_oct2026.md
PubChem wholesale as pool: LB 0.112 (train+COCONUT) → 0.079 (train+PubChem)-30 % relativemediumq6_candidate_space_retrieval/candidate_databases_pubchem_coconut_lotus_npatlas_chembl_coverage_math.md
gated PubChem top-50 injection when library confidence < 0.7 (ungated one-step analog enumeration: +0.000)0.011 MRR@25 gainmediumq6_candidate_space_retrieval/coverage_ceiling_analysis_recall_at_25_and_how_to_raise_it.md
MIST-CF MS2-only formula top-1 on MassSpecGym MCES split (true formula already covers 62–90 % of in-window COCONUT structures)50.5 %highq6_candidate_space_retrieval/formula_assignment_sirius_buddy_mistcf_isotopes_timsTOF.md
MassSpecGym papers with evaluation problems in the June-2026 audit (of 26); one method 82.41 → 6.70 % hit@117 papersmedium-highsourceq5_methods_lineage/foundation_models_dreams_prism_massspecgym_benchmarks.md
spectrum-blind baseline (PubChem default order) hit@1 on MassSpecGym49.98 %medium-highsourceq5_methods_lineage/failure_modes_generalisation_instrument_shift_evaluation_pitfalls.md
error correlation above which rank-averaging two scorers loses to the better one alone (combiners stop helping at ρ ≈ 0.6)0.3 rhomediumq5_methods_lineage/foundation_models_dreams_prism_massspecgym_benchmarks.md
CASMI 2016 six-method consensus Top-10 hits vs best single (131/208)166 of 208highsourceq5_methods_lineage/_QUEEN_DIGEST.md
CASMI 2016 challenges at rank 1 from reference counts alone (popularity prior; inverted on a never-published test)143 of 208highsourceq6_candidate_space_retrieval/candidate_ranking_features_learning_to_rank_ensembles.md
de novo exact-match on structure-disjoint MassSpecGym without formula (18–24 % with true formula + [M+H]+)0 %mediumq5_methods_lineage/denovo_generation_ms2mol_diffms_madgen_spec2mol_msnovelist.md
CSI:FingerID CASMI 2016 Top-1 with training overlap removed (from 70)36 hitshighsourceq4_challenge_intelligence/critiques_fiehn_benchmark_validity_what_the_test_does_not_measure.md
CASMI 2022 best 2D-structure rate on all 222 challenges (headline 38 % was on 55 self-selected; formula 94 %)26 %highsourceq4_challenge_intelligence/casmi_history_2012_2022_winners_methods_lessons.md
hexoside [M+H]+ spectra showing no 162 loss (absence is never evidence)86 %highq3_natural_product_chemistry/class_specific_fragmentation_rules_glycosides_terpenes_alkaloids_polyketides.md
one-step analog reachability in COCONUT: np-examples vs random COCONUT (38.7 %); ~28× pool dilution82.8 %highq6_candidate_space_retrieval/coverage_ceiling_analysis_recall_at_25_and_how_to_raise_it.md
fraction of a timsTOF spectrum deleted by matchms default intensity_from=0.00175 %highq1_instrument_physics/noise_artifacts_isotopes_insource_chimeric.md
decoy-shaped local validation → public LB transfer (another write-up: 0.445 → 0.063)0.456 → 0.112 MRR@25highq2_data_forensics/test_set_decoy_hidden_set_inference_and_domain_shift.md
Kaggle notebook rerun ceiling (one GitHub mirror; unverified on Kaggle)9 hoursmediumsourceq7_compute_execution/ninety_day_plan_milestones_risks_and_kill_criteria.md
Modal A100-40GB price (brief said $1.40; plan file quotes A100-80GB ~$2.50)2.1 USD/GPU-hhighsourceq7_compute_execution/_QUEEN_DIGEST.md
random-ranker MRR@25 = H25/n3.816 / n candidateshighq7_compute_execution/_QUEEN_DIGEST.md
prize pool (first place $16,000; winners must open-source)50000 USDhighsourceq4_challenge_intelligence/casmi2026_rules_timeline_prizes_code_requirements_faq.md
final submission deadline (team-merge 2026-12-07; 5 submissions/day; 2 final picks)2026-12-14 23:59 UTC datehighsourceq7_compute_execution/ninety_day_plan_milestones_risks_and_kill_criteria.md
K2 kill criterion: pool recall@25 on local val below this by 2026-10-25 reassigns week 4 to the pool0.5 recall@25highq7_compute_execution/ninety_day_plan_milestones_risks_and_kill_criteria.md

Glossary

[M+H]+
The commonest charged form: the molecule plus one proton.
for the expert

79 % of test rows; 51 % of enveda-180 rows; 1,324,585 train rows. Precursor survives at 20 eV (~11–13 % of TIC), absent in 63–72 % of spectra at 40 eV and 99 % at 60 eV.

13C satellite / isotope peak
A faint echo of each peak one mass unit higher, from molecules containing a heavy carbon atom.
for the expert

+1.0034 Da partner on 71.1 % of timsTOF fragments ≥ 5 % (control 3.0 %); M+1 leaks into 38.7 % and M+2 into 21.7 % of spectra via the 2 Th isolation window; merged into the parent by the telluric mask.

adduct
The particular charged form of a molecule — with an extra proton, a sodium, or missing a proton.
for the expert

Train has 121 adduct strings; test 7 (or 10 — open); five test adduct strings are exactly isobaric with [M'+H]+/[M'−H]− of another formula; build windows on the molecule-level consensus neutral mass.

calibration
Making a model's confidence mean what it says: 70 % confident should be right 70 % of the time.
for the expert

COSMIC-style Platt classifier on description-matched fillers (AUC 0.82 vs 0.40–0.55 raw); isotonic posterior for seating; gates: monotone CAC, ECE ≤ 0.05, AUROC(absence) ≥ 0.8; freeze first-pass scores.

candidate pool / shortlist
The list of catalogue molecules with the right mass from which the ranker chooses its 25 guesses.
for the expert

train ∪ COCONUT (~710–775k rows, 479,690 + ~56 k canonical keys) at ±10 ppm; adding train IK14s raises window median 36 → ~56 with ~zero coverage gain for novel NPs.

CASMI
The community contest for identifying small molecules from mass spectra, founded 2012, run 2012–2017 and 2022, revived by Enveda in 2026.
for the expert

Critical Assessment of Small Molecule Identification (Schymanski & Neumann); 2016 Cat 2 won on Top-1 count; 2022 (Fiehn/WCMC) 16 groups, best 2D 38 % (21/55) vs 26 % (57/222), formula 94 %; no prior edition used timsTOF.

CASP / AlphaFold moment
The protein-folding contest whose 2020 breakthrough CASMI's organisers hope to repeat for molecules.
for the expert

Allen (C&EN 2026-09-17): CASMI 'hasn't yet had its Alphafold moment'; Fiehn: isolated pure compounds, 'much less ground truth'.

Class 1 / 2 / 3 molecules
Test molecules that have a library spectrum / are in a database but have no spectrum / are in no database at all.
for the expert

Estimated 16 / 45 / 39 % (one write-up; 16 / 27 / 55 in another — open) with channel MRR 0.917 / 0.465 / 0; single-pool ceiling ≈ 0.36.

COCONUT
The one public catalogue of all known natural products, about 480,000 distinct structures.
for the expert

Oct-2026 release: 738,794 rows → 728,390 full InChIKeys → 479,690 IK14; 97,757 formulas; CC0; 28.5 % NP-likeness ≤ 0; 38,771 revoked rows; 1,761-entry C6H5N3O artefact; F and sulfonamide entries deleted at import.

collision energy (eV)
How hard the ion is slammed; higher energy makes smaller pieces.
for the expert

Lab-frame electron-volts. Hidden instrument: fixed 20/40/60 eV (not a 1/K0 ramp). Train's `collision_energy_ev` for Orbitrap rows is NCE·m/z/500·f(z), over-stating NCE ~1.4×.

collision-induced dissociation (CID)
Breaking an ion by slamming it into gas molecules.
for the expert

Kinetic (non-LTE) fragmentation in nitrogen; survival yield depends on lab-frame energy, m/z and structure; no complete theory; hand-written rule sets explain 78–85 % of observed fragment ions.

dark metabolome
The signals a mass spectrometer reports that nobody can assign to a known molecule.
for the expert

Unannotated features in untargeted LC-MS/MS; Enveda claims > 95 % of the molecules it observes are unknown; EnvedaDark = 226 NPs absent from major databases.

de novo generation
Having a model write a brand-new structure from the spectrum instead of picking from a catalogue.
for the expert

0.00 % exact on the structure-disjoint MassSpecGym split without formula; 18–24 % with true formula + [M+H]+ (FRIGID, DualLGD, MS-GPT); Class-3 slot filler only; K3 kills it below 0.15 local MRR by 2026-11-01.

decoy test set
The visible practice file, which is secretly a copy of training rows and reveals nothing about the real molecules.
for the expert

1,213/1,213 rows join uniquely to enveda-180; 400 structures drawn at random; 0/664 dimer rows kept, monomers thinned to 66.7 %; 0/400 in COCONUT; public LB is not scored on it.

dimer adduct
Two copies of the same molecule stuck together on one charge.
for the expert

[2M+Na]+ 16 % and [2M+H]+ 8 % of enveda-180 rows (29 % dimers, 282,074 rows); zero in the visible test; dimers dissociate to monomers at 20 eV.

electrospray ionisation (ESI)
Spraying liquid from a needle at thousands of volts so each molecule picks up a charge.
for the expert

Soft ionisation producing even-electron ions; adduct formation is kinetic, not equilibrium (no Saha-type constant); in-source fragmentation affects 82 % of 214 natural compounds in one Orbitrap study.

Enveda / PRISM
The Boulder drug company hosting the contest, and its secret in-house spectrum model.
for the expert

Host, metric author and answer-key validator (SIRIUS + NMR); revived CASMI 2026-09-15, Series E $311 M on 2026-09-23; PRISM: BERT-family, 1.2 B spectra, closed weights, relative numbers only, no preprint.

enveda-180
Enveda's 1.15-million-spectrum training library, measured on the test instrument but made of synthetic drug-screening compounds, not natural products.
for the expert

Enamine HLL-200 on timsTOF Pro 2, PASEF, 20/40/60 eV, 77 % ESI+; Zenodo 10.5281/zenodo.20436851; 182,941 structures of which 51 in COCONUT; 23 % fluorinated vs 0.4 % in test chemistry; 29 % dimer rows.

enveda-np-examples
The only 250 natural products measured on the test instrument; precious, but every one also has spectra in other libraries.
for the expert

1,184 spectra (Kaggle page says 1,151), 250 structures, 249 in COCONUT (standard keys), 0 exclusive (RIKEN 246, GNPS 233), ';'-separated CE lists at 35/50/80 eV, 42 % outside 245–460 m/z, no public description.

fingerprint (Morgan, MACCS)
A long list of yes/no answers about which chemical building blocks a molecule contains.
for the expert

Predicted from the spectrum (FPNet / MLP heads for Morgan r2 + r3 + MACCS) and compared to candidates by Tanimoto; fingerprint choice is second-order (four variants within 0.007 LB); fingerprints on canonical tautomers differ from database drawings (Δ4-3-keto steroids 41 % re-drawn).

fragment
A piece of the molecule produced when the precursor is broken apart.
for the expert

Product ion; C5–C9 fragments on timsTOF read +0.3–0.6 mDa (+4 to +6.5 ppm) high below m/z 150; 71.1 % of fragments ≥ 5 % carry a 13C satellite at +1.0034 Da.

glycoside
A natural product with a sugar attached; one of the few families with a recognisable sound in the spectrometer.
for the expert

O-hexosides: train's timsTOF stratum holds only 27 (+) / 21 (−) spectra; COCONUT glycosides are 47 % duplicate spellings; hydroxycinnamoyl esters mimic sugar losses at 15–21 mDa.

hold-out / scaffold split
Test molecules kept entirely out of training, including their close relatives, so the exam is honest.
for the expert

Truth-absent hold-out: ~1,000 scaffold- and IK14-disjoint NPs, analogs removed at MCES < 10 / Tanimoto ≥ 0.7, truth deleted from the pool for a random half, test adduct × n-CE mix; fold spec heldout_A 1,977 structures / 13,556 spectra, heldout_B 293.

in-silico fragmenter / forward model
Software that predicts what spectrum a candidate structure would give, for comparison with the real one.
for the expert

MetFrag, CFM-ID 4, ICEBERG 2, GLACIER, MS-FINDER; none models timsTOF CE; MetFrag 24/208 top-1 at CASMI 2016; ICEBERG 2 46.4 spectra/s on 2×A5000; 18 % coverage at a 300 s budget.

InChIKey14
A 14-character code for a molecule's skeleton and connectivity, ignoring handedness; two guesses with the same code count as one.
for the expert

First block of the InChIKey (connectivity hash); the scorer's key is the IK14 of the RDKit-canonical tautomer; `train.inchikey14` is not the scoring key (differs on 1.6 % of train, 10.44 % of COCONUT naive keys).

intensity / base peak
A peak's height; the base peak is the tallest, set to 1.0 in the files.
for the expert

`ms2_normalized_intensities` is base-peak normalised; `base_peak_intensity` carries raw counts for timsTOF (p5/p50/p95 = 2,543 / 201,708 / 14.3 M) and is null for the 7 public libraries — a library barcode.

intensity floor
The smallest peak a library bothered to record; it differs by library and betrays where a spectrum came from.
for the expert

MassBank 1/999, MS-DIAL 0.01, RIKEN 0.01/0.005, pluskal ~7e-5, timsTOF none (9 raw counts); matchms default `intensity_from=0.001` deletes 75 % of a timsTOF spectrum.

ion
A molecule carrying an electric charge, which is what the instrument can steer and weigh.
for the expert

Precursor and fragment ions here are singly charged even-electron species in > 99 % of rows; [M+2H]2+ is 1,717 of 2.54 M train rows.

ion mobility
Sorting ions by their shape as they drift through a gas; the timsTOF does it but the data we get omits it.
for the expert

TIMS (trapped ion mobility) separation and CCS are not exported; whether mobility-separated isomers were merged into one spectrum is unasked (gap).

isomer
Molecules with the same atoms connected differently; they weigh exactly the same.
for the expert

Constitutional isomers share formula and exact mass; MassSpecGym formula-restricted sets are harder than mass-window sets (MIST 9.57 vs 14.64 % hit@1) because same-formula decoys are closer.

Kaggle code competition
Submissions are notebooks re-run offline by Kaggle on the hidden data, under time and memory limits.
for the expert

`is_kernels_submissions_only=True`; ~9 h rerun, internet off (one mirror, medium); Python 3.13, numpy 2.1.3, no RDKit; 29–32 GB RAM; 30 h/week GPU quota; 5 submissions/day; 2 final picks.

leakage
When the answer sneaks into the training or validation data, making scores look better than they are.
for the expert

np-examples structures in RIKEN (246) and GNPS (233); per-spectrum splits put ~94 % of structures on both sides; MassSpecGym audit: ≥ 17/26 papers affected, 82.41 → 6.70 % hit@1, spectrum-blind baseline 49.98 %.

learning to rank (LambdaRank)
A model trained to order candidates well rather than to score each in isolation.
for the expert

LightGBM `lambdarank` over log-LR features (fingerprint sim, analog sim, formula posterior, mass error, provenance); no MS paper with a LambdaMART ablation could be opened; RRF k = 60 is the control it must beat.

library search / cosine similarity
Comparing a new spectrum directly against stored spectra of known molecules.
for the expert

Modified cosine / entropy similarity; LB 0.06–0.10 alone; Class-1 channel MRR 0.917; sensitivity 88.0/91.5 % cross-instrument vs 99.5/99.7 % self.

liquid chromatography (LC)
Pushing a dissolved sample through a tube that separates molecules by how strongly they stick to it.
for the expert

Reversed-phase LC upstream of ESI; Enveda uses 2D-LC mass-targeted isolation; retention time is not exported in train or test.

log-odds / likelihood ratio (LR)
How much more likely a clue is if the candidate is right than if it is wrong; evidence that can be added up.
for the expert

Sum calibrated log-LRs across orthogonal engines; rank-averaging loses at error correlation ρ ≥ 0.3; independence combiners stop helping at ρ ≈ 0.6 → jointly fitted stacker; LR 10³ with n_f 10⁴ gives posterior 0.09.

m/z
Mass divided by charge — the number the instrument actually measures for each ion.
for the expert

Mass-to-charge ratio in Thomson; train precursor m/z median 350.1, with a corrupt max of 2,430,333.

mass spectrometer
An instrument that weighs molecules by giving them a charge and measuring how they move in electric fields.
for the expert

Here a Bruker timsTOF Pro 2: ESI source, TIMS mobility cell, quadrupole isolation (2 Th < m/z 700, 3 Th above), CID cell, orthogonal TOF analyser.

mass window
The tolerance around the measured mass within which catalogue molecules count as candidates.
for the expert

±10 ppm in every ≥ 0.37 public kernel; ±5 vs ±10 LB-neutral in one write-up (0.302/0.303); q2 proposes asymmetric −4/+5 ppm (open); ±10 ppm COCONUT window median 36 candidates, p90 204.

MassSpecGym
The main academic benchmark for this problem, whose published numbers were mostly found to be inflated in 2026.
for the expert

MCES / structure-disjoint split; MIST 14.64 % hit@1 (random 0.37 %); June-2026 audit arXiv 2606.19624; '> 300 peaks' rule discards ~25 % of enveda-180.

merged spectrum
Three spectra at different energies added together, peak by peak, into one.
for the expert

Per-m/z raw-count sum of the 20/40/60 eV scans, re-centroided; effective weights ≈ 0.5/0.4/0.1; 24.6 % of test rows, 68.5 % of test molecules; median 258 peaks vs 88–132 for singles.

metabolome
The complete set of small molecules present in an organism or sample.
for the expert

The small-molecule (< ~1,500 Da) complement of a biological system as observed by LC-MS; here restricted to what ESI-MS/MS can see.

mock-witness test
Try to rank candidates using only the mass and adduct, with no fragments; if that beats chance, the shortlist itself is biased.
for the expert

Ranker on precursor m/z + formula + adduct only; compare against H₂₅/n = 3.816/n; positive result means local scores harvest line-up (pool) bias, not spectral evidence.

molecular formula
The atomic recipe of a molecule, e.g. C15H10O7, without saying how the atoms connect.
for the expert

Inside a ±10 ppm COCONUT window the true formula already covers 62–90 % of structures (median 2–4 formulas); MS2-only formula top-1 is 50–77 %; use as a soft feature, break-even P ≈ 0.75–0.9.

MRR@25 (mean reciprocal rank)
The score: one point if the right answer is first, half if second, a tenth if tenth, nothing beyond 25, averaged over molecules.
for the expert

MRR@25 = mean(1/rank of first correct, 0 if rank > 25); SE ≈ 0.022 on 400 molecules, ≈ 0.04 on ~130; decomposes as c·r·e (pool coverage × ranked-in-25 × mean credit).

natural product
A molecule made by a living organism, often borrowed by humans as a medicine.
for the expert

Secondary metabolite from plants, fungi, bacteria or animals; the hidden set is NPs, hypothesised NPs, analogs and nature-plausible synthetics; NPClassifier pathways: terpenoids, alkaloids, polyketides, shikimates, amino-acid-derived, carbohydrates, fatty acids.

negative mode / [M−H]−
Measuring the molecule after it has lost a proton instead of gained one.
for the expert

16 % of test rows; 465,688 train rows; all class likelihood ratios in the hive are positive-mode; nobody studied NP [M−H]− fragmentation on timsTOF (gap).

neutral loss
The mass difference between a precursor and a fragment, revealing which piece fell off (e.g. a sugar, 162).
for the expert

Hexose 162.0528 LR+ 46, deoxyhexose 146 LR+ 40, pentose 132 LR+ 22, glucuronide 176.0321 LR+ ~270; sensitivity falls 0.33 → 0.02 with precursor mass; 86 % of hexoside [M+H]+ spectra show no 162 loss.

normalised collision energy (NCE)
Thermo's instrument-relative scale for collision energy, not directly comparable to volts.
for the expert

timsTOF 20 eV ↔ NCE 20 (sqrt-cosine 0.82), 40 ↔ 45 (0.70), 60 ↔ 60 (0.61) on 2,421 same-molecule pairs; drive NCE-conditioned models at ~0.7× the eV formula value.

Orbitrap
A different, Thermo-made kind of mass analyser on which most public library spectra were recorded.
for the expert

745,562 train rows labelled Orbitrap; uses NCE; GNPS-Orbitrap spectra show M+1 in 6.1 % and 13C satellites on 14.0 % of strong fragments vs 38.7 % / 71.1 % on timsTOF.

peak
One entry in a spectrum: a mass and how much of it was seen.
for the expert

91.8 % of timsTOF peaks are < 1 % of the base peak; 75.1 % < 0.1 %; peak count tracks sample abundance (Spearman 0.71 with base-peak intensity).

popularity prior
Favouring molecules that have been written about a lot — which backfires on a test of unpublished molecules.
for the expert

PubChem/ChemSpider reference counts: MetFrag top-1 30 → 336 → 420 of 473; 143/208 CASMI 2016 at rank 1 from counts alone; LB +0.004…+0.011 vs monotonically negative (open); rankers learn it implicitly via training-spectrum counts.

positional isomer
Same skeleton, same groups, one group attached in a different place.
for the expert

Median 35 (mean 76, p90 198) same-formula skeletons per NP in 244–459 Da; 58 % face ≥ 25; hesperidin/neohesperidin give no MS/MS difference; a dedicated glycoside-position classifier tops at 84.7 %.

ppm (parts per million)
How far a measured mass is from the true one, as a fraction of the mass.
for the expert

Recomputed signed error with electron mass; enveda-180 99.98 % within ±5 ppm with a +1.9 ppm positive-mode bias; np-examples [M−H]− up to 21.1 ppm; the shipped `precursor_error_ppm` column is an unsigned, electron-free artefact.

precursor
The intact charged molecule the instrument selects before smashing it.
for the expert

Isolated in a 2–3 Th quadrupole window; precursor peak present in 45.9 % of timsTOF MS2 spectra (base peak in 25–29 % of merged [M+H]+); strip > precursor − 1.0 Da before scoring.

PubChem
The 124-million-entry catalogue of all known chemicals; using it wholesale buries the right answer.
for the expert

124,717,116 CIDs (2026-10-07); ±10 ppm window ≈ 10⁴ candidates; train+PubChem LB 0.079 vs train+COCONUT 0.112 (−30 %); gated top-50 injection +0.011; prior-density ratio vs COCONUT ≈ 1,500–2,000×.

public / private leaderboard
The live scoreboard uses a slice of the hidden molecules; the final ranking uses the rest.
for the expert

Public ≈ 33 % (~130 molecules; one kernel, low confidence); 2,703 teams on 2026-10-07; 885 teams (32.7 %) on 16 exact clone scores; 228 teams in 0.42–0.43; derived medal lines gold 0.436 / silver 0.425 / bronze 0.421.

RDKit canonicalisation
The pinned software step that rewrites every guess into one standard form before comparing codes.
for the expert

`MolFromSmiles → TautomerEnumerator().Canonicalize → MolToInchiKey[:14]`, RDKit 2026.03.3 (wheel 37,187,145 B, cp313); `maxTransforms=1000` is hit silently on polyphenols (quercetin 179 ms, rutin 357 ms) but raising it changed zero of 8 IK14s; 17.5 ms mean, > 0.4 s p99.

recall@25 (candidate coverage, c)
The fraction of molecules whose true answer appears anywhere in the 25 guesses.
for the expert

Back-solved ≈ 0.545–0.55 at LB 0.417 (c ≈ 0.55–0.61 truth-in-pool); the ceiling on every method; K2 kill criterion: pool recall@25 < 0.50 by 2026-10-25.

reciprocal rank fusion (RRF)
A simple way to merge several ranked lists by rewarding items that rank high in any of them.
for the expert

RRF k = 60, flat over k = 10–100 (Cormack 2009, +4–5 % over best system); CombMNZ beat it on LETOR 3 (.6107 vs .6051) → bake-off; Borda/rank-averaging lets one engine veto.

SIRIUS / CSI:FingerID
The leading academic identification software, used by Enveda to validate the answers; a licensed web service we cannot run offline.
for the expert

Fragmentation-tree formula assignment + Bayesian fingerprint scoring; CASMI 2016 70 → 36 top-1 with training overlap removed; 68.5 → 34.4 % biodatabase → PubChem; SIRIUS-explainable spectra likely over-represented in the hidden set.

SMILES
A line of text that spells out a molecular structure, the format answers are submitted in.
for the expert

Submission = up to 25 semicolon-separated SMILES per molecule_id; invalid or duplicate-key SMILES silently burn slots; multi-fragment/salt SMILES are not normalised by the scorer; submit single-fragment, parent-stripped SMILES.

spectral entropy
A measure of how spread out a spectrum's intensity is across its peaks.
for the expert

Li 2021 entropy similarity (best of 43, AUC 0.958) uses I^w with w = 0.25 + 0.25·S for S < 3, w = 1 above; whether merged timsTOF spectra sit above S = 3 is open (Morelli vs Li).

standard error (SE)
The typical wobble in a score from chance alone; differences smaller than it mean nothing.
for the expert

Per-molecule RR SD ≈ 0.45 → SE 0.022 (n = 400), 0.032 (200), 0.041 (120); public subset ~130 molecules; byte-identical reruns moved up to 0.105; never promote on an unpaired LB delta < 0.04.

stereoisomer
Mirror-image or left/right-handed versions of the same connectivity; the scorer ignores the difference.
for the expert

InChIKey14 drops the stereo layer; 17.2 % of COCONUT records have no stereo layer and 14,320 IK14 have ≥ 5 stereo variants (max 72); terpenoids median 8 stereocentres.

tandem mass spectrum (MS/MS, MS2)
The list of fragment masses and their heights produced from one selected precursor.
for the expert

Centroid peak list (`ms2_mzs`, `ms2_normalized_intensities`); train median 42 peaks, mean 158, max 73,318; timsTOF rows are raw integer counts with floor exactly 9.

Tanimoto similarity
A 0-to-1 score for how many building blocks two molecules share.
for the expert

Jaccard on fingerprint bits; analog hold-out removal at Tanimoto ≥ 0.7 / MCES < 10; best modified-cosine analog averages Tanimoto 0.45, MS2Query 0.63 at 35 % recall.

tautomer
Two drawings of one molecule that differ only by where a proton sits; the scorer treats them as the same.
for the expert

RDKit `TautomerEnumerator().Canonicalize` merges keto/enol, curcumin, tetracycline, usnic acid, anthrone/anthranol, imine/enamine; not chalcone/flavanone; changes 1.61 % of train keys, merges 1,541 label groups, 12–17 % of NP keys.

telluric mask
Fixed rules that remove instrument-made peaks (the unbroken parent, its heavy-carbon echoes) the way astronomers mask Earth's atmosphere out of starlight.
for the expert

Five steps: delete m/z > precursor − 0.02; merge +1.00335 ± 0.005 satellites (ratio 0.02–0.5); cut < 20 counts; subtract 0.4 mDa below m/z 150; optional frequency weighting; keep if MRR@25 +0.02 on the backtest.

three-mask np-examples backtest
The one local test that predicted the public leaderboard: 250 natural products on the right instrument, with their spectra removed from every library.
for the expert

250 enveda-np-examples structures under three masks, train-key removed across all libraries; predicted LB order for three independent teams while random train splits predicted the opposite; primary model selector.

time of flight (TOF)
Weighing ions by timing how long they take to fly down a tube.
for the expert

Orthogonal TOF analyser; mass accuracy on enveda-180: median 0.98 ppm, 99 % < 4.4 ppm, truncated at |5.13| ppm; errors correlated within a run.

timsTOF
The Bruker instrument that measured every hidden-test spectrum.
for the expert

timsTOF Pro 2 with PASEF; all 1,154,969 train timsTOF rows are Enveda's own (enveda-180 + enveda-np-examples); no published learned method has a timsTOF evaluation.