2026-10-07 01:20 [emperor] — see research/_meta/train_eda.md

Element coverage and carbon-count histograms of the 275,810 train structures are in dashboard/src/data/eda_train.json (keys element_coverage_structures, carbon_count_hist_structures). enveda-np-examples = 250 NP structures on timsTOF, all also in public libraries.

2026-10-07 — [q6_candidate_space_retrieval] -> q3_natural_product_chemistry: which databases under-cover which classes

1. There is effectively one open NP map, not several. COCONUT Oct-2026 = 479,690 unique InChIKey14 (computed from the download); LOTUS v11 adds 254; the 29-source THEOBROMA union reaches only 486,215 IK14 families. NPAtlas 3.0 (36,545 microbial NPs, v2024_09) and SuperNatural 3.0 / NPASS 3.0 add zero IK14 coverage over COCONUT and are CC BY-NC, so they are not worth shipping. Practical consequence: class-level gaps in COCONUT are gaps in every pool we can build.
2. Mass-window density is class-relevant: +/-10 ppm in COCONUT gives median 36 candidates, but p90 204 and a worst case of 1,761 (a curation artefact at C6H5N3O). Classes with many isomeric variants (glycosides, terpenoid scaffolds with the same formula) will sit in the dense tail.
3. A public write-up reports 70.8 % of in-window candidates already share the true formula, so isomer discrimination - not formula - is the chemistry problem that decides our rank.

Ask for q3: two concrete questions. (a) Which NP superclasses are systematically absent or thin in COCONUT — microbial post-PKS tailoring products, marine halogenated metabolites, plant specialised metabolites from under-sampled taxa? I need a class list to decide where analog enumeration must cover for the ~40 % of hidden molecules likely outside every database. (b) For "one-step analogs", which biosynthetic transformations are the realistic ones to enumerate (hydroxylation, O-methylation, glycosylation, prenylation, halogenation) and what mass shifts do they imply? I want an enumeration operator grounded in biosynthesis, not in generic RDKit substituent swaps. — [q6_candidate_space_retrieval]

2026-10-07 (tick 2) — [q6_candidate_space_retrieval] -> q3_natural_product_chemistry: our numbers agree; two refinements; one ask

1. Agreement. Your worker and mine both recomputed COCONUT Oct-2026 to 479,690 IK14 from 738,794 rows. Minor difference: you find NPAtlas adds 138 new IK14, my worker reported ~0 — immaterial (Δc < 0.001). Your "27 median competitors at ±5 ppm, p90 193" and my "36 median at ±10 ppm, p90 204" are consistent.
2. Your 95.7 % formula-level parent / 247-of-250 figure vs my structure-level 82.8 %. Different definitions: mine requires the one-step product to be an exact COCONUT IK14 (15 RDKit operators), yours requires a formula-level relative. The gap (97 % → 83 %) is the regiochemistry combinatorics you flagged. On a random COCONUT sample (not the 250 well-studied NPs) my reachability falls to 38.7 % — so the 250 overstate analog reach by ~2×. Operators ranked by how many of the 250 they reach: dehydroxylate 113, sp3-hydroxylate 106, aromatic-hydroxylate 83, O-methylate 83, O-hexosylate 79, O-acetylate 61 (coverage_ceiling_...md §4). Cost: 33 products/parent mean → ±10 ppm pool inflates ~28× (median 35 → ~1,050).
3. Ask (refines my tick-1 question). You say sesquiterpene/diterpene formulas carry 600–1,300 isomers. Which NPClassifier superclasses/classes populate the dense tail (p90+ windows) in the 245–460 m/z test band? I want to restrict analog enumeration to classes where (a) one-step relatives are plausible and (b) the window is not already saturated with isomers, and to build a per-class "sensor sweep width" for the ranker. — [q6_candidate_space_retrieval]

2026-10-07 — [q3_natural_product_chemistry] -> q6_candidate_space_retrieval: answers to (a) thin classes and (b) biosynthetic enumeration operator

(a) Which classes are thin or absent in COCONUT — and the one that matters is absent by policy, not by sampling.

1. COCONUT deletes fluorinated and sulfonamide-bearing entries at import (COCONUT 2.0 paper; confirmed by our enveda_platform worker). Enveda's own analog patents (withaferin A, isoandrographolide, nootkatone families) put F at C-4, on side-chain CH2F and as aryl-F, plus heteroaryls/NMe2/nitriles. So the "synthetics plausibly in nature / analogs" slice of the hidden 400 (our estimate 15–25 %) is unreachable in COCONUT at any pool size. Train shows the contrast: enveda-np-examples 0.4 % F / 3.6 % S vs enveda-180 23.0 % F / 28.3 % S. Mitigation: a small side pool of enumerated F/Cl/Br/heteroaryl children of top-k COCONUT scaffolds, or scaffold-restricted PubChem (same Murcko scaffold as a COCONUT hit) — not full PubChem. File: research/q3_natural_product_chemistry/enveda_platform_isolation_profiling_what_analogs_means.md §6.
2. Kingdom skew, not class absence, is the second gap. Terpenoids are 46.5 % of plant structures but 6.2 % of bacterial; peptides 1.5 % vs 27.5 %. NPAtlas (microbial) adds only 138 new IK14 over COCONUT, so microbial post-PKS/NRPS products are present but the COCONUT-wide class prior is plant-shaped; 10.8 % of NPAtlas entries are pathway hybrids (PKS-NRPS 1,997; alkaloid-peptide 891). Marine halogenated: CMNPD is inside COCONUT (30,700 rows / 26,538 IK14) — we found no evidence it is systematically thin, but note train has only 8,522 Br-containing and 369 I-containing structures, so the spectral model, not the pool, is the weak link for marine halogenated NPs. Files: np_classes_biosynthesis_npclassifier_classyfire.md §2–3, np_chemical_space_coconut_lotus_npatlas_sizes_overlap.md §1.
3. COCONUT's alkaloid pathway (23 % of COCONUT vs 11–13 % of LOTUS/NPAtlas) has mean NP-likeness 0.07 — it is diluted with synthetic-looking N-heterocycles from bulk imports (Super Natural II is the largest contributor: 311,885 rows, 110,432 sole-sourced). Treat alkaloid-window candidates with NP-likeness < 0 as filler; your 1,761-candidate C6H5N3O bucket is the same artefact.
4. Lipids/fatty acids: 44.6 k enumerated triacylglycerols inflate COCONUT's fatty-acid pathway (MW median 807 Da). Over-represented in the pool, implausible in a 245–460 m/z timsTOF test window — these are q7's "wrong fillers".

(b) The enumeration operator, grounded in measured COCONUT density. For every COCONUT IK14 we asked whether another COCONUT formula sits exactly one tailoring delta away (fraction with such a relative): +CH2 (O/N-methylation, 14.01565) 84.1 %; −H2 (desaturation, 2.01565) 80.2 %; +O (hydroxylation/epoxidation, 15.99491) 79.8 %; +C2H2O (acetylation, 42.01057) 78.0 %; +CH2O (methoxylation, 30.01056) 77.7 %; +C5H8 (prenylation, 68.06260) 68.9 %; +C3H2O3 (malonylation, 86.00039) 58.0 %; +C5H8O4 (pentosylation, 132.04226) 46.6 %; +C6H10O5 (hexosylation, 162.05282) 33.7 %; +C6H8O6 (glucuronidation, 176.03209) 25.5 %; +SO3 (sulfation, 79.95682) 4.1 %. Any of the eleven: 95.7 %. On the 250 enveda-np-examples molecules, 247/250 have ≥1 such relative in COCONUT (+O 237, −O 234, +CH2 233, +C2H2O 229, +H2 229, +C6H10O5 200). Add halogenation (+Cl−H 33.96103, +Br−H 77.91051) for the analog slice only. Expected gain if gated (fire only when in-window IK14 pool < ~15 or best in-pool score is low): +0.01…+0.04 MRR@25; ungated it dilutes the window 10–100× (a 6-OH parent yields 6 O-methyl regio-isomers per delta) and will cost score, same mechanism as PubChem. File: np_chemical_space_coconut_lotus_npatlas_sizes_overlap.md §Forced analogy.

(c) Fact you did not ask for but need: the neutral-loss "timbre" that would break regio-isomer ties is nearly absent from the timsTOF stratum of train. Our fragmentation worker found only 27 ([M+H]+) and 21 ([M−H]−) timsTOF spectra of terminal O-hexosides in 1.16 M timsTOF rows, and zero oleanane triterpenes matching a strict SMARTS — the timsTOF rows are enveda-180 synthetics. Class rules (hexose 162.0528 LR+ 46, glucuronide 176.0321 LR+ ~270, tropane 124.1121 LR+ 42, desosamine 158.1176 LR+ 369, C-glycoside cross-ring 120.0423 LR+ 66–76; water/terpene losses LR+ ≤ 2.6 = useless) were measured on Orbitrap/QTOF rows and meet timsTOF for the first time in the hidden set. 19-rule soft log-LR reranking table with SMARTS: class_specific_fragmentation_rules_glycosides_terpenes_alkaloids_polyketides.md §8 (+ .json, table reranking_rules). Use as promote-only evidence inside the formula window; never as a hard filter (86 % of hexoside [M+H]+ spectra show no 162 loss at ≥5 %).

Question for q6: when you build the window, can you emit per-candidate NPClassifier pathway + is_glycoside + NP-likeness columns from the COCONUT CSV so the reranker can apply the rules without re-running SMARTS on a million structures? — [q3_natural_product_chemistry]

2026-10-07 — [q3_natural_product_chemistry] -> q6: the metric-key numbers for your pool (addendum)

1. Re-key the pool with the grader's own function, not plain InChIKey14. K(s) = InChIKey(TautomerEnumerator().Canonicalize(flat, largest-fragment, proton-neutral mol))[:14], rdkit pinned 2026.03.3 (our 60 NP pair tests were bit-identical on 2026.03.3 and 2026.03.6; 2026.09.x changes the tie-break). On train it merges 1,541 groups of distinct stored inchikey14 (3,167 skeletons, 58,954 spectra) and splits 11 skeletons (large peptides, bilins, thiopeptides, rifamycins — enumeration hits maxTransforms). Key changes for 1.61 % of train SMILES overall but 6.9 % of COCONUT-member structures (ansa-macrolides 79 %, dihydroflavonols 52 %, pregnane steroids 41 %, naphthoquinones 41 %). Cost: 17.5 ms mean, >0.4 s p99 → ~3.6 CPU-h for COCONUT; do it once offline on Explorer, ship key → representative flat SMILES.
2. Your 479,690 is the right denominator; raw entries are 35 % spellings (steroids 54 %, macrolides 59 %, glycosides 47 %; one skeleton has 72 stereo entries). Flavonoid hexoside formulas are ~68 % enharmonic (C21H20O11: 449 entries → 142 keys; rutin C27H30O16: 495 → 157), triterpene acids C30H48O3: 890 → 385, whereas aglycone formulas are 2–5 % (quercetin C15H10O7: 61 → 59). A top-25 drawn from raw entries carries only 8–17 distinct keys in those classes.
3. Positional-isomer density inside the test window (244–459 Da neutral mass): median 35 same-formula skeletons per structure, mean 76, p90 198; 58 % of skeletons have ≥25 same-formula neighbours. This is the recall@25 ceiling you back-solved — not stereo, not formula.
4. Train SMILES carry zero stereo and zero multi-fragment SMILES; COCONUT holds 4,247 multi-fragment and ~2,400 metal-containing entries, which change the key. Strip to largest organic fragment; leave quaternary cations charged (berberine is keyed as the cation). Files: research/q3_natural_product_chemistry/beyond_2d_tautomers_stereo_isomers_what_the_metric_cannot_see.md (+ .json: 60-pair table, per-formula density, redundancy by superclass). — [q3_natural_product_chemistry]

2026-10-07 01:20 [emperor] — official train.parquet EDA now on disk

Read research/_meta/train_eda.md (numbers from the official Kaggle download, not a mirror). Headlines: 2,539,608 spectra / 275,810 InChIKey14 / 51,291 formulas; 11 libraries; timsTOF rows = 1,154,969 and ALL of them are enveda-180 (1,153,785) + enveda-np-examples (1,184); enveda-np-examples has 250 structures, 0 exclusive (all also present in public libs) — it is the only NP-on-timsTOF proxy for the hidden test; visible test.parquet rows match enveda-180 rows 1213/1213 on (precursor_mz, num_peaks, base_peak) — decoy confirmed. 64.6 % of train precursors fall in the test's 245–460 m/z window.

2026-10-07 — [q6_candidate_space_retrieval] -> q2_data_forensics: three pool numbers to anchor on

1. At the metric's resolution the natural-product universe is small. My worker downloaded the Oct-2026 COCONUT CSV: 738,794 rows collapse to 479,690 unique InChIKey14. LOTUS v11 (151,885 IK14) is 99.83 % inside COCONUT and contributes only 254 new connectivities; THEOBROMA (29-source union, bioRxiv 2026.06.12.731585) dedups to 486,215 IK14 families. Unioning NP databases does NOT enlarge the map. File: research/q6_candidate_space_retrieval/candidate_databases_pubchem_coconut_lotus_npatlas_chembl_coverage_math.md.
2. Window sizes, computed on IK14-deduplicated COCONUT: a +/-10 ppm neutral-mass window yields median 36 candidates per molecule, mean 87, p90 204, max 1,761. In the 300-400 Da band the median is 64; 400-500 Da it is 49; 500-600 Da it is 20. Since your EDA says 64.6 % of train precursors sit in the test's 245-460 m/z window, the hidden test lands in the densest part of the window-size distribution, not the sparsest.
3. Formula is a weak filter here: in one public write-up's pool, 70.8 % of in-window candidates already carry the true molecular formula, so perfect formula prediction buys under 30 % pool reduction. See research/q6_candidate_space_retrieval/formula_assignment_sirius_buddy_mistcf_isotopes_timsTOF.md.

Ask for q2: enveda-np-examples is 250 NP structures on timsTOF with 0 exclusive structures — can you publish (a) their InChIKey14 list and (b) the precursor_error_ppm distribution restricted to instrument_type == timsTOF, split by adduct? I need the real ppm spread to set the retrieval window; vendor specs are the wrong source. Also: how many of the 275,810 train IK14 are inside COCONUT? That number sets P(true structure in pool) for any COCONUT-only strategy. — [q6_candidate_space_retrieval]

2026-10-07 (tick 2) — [q6_candidate_space_retrieval] -> q2_data_forensics: one contradiction to settle, two facts

1. Contradiction that decides my coverage estimate. Your enveda180_and_np_examples_anatomy.md says the enveda-np-examples/COCONUT overlap "swings from 8 of 250 to 99.6 % across write-ups depending on how structures were keyed". My coverage-ceiling worker assumed 99.6 % (and measured 82.8 % one-step reachability on those 250). If the true IK14 overlap after RDKit tautomer canonicalisation is anywhere near 8/250, every c-estimate in my digest shifts and the analog-enumeration case gets stronger. Request: compute, with the metric's own pipeline (RDKit TautomerEnumerator().Canonicalize() → InChIKey14), how many of the 250 np-example IK14s and how many of all 275,810 train IK14s fall in COCONUT Oct-2026 (C:\Temp\casmi26\npdb\coconut\). q3's worker independently got 51/182,941 enveda-180 in COCONUT; that part is settled.
2. Fact: train structures are mostly pool dilution, not coverage. enveda-180 is Enamine HLL-200 chemistry (0.03 % in COCONUT). Adding the 275,810 train IK14s to COCONUT raises the ±10 ppm window median from 36 to ~56 (q1's write-up number) with ~zero coverage gain for a novel-NP test. Keep train structures in the pool only for the ~16 % Class-1 library-hit channel, and give them a per-source prior, not equal footing.
3. Fact: cleaning asymmetry. Your finding that 62,834 train rows exceed 20 ppm while the test was cleaned on this criterion means the retrieval window must be set from the cleaned timsTOF slice (|ppm| ≤ 10), not from the raw column. q1 reports (unverified, one notebook) 99 % of enveda-180 errors < 4.4 ppm — if you can confirm the p99 on timsTOF rows I will set the window at ±5 ppm with a ±10 ppm fallback. — [q6_candidate_space_retrieval]

2026-10-07 — [q2_data_forensics] -> q6_candidate_space_retrieval: answers to your three asks (part 1 of 2)

Computed from the official local train.parquet joined to your COCONUT CSV (coconut_csv_lite-10-2026, 738,794 rows -> 479,690 unique IK14, same number you got). Keys are the shipped inchikey14 / COCONUT standard_inchi_key[:14], NOT re-canonicalised with the RDKit 2026.03.3 tautomer protocol (expect a few % drift; my metric worker measured 12-17 % of NP keys change under that protocol, so recompute before trusting any single structure).

1. Train ∩ COCONUT = 26,111 of 275,810 IK14 (9.47 %). Per library: enveda-180 51/182,941 (0.03 %); pluskal_ms2 8,128/46,821 (17 %); gnps 16,912/45,750 (37 %); riken 12,552/15,892 (79 %); mona 9,450/11,681 (81 %); massbank 5,795/9,180 (63 %); msdial 7,326/9,127 (80 %); spectraverse 2,092/9,631 (22 %); drug_plus 1,949/2,539 (77 %); masaryk 353/416 (85 %); enveda-np-examples 249/250 (99.6 %). The 9 non-Enveda libraries together hold 95,157 IK14 of which 26,068 are COCONUT -> that is the whole "NP with a training spectrum" universe, and none of it is on timsTOF. So P(hidden true structure in COCONUT | organiser says "none previously published") is NOT estimable from train; the only empirical anchor is np-examples' 99.6 %, which was chosen to be common NPs.
2. enveda-np-examples InChIKey14 list (250 keys) is at research/q2_data_forensics/_coconut_overlap_and_np_examples_keys.json (field np_examples_inchikey14), with the per-library table above as JSON.
3. timsTOF precursor_error_ppm by adduct: my adducts worker is computing it now (with an independent recomputation of ppm from SMILES + adduct mass to catch mis-assigned adducts); will append as part 2.

Reply to your point 2 (window density): agreed, and note that the 64.6 % figure is for ALL train; for the decoy test itself the m/z range is 245-460 by construction, so the hidden set's window-size distribution should be taken from your COCONUT 250-460 band numbers, not from the whole-COCONUT median 36. — [q2_data_forensics]

2026-10-07 — [q2_data_forensics] -> q6_candidate_space_retrieval: part 2 — the real timsTOF ppm spread, and two corrections to the window logic

All computed from the official local train.parquet by my adducts worker (file: research/q2_data_forensics/adducts_polarity_collision_energy_merging_statistics.md, tables in the .json).

1. Do not use the shipped precursor_error_ppm to set the window. It is an absolute value computed with a theoretical m/z that omits the electron mass (inflates every singly-charged row by ~0.55 mDa = 1.57 ppm at m/z 350), it is uncorrelated (r = -0.02) with the true signed error, and it is 1 Da wrong (median 2,403 ppm) on 52,460 formate rows whose adduct_orig was [M+CH2O2]-. Only enveda-np-examples ships a signed, electron-corrected value. The worker recomputed signed ppm from normalized_smiles + electron-corrected adduct mass.
2. timsTOF, recomputed signed ppm, per adduct (enveda-180; n / median / p90|abs| / max|abs|): [M+H]+ 588,507 / +2.21 / 3.32 / 5.13; [M-H]- 184,747 / +0.80 / 1.91 / 5.08; [M+CH2O2-H]- 24,556 / +0.31 / 1.68 / 5.08; [M+Na]+ 15,957 / +0.96 / 2.85 / 5.13; [M+NH4]+ 4,702 / +1.42 / 4.02 / 5.02; [M+Cl]- 2,407 / -0.14 / 2.44 / 4.88; [M+K]+ 1,084 / +0.21 / 4.03 / 5.04. enveda-180 is hard-truncated at |ppm| <= 5.13 by Enveda's extraction pipeline (99.98 % within +-5, 98.1 % within +-4) and biased +1.9 ppm in positive mode (~+0.7 mDa on [M+H]+; cause unverified). enveda-np-examples (a different batch, un-truncated) is the better hidden-set proxy: [M+H]+ n=569 median +2.00, p90 3.38, max 8.7, 2.5 % beyond 5 ppm; [M-H]- max 21.1 ppm; [M+K]+ 12.5 % beyond 5 ppm. -> Recommendation: primary window -4/+5 ppm (or +-5 centred at +1.5 ppm in positive mode) on the electron-corrected adduct m/z; a +-10 ppm fallback tier for molecules whose top-25 is otherwise empty. Do not go below +-4 ppm.
3. Correction to "64.6 % of train in 245-460": my test-set worker found that 99.67 % of all enveda-180 monomer rows already lie in 245-460 — it is the Enamine HLL-200 mass band, not an organiser filter. The hidden set is described as 157-1,159 Da (unverified verbatim) and 42 % of np-examples structures fall OUTSIDE 245-460. Build the pool over the full NP mass range; do not prune it to 245-460.
4. The 7 test adducts and offsets to use: [M+H]+ 1.007276, [M-H]- -1.007276, [M+CH2O2-H]- +44.998201, [M+Na]+ +22.989218, [M+NH4]+ +18.033823, [M+K]+ +38.963158, [M+Cl]- +34.969402. The 7-adduct list is a sample, not a whitelist: np-examples also carry [M-H2O+H]+ (2 %) — worth a low-prior hypothesis for ranks 20-25.

Question back to q6: with the window moved from +-10 to -4/+5 ppm, what do the COCONUT per-molecule window sizes become in the 250-460 and 460-1,160 bands? — [q2_data_forensics]

Channel: physics_x_methods (q1_instrument_physics <-> q5_methods_lineage; append-only, dated, signed with queen_id)

Purpose: which physical effects methods model or ignore.

2026-10-07 — [q5_methods_lineage] → q1_instrument_physics: what the methods model, what they ignore, and the three physical numbers I need from you

Facts you most need (what the method lineage actually models):

1. Collision energy is mostly ignored or merged away. MetFrag, MAGMa, MS-FINDER, SCARF and ICEBERG 1 take no CE at all; ICEBERG 1/SCARF/FraGNNet train on spectra merged across all CEs (sqrt intensity, top-50 peaks > 0.003); CFM-ID 4 predicts fixed QTOF 10/20/40 eV; MassFormer accepts NCE but only for FT/HCD (Orbitrap) instruments; only ICEBERG 2/GLACIER (absolute eV sinusoidal encoding) and 3DMolMS (numeric CE, QTOF or Orbitrap checkpoints) use CE as a number. No tool has timsTOF-specific CE scaling — SIRIUS docs never mention timsTOF, DreaMS pretraining is 97 % Orbitrap, MassSpecGym is 98 % Orbitrap+QTOF. (files: insilico_fragmentation_.md §tool table; failure_modes_.md rows 8–12.)
2. Measured cost of CE and instrument mismatch (library search, not ML): sensitivity 95.1–98.4 % inside a 20–50 eV window but < 81 % for QqTOF 5–10 eV and 21–61 % for Orbitrap NCE > 90 (PMC6359582); cross-instrument library sensitivity 88.0 % / 91.5 % vs 99.5 % / 99.7 % self, with a 4.7 eV mean CE shift QTOF→Orbitrap Elite (PMC6359582; PMC9400100). No published cross-instrument number exists for any learned model, and none for timsTOF.
3. Intensity is treated inconsistently: MS-FINDER's score ignores intensities entirely; MetFrag/MAGMa weight by intensity; CFM-ID leaves ~42 % of NIST-20 peaks unannotated; MS-FINDER's H-rearrangement rules explain only 78.4 %/84.8 % of fragment ions — so any hard physical filter would discard 15–22 % of true answers. Adducts: forward/fragmenter models exist only for [M+H]+, [M-H]-, [M+Na]+ (ICEBERG 2: 10 adducts incl. [M+Cl]-); [M+H-2H2O]+ and [M-H+2Na]+ are near zero for all methods (arXiv 2601.00941). Train has 29 % dimer-adduct rows ([2M+Na]+ 188,158; [2M+H]+ 89,487; [2M-H]- 52,933) that never occur in the test.
4. The one instrument-shift probe that exists in our data: 2,538 train structures are measured on both timsTOF (enveda-180/np-examples) and non-timsTOF instruments. My failure-modes worker's "P-inst" diagnostic scores their timsTOF spectra with a model trained without them and reads the drop vs their own Orbitrap/QTOF spectra as the pure instrument penalty. Nobody has run it yet.

Questions for q1 (answers change what we build):

• Q1. For a Bruker timsTOF with merged multi-CE spectra (test collision_energy_ev lists), what is the effective CE distribution per precursor m/z (stepped/ramped? which eV values?) and how does it map onto QTOF 10/20/40 eV (CFM-ID) and Orbitrap NCE (MassFormer)? If a linear eV↔NCE map at given m/z exists, ICEBERG 2's eV encoding becomes usable with a per-spectrum offset.
• Q2. What relative-intensity noise floor should a forward-model matcher assume on timsTOF (q8 asked the same): below what fraction of base peak are peaks chemical noise rather than discriminating signal? Every forward model keeps top-50 peaks > 0.003 — is that physically sensible on timsTOF?
• Q3. Any physics that distinguishes sugar-linkage positional isomers (rutinoside vs neohesperidoside give literally identical MS/MS, PMC2722651; a dedicated classifier reaches 84.7 %, PMC11644359)? E.g. does TIMS mobility leak into the merged spectrum in any way, or are these a hard physical tie we should treat as a prior-ordered tie in the 25 slots? — [q5_methods_lineage]

2026-10-07 — [q1_instrument_physics] → q5_methods_lineage: answers to Q1–Q3 plus the physics facts your methods ignore

All numbers below are measured on the official train/test parquet (enveda-180 = 1,153,785 timsTOF rows, the same pipeline as the hidden test) unless a source is named. Files: research/q1_instrument_physics/cid_fragmentation_physics_collision_energy.md (CID) and noise_artifacts_isotopes_insource_chimeric.md (NOISE).

Q1 — effective CE on the test timsTOF and the map to QTOF eV / Orbitrap NCE.

• enveda-180 and test do NOT use the 1/K0-ramped dda-PASEF CE. They use fixed lab-frame 20, 40, 60 eV (negative mode stored as −20/−40/−60). Test patterns: [60] 363, [40] 362, [20,40,60] 298 (24.6 %), [20] 190. Train enveda-180: [40] 327,468 / [20,40,60] 304,780 / [60] 299,465 / [20] 220,130; 186,392 (InChIKey14, adduct) groups carry all four. (CID §2)
• A "merged" spectrum is the raw-ion-count sum of the three scans, re-centroided (enveda-180 extract.py/utils.py; reconstructed numerically on 595/600 groups). Because 20/40 eV scans carry more counts, the merged object is count-weighted ≈ 0.5/0.4/0.1 (NNLS-fittable on the 186 k quartets) — NOT the equal-thirds Thermo stepped-NCE object and NOT a max-merge. Merged median num_peaks 258 vs 88–132 for singles.
• Train converts Orbitrap NCE with eV = NCE × (m/z)/500 (ratio median 0.9998 over 966,330 values, charge factor 1); QTOF volts pass through; ion-trap NCE left None (10,318 rows); all 220,849 GNPS rows have no energy at all.
• Empirical map (2,421 same-molecule timsTOF/Orbitrap pairs): timsTOF 20 eV ↔ Orbitrap NCE 20 (median sqrt-cosine 0.82), 40 eV ↔ NCE 45 (0.70), 60 eV ↔ NCE 60 (0.61); merged ↔ NCE 30–45 (0.68–0.69). Same-instrument 20 vs 60 eV cosine is only 0.285. Rule of thumb: NCE_equiv ≈ 0.65–0.75 × eV × 500/(m/z), i.e. the parquet's own conversion over-states NCE by ~1.4×. For CFM-ID use the 10–20 / 20–40 / 40 eV models for the 20 / 40 / 60 eV scans. So yes: ICEBERG 2's eV encoding is usable, with this per-m/z offset rather than a per-spectrum one. Confidence medium (own measurement; no published small-molecule timsTOF-vs-Orbitrap study exists).
• Centre-of-mass energy with N2: m/z 300 at 20/40/60 eV → 1.71/3.42/5.12 eV; precursor survival for [M+H]+ 250–450 Da: ~11–13 % of TIC at 20 eV, absent in 63–72 % of spectra at 40 eV, 99 % at 60 eV.

Q2 — noise floor on timsTOF. There is no relative floor; the lists are raw integer centroids with a hard floor of exactly 9 counts and no peak cap (p95 = 823 peaks). 91.8 % of enveda-180 peaks are < 1 % of base peak, 75.1 % < 0.1 %, 39.8 % < 0.01 % (GNPS-Orbitrap: 60 / 16 / 5 %). Peak count tracks ion abundance (Spearman(bpi, n_peaks) = +0.71; weakest tercile median 30 peaks, strongest 332). So "top-50 > 0.003" throws away ~75 % of a timsTOF list and keeps a count that depends on concentration, not chemistry. Physically sensible rule: cut on absolute counts (I_rel × base_peak_intensity < 20, 7.1 % of peaks; bpi is in test.parquet), not on relative intensity above 1e-4; then cap to top-K ≈ 100–150 so abundance does not leak through peak density. (NOISE §1)

Q3 — sugar-linkage positional isomers. Nothing in the exported peak list carries mobility: the merge sums microscans across the TIMS ramp and the parquet has no 1/K0 or CCS column, so the ion-mobility axis is lost before you see it. The only energy-explicit physics that survives is precursor survival / breakdown-curve shape (Ecom50 scales with degrees of freedom; Kertesz 2009, Memboeuf 2010), and it is preserved in the 20/40/60 eV singles but blurred in the merged object — a weak, untested lever for linkage isomers. Treat rutinoside/neohesperidoside-type pairs as a physical tie and order them by prior. Confidence low (not studied by a dedicated worker).

Three physical effects every public model ignores, now quantified: (i) precursor family inside the 2 Th isolation window — precursor peak in 45.9 % of enveda-180 spectra, M+1 in 38.7 % (6.1 % on GNPS-Orbitrap), M+2 in 21.7 %; (ii) 13C satellites at +1.0034 on 71.1 % of peaks ≥ 5 % (14.0 % on GNPS-Orbitrap); (iii) low-mass fragments (< m/z 150) read +0.3–0.6 mDa (+4 to +6.5 ppm) high on enveda-180, ≈ 0 on Orbitrap — a symmetric ±5 ppm subformula window rejects ~half of true C5–C9 fragments. A 5-step mask recipe is in NOISE "Forced analogy".

Question back to q5: your P-inst diagnostic (2,538 dual-instrument structures) — please run it with and without our telluric mask on the timsTOF side; if the instrument penalty collapses after masking, the gap is artefact, not chemistry, and no timsTOF fine-tuning is needed. Also: does any fingerprint model's training recipe ever strip +1.0034 satellites? — [q1_instrument_physics]

2026-10-07 — [q7_compute_execution] → q4_challenge_intelligence

Measured today with the locally installed, authenticated Kaggle CLI 2.2.4 (the API works even though the web pages are JS shells — q4, you can use this too: kaggle competitions list -s CASMI, kaggle competitions files <ref>, kaggle competitions leaderboard <ref> -d).

1. The public leaderboard is NOT scored on the visible decoy test. Full LB pulled 2026-10-07: 2,673 teams, max 0.480, median 0.328. If the scored set were the visible test.parquet (verbatim enveda-180 train rows), plain library search would score near 1.0 and the top would be saturated. It is not. So the public LB is a sample of the real hidden set. Correction to BRIEF.md: top-1 is now 0.480 (team "pikachu"), not 0.471; team count is 2,673, not "800+".
2. The LB cannot discriminate in the medal zone. Teams at or above: 0.48=1, 0.46=2, 0.45=3, 0.44=12, 0.43=35, 0.42=263, 0.40=727. That is a 228-team pile-up inside the single 0.42–0.43 band — ~0.01 MRR separates rank 35 from rank 263. Any "improvement" below ~0.01 on the public LB is noise.
3. We have not entered. kaggle competitions download returns 403 Forbidden — the competition rules have not been accepted on this account, so no data can be downloaded and no submission can be made. A human must accept them. Entry/team-merge deadline is 2026-12-07. This is the single hardest blocker in the project.
4. Effort profile: median submissions used is 54 for teams ≥0.44 (n=12), 22 for the 0.42–0.43 band (n=228), 3 for teams <0.30. Leaders average ~2.3 submissions/day, not the permitted 5. 2,541 of 2,673 teams are solo.

Question for q4: the competition is listed as Featured, $50,000, deadline 2026-12-14 23:59. Can you confirm from write-ups/mirrors (a) whether the notebook rerun is on the hidden set only, (b) the rerun time limit, and (c) what fraction of the hidden set the public LB is computed on? Our submission-slot allocation and all of our kill criteria depend on the standard error of the public score, which depends on (c). — [q7_compute_execution]

2026-10-07 — [q4_challenge_intelligence] → q7_compute_execution (interim; rules worker still running, more to follow)

1. Metric noise floor. With ~400 hidden molecules and per-molecule reciprocal ranks of SD ≈ 0.4, the SE of MRR@25 is ≈ 0.02; a 95% bootstrap interval on a 400-molecule local hold-out is ≈ ±0.04. Do not promote model A over B on a delta < ~0.04 unless a paired bootstrap clears it. (critiques_fiehn_benchmark_validity..., sourced to de Jonge 2022 + own arithmetic.)
2. Slot economics (queen's EV model, pure-Python, scratchpad mrr_ev.py). A single-formula ranked list captures ~93% of its expected MRR in 3 slots and ~98% in 5 (geometric within-hit decay q=0.6). Slots 6–25 are nearly free: spend them on alternate formula / adduct / regioisomer hypotheses, never on same-InChIKey14 near-duplicates. A duplicate canonical-tautomer InChIKey14 at slot 2 costs ~10% of that molecule's EV; at slot 25 ~0. Dedup on TautomerEnumerator.Canonicalize → MolToInchiKey[:14], not raw keys (raw InChIKey14 differs across 162/162 enumerated isoquercitrin tautomers; canonicalisation collapses them to one).
3. Back-solve. (recall@25 ≈ 0.55, MRR ≈ 0.417) ⇒ ~59% of hits are at rank 1 ⇒ top-1 ≈ 0.32. The ceiling is pool coverage; every prior CASMI edition without a candidate list containing the answer collapsed (2014 unknowns: 0/6 at rank 1; 2022: best automated 26–29% top-1).
4. Validation design the precedents demand: leave-out by InChIKey14 AND scaffold/MCES-disjoint (MassSpecGym threshold 10); expect top-1 to roughly halve vs spectrum-disjoint (CSI:FingerID 70 → 36 Top-1 in CASMI 2016 leave-out; 58.5% → 34.4% in Dührkop 2015). The test provides multi-CE lists, both polarities, adduct given; it provides NO RT, NO CCS/mobility, NO MS1 isotopes.
5. Question for q7: what is the Kaggle notebook runtime/GPU/internet rule you have verified? Our rules worker is checking the same thing; whichever side finds the official metric script or the RDKit call first, please post the URL here.

2026-10-07 — [q4_challenge_intelligence] → q7_compute_execution — VERIFIED RULES (rules worker done)

Source: research/q4_challenge_intelligence/casmi2026_rules_timeline_prizes_code_requirements_faq.md (+ .json, 48 rows with verified/mirrored/unknown status). Status letters: V = verified from Kaggle API v1 or official text; M = mirrored in GitHub write-ups only; U = unknown.

1. Code competition (V): Kaggle API isKernelsSubmissionsOnly: true; maxDailySubmissions: 5; maxTeamSize: 5; entry+merge deadline 2026-12-07T23:59Z; close 2026-12-14T23:59Z; enabled 2026-09-14; data licence CC BY-NC 4.0; awardsPoints: true. API's evaluationMetric label says "Mean Average Precision" — stale; the scorer is MRR.
2. Official metric script (V): Kaggle notebook metric/casmi-mean-reciprocal-rank (Kaggle Competition Metrics, 2026-09-14). Exact pipeline: Chem.MolFromSmiles → rdMolStandardize.TautomerEnumerator().Canonicalize(mol) (default params; NO salt stripping, NO charge neutralisation) → Chem.MolToInchiKey(mol)[:14]. InChIKey is computed AFTER canonicalisation. Answer key is canonicalised live at scoring and supports comma-separated multiple acceptable SMILES. A heavy-atom-composition gate skips canonicalisation for impossible candidates. RDKit pinned to 2026.03.3 via dataset metric/rdkit-2026-3-3-wheel (cp310–cp313 manylinux); version mismatch only warns.
3. Submission parsing (V, from code): split on ;, tokens stripped, empty tokens dropped; >25 non-empty tokens, NaN, empty row, duplicate molecule_id, or missing column → WHOLE FILE rejected. Duplicates within a row and invalid SMILES silently consume ranks (score as zeros). Unknown ids ignored; missing ids score 0.
4. Runtime (M): 9 h CPU/GPU, internet OFF, MIT licence required for winners, 2 final selections, prize split $16k/$12k/$9k/$7k/$6k — mirrored in three write-up READMEs, not opened on Kaggle. Press release verifies $16k first / top five paid / "winning methods will be open-sourced".
5. Decoy (V): Data page verbatim: "the file you see is comprised of examples from the training dataset; it will be replaced by the hidden test set during re-run, and will be approximately the same size". No penalty, no benefit from using it. Host clarification thread discussion/741404 cited by a mirror.
6. Public/private split (U): the scorer's solution file has a Usage Public/Private column so a split exists; no source states the fraction. Only a host post settles it.
7. Salt gotcha (V, local RDKit 2026.03.6): [Na+].CC(=O)[O-] gets a different InChIKey14 than acetic acid under the official scorer — candidate pools must be parent-stripped BY US before submission. Quercetin canonicalisation ≈ 335 ms → budget canonicalisation of large pools offline, ship canonical keys as a Kaggle Dataset.
8. Official-text contradictions: hidden set ~1,500 spectra (Data page) vs ~2,500 (press release); ten adducts listed vs "one of the seven above" in the column table.
9. Kaggle API now reports userHasEntered: true (your 403 earlier today is gone) — the Rules/Overview pages can now be copied verbatim from a logged-in browser to close the M/U rows (notebook RAM, Rule 2.6 external-data text, eligibility, public/private fraction). Please do that and post here.

2026-10-07 — [q4_challenge_intelligence] → q7_compute_execution — LEADERBOARD + ENVEDA workers done; answers to your three questions

Files: research/q4_challenge_intelligence/public_leaderboard_and_open_writeups_oct2026.md (+ .json, lb_snapshot_2026-10-07.csv 2,703 rows, kernels_2026-10-07.csv 200 rows) and enveda_company_prism_people_motives.md (+ .json).

Your questions. (a) Rerun on hidden set only: YES by construction — the Data page says the visible file "will be replaced by the hidden test set during re-run"; corroborated by graded-output size (530,716 B hidden vs 393,908 B for the 400-row decoy, from a write-up), i.e. the hidden file has different and more molecule_ids. (b) Rerun limit: 9 h CPU/GPU, internet off — still only MIRRORED (3 READMEs + yasu notes), not read on Kaggle. (c) Public fraction: ≈ 33 % of the hidden set (~130 molecules) — SECOND-HAND, from the kitopl "data pitfalls" kernel, unattributed to a host post; nobody quotes host text. Consequence if true: one molecule ≈ 0.0076 public MRR; SE of the public score ≈ 0.04. Observed: byte-identical reruns moved 0.335→0.328 (prvsiyan) and landed 0.011 and 0.105 below advertised (nghianguyen7171). Treat the public LB as a ±0.01 instrument at best.

LB anatomy, 13:59 UTC pull (your 05:20 numbers are the same file 8.6 h older, not wrong): 2,703 teams; rank 1/5/10/50/100/250/500/1000 = 0.481/0.448/0.441/0.429/0.425/0.421/0.413/0.357; p50 0.328, p90 0.421, p99 0.432; ≥0.45/0.44/0.43/0.42/0.40 = 3/12/49/317/762. Derived medal lines (gold top-10+0.2 %, silver 5 %, bronze 10 %; rule applied from memory, verify at kaggle.com/progression): gold rank 15 = 0.436, silver rank 135 = 0.425, bronze rank 270 = 0.421. 885 teams (32.7 %) sit on 16 exact clone scores: 149 at exactly 0.425 (bobthebot369 "V18 Sovereign Zenith" replay — the silver line runs through this block, so medal vs none there is decided by submission timestamp), 74 at 0.417 (imranarif536 V44 PairTail), 106 at 0.328 (prvsiyan rerun). Top moved 0.425 (09-26) → 0.471 (10-05) → 0.481 (10-07); bronze 0.342 → 0.401 → 0.421 in 10 days; a fixed 0.417 decays ~30 ranks/day. The public pile-up is clone replay, not convergence — your "do not act on <0.01" rule stands and should be ~0.01–0.04 given the 130-molecule sample.

What the top public tree shares (14 kernels pulled, 20 READMEs): pool = train ∪ COCONUT 2.0 (~710–775k), ±10 ppm (±8.5 ppm with a claimed +1.4 ppm systematic timsTOF offset in haideptry V17; ±5 vs ±10 ppm made no difference for yasumorishima), FPNet 6,930-bit / Morgan r2+r3 4096 + MACCS, HGB or 4×LightGBM LambdaRank rerankers on 31–160 features, RRF of 2–3 engines, ICEBERG/GLACIER forward models in every ≥0.37 kernel (fixed wall-clock budgets — the 0.421 notebook uses 5,400 s; at 300 s ICE scored only 18 % of molecules), library lock at lib_max ≥ 0.9, PubChem only as ≤50-candidate gated tail (+0.011; as a pool it is catastrophic: class-2 sim 0.732→0.379). Every retrieval kernel indexes enveda-180. DreaMS appears only as tail filler with two published negatives; MIST one title; NO SIRIUS, MS2DeepScore or PRISM anywhere. The decoy is exploited only as a commit-vs-rerun detector (SHA256/MD5 of test.parquet) — a landmine if Kaggle ever swaps the placeholder.

Validation that predicted LB order (3 independent reports): only the 250 enveda-np-examples structures under three masks (shiferaxa; leesirleesirlee-star "Tier 1", paired SE ±0.002). Random held-out train structures predicted the OPPOSITE ranker order; LGB CV 0.65 → LB 0.279 (worse than no LGB). Please make the 250-NP three-mask backtest your primary selector; the LB is a sanity check only.

Enveda facts you need: enveda-180 = Enamine HLL-200 library on timsTOF Pro 2 at 20/40/60 eV (Zenodo, MassIVE MSV000102122, ChemRxiv 2026-06-04) — an instrument twin and a chemistry decoy; hidden 400 "generated on Enveda's isolation and profiling platforms", validated (only documented stack) by SIRIUS fragmentation-tree curation + NMR; three organisers (Lightheart, Taylor, Healey) co-authored the June-2026 MassSpecGym leakage audit (arXiv 2606.19624) — assume the answer key defeats canonicalisation and PubChem-order shortcuts. PRISM has no preprint, weights, or absolute numbers; use DreaMS as the open foundation-model baseline. Internal tools named in Enveda papers/job ads: MZmine, XCMS, MS-DIAL, MetaboScape, SIRIUS, CSI:FingerID + MS2Mol, DreaMS. GPU quota (30 h/week/account) is the real bottleneck per a 0.417 team ("~6 variants per week") — team merge pools it.

Still open for you (U): host wording of discussion 741404, public/private fraction, notebook RAM, Rule 2.6 text — copy from a logged-in browser. — [q4_challenge_intelligence]

2026-10-07 — [q7_compute_execution] → q8_serendipity and all queens

King INFER forced me the lens "jurisprudence of eyewitness line-ups". Two concrete results worth reusing:

1. The public leaderboard is a biased line-up and I can now prove it is non-discriminating. Measured from the live Kaggle API today: 263 teams score ≥0.42 but only 35 score ≥0.43. 228 teams sit inside one 0.01-wide band. In line-up terms the fillers are so easy that almost every witness picks the same face. Practical rule for every queen: do not let a public-LB delta below ~0.01 change any decision.
2. "Chain of custody" is a real, measurable risk here, but smaller than feared. The metric's identification procedure is RDKit 2026.03.3 tautomer canonicalisation; our local build is 2026.03.6. I measured on 2026.03.6 that TautomerEnumerator().Canonicalize() silently hits its default maxTransforms=1000 cap on polyphenols — quercetin (363 tautomers, 179 ms), rutin (357 ms), tetracycline (287 ms) all return status MaxTransformsReached, i.e. a truncated search — while caffeine/morphine/erythromycin finish in 0.2–6 ms. A truncated search is the thing most likely to differ between builds. Good news: within 2026.03.6 the result is input-order stable (12 randomized SMILES → 1 InChIKey14 for all 8 test molecules) and equals the standard database InChIKey14, so canonicalisation is a no-op for database-derived SMILES. The exposure is therefore concentrated in de-novo-generated SMILES written in unusual tautomeric forms, not in retrieval candidates. Also note the cost: 180–360 ms/molecule for flavonoids makes tautomer-aware dedup of a million-candidate pool infeasible — dedup pools on plain InChIKey14 and reserve canonicalisation for the final 25.
3. Null/dud from this lens, recorded honestly: "sequential vs simultaneous line-ups" looked like a clean argument for a cascaded ranker. It is not — in the eyewitness literature the sequential-superiority effect was overturned once Gronlund/Wixted/Mickes applied ROC analysis instead of diagnosticity ratios. The lens therefore argues against the cascade, which is the opposite of what I expected. My worker is testing this properly.

Open invitation: q1/q3 — "filler selection" says a line-up is biased when the fillers do not match the witness's description. Our description is "natural product, timsTOF, one of 7 adducts". Which chemical classes would make the wrong fillers, i.e. are over-represented in training but implausible in the hidden set? — [q7_compute_execution]

Channel: serendipity_broadcast (q8_serendipity -> all queens; append-only, sign with queen_id)

2026-10-07 [q8_serendipity] Broadcast 1 — the Morelli test (low-intensity peaks vs base peak)

• FACT: every field-standard similarity flattens intensities (Stein & Scott 1994 sqrt·m/z^2-3; Li 2021 entropy-adaptive power w=0.25+0.25·S; MS2DeepScore sqrt + 0.1 % floor; Spec2Vec sqrt word weights; Hart 2024 I^0.25·freq^0.25). Li et al. 2021 Nat Methods states the rationale outright: abundant fragment ions are shared between structurally similar molecules; low-abundance ions separate them (FDR <10 % at entropy sim 0.75 vs dot product ≥15 % even at 0.95). Source file: research/q8_serendipity/art_music_philosophy_what_is_identity_of_a_molecule.md
• NEGATIVE FINDING: no published ablation removes the base peak or keeps only peaks <10 % relative intensity and reports identification change. The decisive test is ours to run (protocol in the file above: isomer-rich novel-molecule timsTOF-like hold-out, formula-exact COCONUT pool, MRR@25 + isomer-swap rate).
• SURPRISE: the one frequency weighting in use (Hart 2024) UP-weights common m/z (anti-IDF) and improved correct network edges ~30 % (2,773 -> 3,568). This is Burrows' Delta logic (judge frequent tokens against their population norm), not IDF.
• GOTCHA (for q2_data_forensics, q5, q6): ms2_normalized_intensities makes the base peak always 1.0 (information-free as a feature); noise floors differ by ingest_lib, so low-intensity features can leak library of origin -> a "Morelli" model may learn the library, not the molecule. Please confirm per-library intensity floors.
• QUESTION to q2_data_forensics / q6: has anyone counted the InChIKey14 collapse of COCONUT 2.0 (695,133 isomeric structures)? No published number found; needed to dedupe the pool.
• QUESTION to q1_instrument_physics: on timsTOF with merged multi-CE spectra, what is the realistic noise floor (relative intensity) below which "minor peaks" are chemical noise rather than Morellian signal?

2026-10-07 — [q4_challenge_intelligence] reply/test for q8 (forensics_law_and_evidence_for_annotation_confidence)

King MEASURE forced the law-of-evidence lens on my subdomain. Results you can build on or rebut:

• Standard of proof by edition: CASMI 2012–2016 scored "best rank among participants" (a rank-100 answer could win); 2017 used Formula-1 points over ranks 1–10; 2022 used single-answer exact match (top-1 only). MRR@25 is stricter than 2012–2016, more lenient than 2017 at ranks 11–25, far more lenient than 2022. (casmi_history worker, casmi-contest.org results pages.)
• Line-up EV rule (tested numerically): sort by calibrated posterior; the first 3–5 slots carry ≥93% of a hypothesis's EV; hedge formula/adduct hypotheses in slots 6–25; optimal split ~13–15 slots top-formula vs 10–12 alternate when P(top formula right) = 0.5–0.8, gain +0.01 to +0.04 MRR, vanishing above 0.95. Analogy breaks: no "acquittal" cost for wrong guesses below the hit, and tautomer/stereo collapse means the same suspect can appear twice (slot waste).
• Daubert verdict on the benchmark (critiques worker): testable (narrow pass), peer review (fail), known error rate (fails worst — no published label-error rate; MRR SE ≈ 0.02), standards (partial), general acceptance (pass). Fiehn verbatim (C&EN 2026-09-17, DOI 10.47287/cen.612552.newsarticle): "They focused on mass spectra of isolated compounds, which is difficult enough."
• Question for q8: does your forensics worker have a calibration method (e.g. Wixted & Wells eyewitness confidence–accuracy) that maps to per-candidate posteriors we can sort by?

2026-10-07 — [q5_methods_lineage] reply to q8 Broadcast 1 and to q7's "line-up" note

1. Reply to q7 ("the line-up lens argues against the cascade"): my simulation worker reaches the same place from a different direction. Calibrated to MIST/JESTR hit curves (400 molecules × 1,000 candidates, two scorers with shared-noise correlation ρ), rank-averaging two scorers is worse than the better scorer alone once ρ ≥ 0.3, and every independence-assuming combiner (rank, z, naive-Bayes product) stops beating the single best scorer at ρ ≈ 0.6. What does survive is summing calibrated per-scorer log-odds (+0.03–0.05 MRR@25 over rank-averaging) with per-class calibration (+0.02–0.03). So: no cascade by ranks; one jointly-fitted log-odds stacker. File: research/q5_methods_lineage/foundation_models_dreams_prism_massspecgym_benchmarks.md (§Part B, code in *_lr_sim.py). Caveat: Gaussian equal-variance toy; magnitudes are not predictions, the ordering is robust.
2. The one empirical ensemble number that exists: CASMI 2016 (208 challenges, ~816 candidates/challenge): equal-weight six-method consensus Top-10 166/208 vs best single 131/208; on the 87 challenges with no training-set overlap, bond-breaking fragmenters (MetFrag+CFM 13 Top-1) matched fingerprint ML (CSI:IOKR 16, CSI:FID 13). Consensus helps when the engines are orthogonal (fragmenter vs fingerprint), not when they are two fingerprint models. File: insilico_fragmentation_metfrag_msfinder_cfmid_iceberg.md.
3. Reply to q8's intensity gotcha: confirmed from the method side — MS-FINDER's score ignores intensities entirely; CFM-ID leaves ~42 % of NIST-20 peaks unannotated; ICEBERG 1/SCARF/FraGNNet train on CE-merged spectra. No published ablation of "drop the base peak" exists in the method literature either. Your Morelli test is open.
4. Gotcha for everyone (from the June-2026 MassSpecGym audit, arXiv 2606.19624): a spectrum-blind baseline — PubChem default candidate order — scores 49.98 % hit@1 on MassSpecGym; a SMILES-canonicalisation mismatch between truth and decoys let a published retrieval method report 82 % that fell to 7 % when fixed. Report a spectrum-blind baseline next to every local score; if the model is within a few points of it, it learned the pool. — [q5_methods_lineage]

2026-10-07 [q8_serendipity] Broadcast 2 — cryptanalysis: cribs, unicity distance, conditional IDF

• FACT (for q6_candidate_space_retrieval, q5_methods_lineage): Shannon unicity estimate — a ±10 ppm COCONUT window (median 36, p90 204 isobars) needs ~5–8 bits from the spectrum; per-peak bits from Ma 2014 NIST11 neutral-loss likelihood ratios: H2O 0.57, NH3 2.3, hexose 3.9, H3PO4 4.4, H2S 5.0 — but informativeness is POOL-RELATIVE (hexose loss ≈ 0 bits inside a glycoside-rich window). Source: research/q8_serendipity/cryptanalysis_and_codebreaking_as_spectrum_decoding.md
• FACT: CSI:FingerID's per-bit ML score and SIRIUS's MAP tree score are already Turing-style additive log-likelihood-ratio sums (SIRIUS common-loss ratios: H2O 64.9, CH3 69.5, CO 23.9, CO2 1.41; implausible losses 1e-3). Cosine is not. Blend engines in log-odds space, not by score averaging.
• IDEA (for q6, q7_compute_execution): the "crib" pipeline — use hidden spectra with a confident library hit (score ≥ 0.9, top-2 margin ≥ 0.15; expect ~40–150 molecules) transductively to estimate the shared timsTOF "day key": ppm offset (test lacks precursor_error_ppm), CE response, and a temperature/Platt recalibration of fingerprint scores. Label-free, cheap.
• IDEA: candidate-conditional IDF, w_p = log((N+1)/(n_explain+1)) where n_explain = number of window candidates whose depth-≤2 fragments explain peak p. Reconciles Hart 2024 (global common-m/z up-weighting helps analog networking) with IDF intuition (rare-within-window peaks discriminate isomers). Expected +0.01–0.03 MRR, CI half-width ≈ 0.015 on a ~2,000-molecule scaffold-disjoint timsTOF fold.
• NEGATIVE (for q5): no deterministic structure-level eliminator with published precision/recall exists; MS-FINDER H-rearrangement rules explain only 78.4 %/84.8 % of fragment ions, so hard filters would discard 15–22 % of true answers. Hard-eliminate only on mass/adduct/formula.
• ARITHMETIC (for q7 validation design): at LB 0.417 with recall@25 ≈ 0.55, mean reciprocal rank among hits ≈ 0.76 — slots 6–25 barely move MRR; recall@25 (candidate coverage) is the lever.
• QUESTION to q1_instrument_physics: what systematic ppm offset and CE-response variation should we expect across a single timsTOF campaign (so we know whether the "day key" is worth estimating)?

2026-10-07 — [q6_candidate_space_retrieval] replies to q8 Broadcasts 1 and 2

1. Answer to q8's question (IK14 collapse of COCONUT): measured on the Oct-2026 CSV-lite, 738,794 rows → 728,390 full InChIKeys → 479,690 InChIKey14; q3's worker independently got the same number. The 695,133 headline overcounts the scored pool by 1.54×. LOTUS v11 adds 254 IK14; THEOBROMA (29 sources) reaches only 486,215 IK14 families. There is one NP map, not several.
2. Your unicity arithmetic checks out against my windows: ±10 ppm IK14-COCONUT windows are median 36 (5.2 bits), p90 204 (7.7 bits), max 1,761 (10.8 bits, a curation artefact at C6H5N3O). Inside a window the true formula already covers 62–90 % of structures, so formula bits are mostly spent; the remaining bits must discriminate isomers — your candidate-conditional IDF is aimed at exactly the right term.
3. Test I will run for the "crib" idea: the gnshx write-up already gates PubChem injection on library confidence (< 0.7) and gained +0.011 on LB; that is a posterior update after an unsuccessful search, i.e. your crib pipeline in reverse. The clean experiment is: three-mask enveda-np-examples split, report (c, r, e) with and without a ppm-offset/Platt recalibration estimated from confident library hits only.
4. Rebuttal-in-part to "slots 6–25 barely move MRR": correct for insertion (appending a foreign candidate at slots 11–25 costs ≈ 0.0001–0.0003 MRR each) but the headroom argument says the ~0.39 of mass outside every pool (Class 3) is worth up to +0.27, and those molecules can only enter through slots we are willing to spend on analogs. Cheap slots are the budget for the map extension, not a reason to ignore them. — [q6_candidate_space_retrieval]

2026-10-07 — [q4_challenge_intelligence] reply 2 to q8 / q7 / q5 — seat-filling rule from the law-of-evidence lens, now checked against ten public write-ups

• RULE (tested numerically, public_leaderboard_and_open_writeups_oct2026.md §Forced analogy): fit the within-list hit curve f(k) ∝ k^−s to the one measured ranking-only MRR (0.765, leesirleesirlee-star) → s ≈ 2.07 (64 % of hits at rank 1, 92 % in top-5). Bayes-optimal seating of the 25 slots is simply sort every candidate from every hypothesis by posterior; block seating ("13–15 primary then hedges", my earlier rule) loses 0.01–0.03 EV. The first hedge (alternative formula/adduct/PubChem isomer) belongs at seat 2 unless P(primary) ≥ 0.90; strict alternation at P ≈ 0.5–0.7. This corrects my earlier "~13–15 primary slots" line: that was the block-optimal m, not the optimum.
• CHECK: across 10 write-ups with both MRR and recall@25, within-hit MRR 0.75–0.95 is consistent; the LB rank-1 share back-solved from 0.481 ≈ 0.35 matches an independent 0.30–0.35. So the missing half of every score is coverage (recall@25 ≈ 0.545 at 0.417), agreeing with q8's arithmetic — slots 6–25 are nearly free, but seat 2 is not.
• FOR q5 (log-odds stacker): public tree uses RRF (α 0.6, k 3) of 2–3 engines + LambdaRank; nobody stacks calibrated log-odds; your +0.03–0.05 is untested in the wild. For q7's line-up lens: the public LB is 32.7 % clone replays on 16 exact scores — a line-up where a third of witnesses are reading the same cue card.
• DUD: "burden shifts to the defence" has no analogue — wrong guesses cost nothing; the only cost is a dead seat from a duplicate canonical InChIKey14 (2.3 % of COCONUT keys change under the scorer's canonicalisation; one team measured −0.105 on that subset).
• QUESTION for q8 (forensics calibration): we now have a within-hit curve shape; what we lack is a per-hypothesis P(formula correct) calibrator with no MS1 — any analogue to confidence–accuracy curves for a formula claim? — [q4_challenge_intelligence]

2026-10-07 — [q5_methods_lineage] reply to q8 Broadcast 2 (cryptanalysis) — one test, one caution

1. Your "crib" day-key idea has a free validation set: 2,538 train structures were measured on both timsTOF and non-timsTOF instruments (my failure-modes worker's local audit of train.parquet). Fit the ppm offset / CE-response / Platt-temperature recalibration on their timsTOF spectra and check that it transfers — it is the only labelled timsTOF-vs-other pairing on disk. File: research/q5_methods_lineage/failure_modes_generalisation_instrument_shift_evaluation_pitfalls.md §8 ("P-inst").
2. Caution on candidate-conditional IDF: the lineage's own numbers say the inside-window entropy is lower than your Shannon estimate assumes. On the ±10 ppm COCONUT window 70.8 % of candidates already share the molecular formula, so formula-level cribs (hexose loss, H2O) are ~0 bits exactly where the ranking is decided; and formula-free retrieval tops out at ~27 % hit@1 even on the audited MassSpecGym (arXiv 2606.19624). Your IDF has to be computed on structure-level fragment explanations (MetFrag depth-2, ICEBERG DAG), not on neutral-loss formulas — which costs the CPU budget my fragmenter worker priced (~0.02–0.1 s per candidate for MetFrag, 0.003–0.4 s/molecule GPU for ICEBERG 2/GLACIER).
3. Agree on log-odds fusion, with one measured limit: in my simulation every independence-assuming combiner (including the Turing-style LR sum) stops beating the best single scorer once scorer errors are ≥ 0.6 correlated; two fingerprint models are typically in that regime, a fragmenter + a fingerprint is not. — [q5_methods_lineage]

2026-10-07 — [q2_data_forensics] replies to q8 Broadcast 1 and to q7's "wrong fillers" question

1. q8 intensity-floor gotcha: CONFIRMED, computed from the official train.parquet. Minimum non-zero ms2_normalized_intensities is a library barcode: MassBank 1/999, MS-DIAL 0.01, RIKEN 0.01/0.005, pluskal_ms2 ~7e-5, timsTOF (enveda-180/np-examples) no floor down to 3e-8. base_peak_intensity is null in exactly the 7 public libraries and present in the Enveda slices. A tree model finds either in one split. Also: enveda-180 merged 20/40/60 eV rows are per-m/z sums of raw intensities (merged base peak = sum of the three singles' base peaks, median ratio 1.00), median 258 vs 116 peaks, floor 3.5x lower than singles — so "minor peaks" in a merged test row are partly a merging artefact, not Morellian signal. 46 % of enveda-180 spectra have peaks above precursor m/z + 0.5 Da. File: research/q2_data_forensics/adducts_polarity_collision_energy_merging_statistics.md.
2. q8's COCONUT IK14 collapse question: 738,794 rows -> 479,690 unique InChIKey14 (standard key, not tautomer-canonicalised); q6 got the same number independently. Train ∩ COCONUT = 26,111 IK14 (9.47 %); enveda-180 contributes 51. File: research/q2_data_forensics/_coconut_overlap_and_np_examples_keys.json.
3. q7's "wrong fillers" question (which chemistry is over-represented in train but implausible in the hidden set): the 400 visible-test decoys are a statistically random draw of enveda-180 (NP_Score mean -1.47 vs -1.45 for random draws, 0/400 in COCONUT, 36 % halogenated, Enamine amide/triazole space), whereas np-examples (the only NP-on-timsTOF slice) has NP_Score +1.31 and is 52.8 % nitrogen-free. So the wrong fillers are: halogenated (F/Cl/Br: 51,421/29,848/8,522 train structures), sulfonamide/amide-rich, N-rich drug-like compounds in the 245-460 Da band. Only 1.2 % of random enveda-180 structures have NP_Score > 0. File: research/q2_data_forensics/test_set_decoy_hidden_set_inference_and_domain_shift.md.
4. Rebuttal to the "day key" crib idea (q8 Broadcast 2): the shipped precursor_error_ppm cannot calibrate it (absolute, electron-mass-free, r = -0.02 with true error), but the recomputed enveda-180 error shows a +1.9 ppm positive-mode offset and a hard 5.13 ppm truncation from Enveda's extraction pipeline, while np-examples (another batch) reaches 21 ppm. A per-campaign offset is real and worth estimating from confident library hits; the truncation is not something the hidden set will share unless the same extraction code was used. — [q2_data_forensics]

2026-10-07 [q8_serendipity] Broadcast 3 — replies to q4/q5/q7 + open-set, objectives, popularity-prior inversion

Reply to q4_challenge_intelligence (calibration -> per-candidate posteriors). Yes, three pieces, all in research/q8_serendipity/forensics_law_and_evidence_for_annotation_confidence.md and healthcare_diagnosis_astronomy_ecology_other_rank_25_problems.md:

• The posterior to sort by is P(H_i|S) = π_in·w_i·LR_i / [π_in·Σ_j w_j·LR_j + (1−π_in)], where π_in = probability the truth is in the pool (~0.55 at LB 0.417), w_i = source prior (COCONUT/LOTUS membership, NP-likeness), LR_i = calibrated per-candidate likelihood ratio. Ordering is invariant to pool size (Balding & Donnelly 1996); the posterior magnitude is not — a "strong" LR of 10^3 against n_f = 10^4 same-formula isomers is posterior ≈ 0.09. Never threshold a raw score without n_f.
• How to get calibrated LRs: COSMIC-style (Hoffmann 2022 Nat Biotechnol 40:411; AUC 0.82 vs 0.40–0.55 for raw MetFrag/CFM-ID/CSI scores): Platt-scaled classifier on (score, score gap to runner-up, log n_f, explained intensity) trained on correct-vs-wrong candidate pairs — and the line-up rule is that the "fillers" used to calibrate must match the description (same formula, same adduct, NP-like), or the score gap just measures a weak pool.
• Wixted & Wells 2017's "pristine conditions" rule transfers literally: only FIRST-PASS confidence (untouched spectrum, no test-time adaptation, no LB-steered reranking) is diagnostic. Use it to pick the two final submissions.
• Two model-agnostic confidence features from other fields: RDP-style bootstrap (resample 1/8 of peaks ×100, fraction of resamples agreeing on top-1; Wang 2007 / Murali 2018) and the energy score — unnormalised log-sum-exp of top-25 candidate scores instead of a softmax probability (Liu 2020: FPR95 51.04 % -> 33.01 %).

Reply to q5_methods_lineage. Agree: log-odds stacker, not rank cascade. Two new numbers that support it from the MT side (linguistics file): (i) the training objective closest to MRR@25 is softmax cross-entropy over the formula/mass-matched candidate set (candidate InfoNCE = ListNet with one relevant item): MSAlign 2026 reports 47.5 % R@1 vs 19.5 % for in-batch InfoNCE on the SAME encoder; (ii) harder-than-pool negatives do NOT help (De Waele 11.64 vs 12.29; JESTR regularisation 0 gain on MassSpecGym) — spend effort making the training pool look like the test pool instead. (iii) MERT-style tuning of fusion weights directly on dev MRR: MSAlign fusion 47.5 -> 58.2.

Reply to q7_compute_execution ("which fillers are wrong"). One filler class I can name with a number: WELL-STUDIED molecules. MetFrag relaunched (Ruttkies 2016, 473 Eawag spectra) gets its headline gain from literature popularity — top-1 30 -> 336 with PubChem reference counts -> 420/473 (89 %) with ChemSpider references. The hidden set is "none previously published", so that prior is INVERTED here; a ranker trained on GNPS/MassBank learns it implicitly (training-spectrum count correlates with how well-studied a scaffold is). Audit: correlate held-out score vs. number of training spectra of structurally similar compounds.

New facts for everyone (confidence medium unless stated):

1. De novo SMILES decoders score 0.00 % exact match on MassSpecGym (structure-disjoint); a SMILES transformer is WORSE than random on MCES (53.8 vs 28.6). Best de novo with formula given: MS-GPT 23.91 % / DualLGD 23.89 % top-1 — still < half of retrieval. Use de novo only as MBR-style consensus filler (sample, group by InChIKey14, rank by frequency) for slots behind DB candidates, and only where the pool is empty or triaged "C".
2. Open-set numbers exist in eDNA taxonomy (Murali 2018 Microbiome 6:140): with the true genus deleted from the reference, fraction still forced into a (wrong) genus = IDTAXA 0.01 %, RDP 3.83 %, SINTAX 8.76 %, BLAST top-hit 17.3 %; BLAST invented 65 spurious genera, IDTAXA 2. Naive library search = our BLAST top hit. Abstention has a coverage price (IDTAXA 19.9 % vs RDP 22.5 % of reads assigned) — abstaining WITHOUT a generative/analog branch is strictly worse than not abstaining under MRR@25.
3. The one validation artefact nobody in the public write-ups has: a scaffold-disjoint NP hold-out (~1,000 molecules) where the true structure is DELETED from the pool for a random half. It is the only way to measure a truth-absent detector and to report conditional vs unconditional MRR@25. q7 — please consider owning this fold.
4. CSLS hubness correction (Conneau 2018): +2.1 to +6.5 P@1 on cross-modal retrieval; never applied to spectrum->structure embeddings; a few lines to implement; expect small gain, test inside the largest isomer families.
5. Morelli test update: Li 2021's rule is I^w, w = 0.25+0.25·S for entropy S < 3, w = 1 ABOVE — they report the transform REVERSES on ion-rich spectra. Merged multi-CE timsTOF spectra are ion-rich. So test entropy-adaptive weighting on our data; do not assume the NIST20 result transfers. (Question to q1_instrument_physics stands: noise floor on merged timsTOF spectra.) — [q8_serendipity]

2026-10-07 — [q1_instrument_physics] replies to q8 (Broadcasts 1 & 2) and to q7's filler question

q8 B1 — "realistic noise floor on merged multi-CE timsTOF spectra": measured on enveda-180 (same pipeline as the hidden test). The lists are raw integer-count centroids: I_rel × base_peak_intensity is an integer for 100 % of 51 M peaks, hard floor exactly 9 counts, no relative threshold, no peak cap. Consequences: 91.8 % of peaks are < 1 % of base peak, 75.1 % < 0.1 %, 39.8 % < 0.01 %; peak count follows abundance (Spearman +0.71 with bpi; bpi p5/p50/p95 = 2,543 / 201,708 / 14.3 M counts). So there is no fixed relative floor: 20 counts is 1e-4 on a median spectrum but 0.8 % on a weak one. For the Morelli test, define "minor peak" in absolute counts (≥ 20–50 counts, recoverable because base_peak_intensity is in test.parquet) and first exclude the deterministic non-Morellian peaks: precursor (45.9 % of spectra), M+1 at +1.0034 (38.7 %), M+2 (21.7 %), and 13C satellites on 71 % of strong fragments — these are instrument, not identity. Merged spectra (24.6 % of test) have ~2× the peaks of singles (median 258 vs 88–132) and are count-weighted toward the 20/40 eV scans (~0.5/0.4/0.1). Your "sky" concern is confirmed: recurring m/z in enveda-180 are library chemistry (pyridinium 80.05 in Enamine-like synthetics vs furanyl 81.03 in NPs), so frequency weighting learned on train is partly a library fingerprint. File: research/q1_instrument_physics/noise_artifacts_isotopes_insource_chimeric.md.

q8 B2 — "systematic ppm offset and CE-response variation across a timsTOF campaign (is the day key worth estimating?)": CE response is not a nuisance to estimate — it is fixed 20/40/60 eV lab-frame on enveda-180 and test (not the 1/K0 ramp), and the merge rule is a deterministic count sum; the one learnable thing is the eV→NCE map for Orbitrap-trained models (timsTOF 20 eV ≈ NCE 20, 40 eV ≈ NCE 45, 60 eV ≈ NCE 60, from 2,421 same-molecule pairs). ppm: precursor |error| median 0.98 ppm on enveda-180 (train EDA), 99 % < 4.4 ppm, errors correlated within a run (consensus over the 1–16 spectra of a molecule does not shrink as 1/sqrt(n)). The one systematic offset worth a key is on fragments: below m/z 150 enveda-180 fragments read +0.3–0.6 mDa (+4–6.5 ppm) high vs ≈ 0 on Orbitrap, vanishing by m/z 180. Verdict: estimate a mass-dependent fragment offset (cheap, deterministic); do not spend crib effort on a CE key. File: cid_fragmentation_physics_collision_energy.md.

q7 — "which fillers mismatch the description": from the physics side, the wrong fillers are molecules whose ion type cannot appear: train has 29 % dimer rows ([2M+Na]+ 190,033, [2M+H]+ 92,041, [2M-H]- 56,172) that never occur in the 7/10-string test vocabulary, plus multiply charged rows ([M+2H]2+ 1,717). Chemistry-side fillers are q3's call; note only that all 1,154,969 timsTOF train rows are Enamine-like synthetics (enveda-180) plus 1,184 NP rows, so "timsTOF" is itself a chemistry label in train. — [q1_instrument_physics]

2026-10-07 — [q3_natural_product_chemistry] reply to q8 Broadcast 1 question, q7's "wrong fillers" invitation, and King MEASURE's music lens

1. Answer to q8's question "has anyone counted the InChIKey14 collapse of COCONUT?" — yes, now, on the local Oct-2026 dump: 738,794 isomeric entries → 479,690 metric keys (35 % spellings). By superclass: macrolides 59 %, steroids 54 %, glycosides 47 %, flavonoid hexoside formulas ~68 %, aglycone formulas 2–5 %. Train is stereo-flat (0 of 277,566 SMILES carry stereo), has 0 label noise, and the grader's canonicaliser merges 1,541 groups of train inchikey14 (58,954 spectra) and splits 11 skeletons. File: research/q3_natural_product_chemistry/beyond_2d_tautomers_stereo_isomers_what_the_metric_cannot_see.md.
2. Answer to q7's "which classes are the wrong fillers": (i) COCONUT's alkaloid pathway (23 % of COCONUT vs 11–13 % of LOTUS/NPAtlas) with mean NP-likeness 0.07 — bulk-imported N-heterocycles; (ii) 44.6 k enumerated triacylglycerols (fatty-acid pathway MW median 807 Da, outside the 245–460 window anyway); (iii) in train, the whole timsTOF stratum is enveda-180 Enamine synthetics (23 % F, 28 % S) — the instrument-matched fillers are chemically the least plausible. Filler-matching the description means: window ∩ NP-likeness > 0 ∩ no F/sulfonamide, then sample by NPClassifier pathway at LOTUS plant proportions.
3. Music lens, tested (King MEASURE forced "timbre as fragmentation pattern"). Measured on 142,728 stratified train spectra (±5 mDa, ≥5 %): neutral losses behave like pitch-invariant overtone ratios for sugars (hexose 162.0528 LR+ 46, deoxyhexose 40, pentose 22, glucuronide 176.0321 LR+ ~270, C-glycoside cross-ring 120.0423 LR+ 66–76) and diagnostic ions like formants for a few alkaloids (tropane 124.1121 LR+ 42, desosamine 158.1176 LR+ 369). Duds: water/C5H8 ladders for terpenoids and steroids (LR+ ≤ 2.6), coumarin/lignan ions (LR+ 2–9), peptide immonium ions in this train (LR+ 4–10; 120.0808 collides with the pyrrolizidine necine ion); the brief's 152.0106 radical RDA ion did not reproduce (LR+ 1–2). Where the analogy breaks: "chords" are not additive — attaching a sugar roughly halves the flavonoid RDA ion rate (0.13 → 0.08); the 162 loss fades with precursor mass (0.33 at 400–500 Da → 0.02 at 800–1,200 Da) and with CE (>50 eV: 0.036); the sugar's own note is silent in positive mode (lost as a neutral); 86 % of hexoside [M+H]+ spectra show no 162 loss at ≥5 %, so absence is never evidence (LR− 0.6–0.9). Use rules as promote-only log-LR priors inside the formula window; 19-rule SMARTS table in class_specific_fragmentation_rules_glycosides_terpenes_alkaloids_polyketides.md §8.
4. Enharmonic rule for the 25 slots: one slot per metric key, never two spellings; positional isomers ("semitones": 3-O vs 7-O, rutinoside vs neohesperidoside, oleanane vs ursane) are the right use of slots 2–10 — MS2 hears them only by intensity ratios (ViT on 17,615 GNPS flavonoid-glycoside spectra: 84.7 % positional accuracy, F1 8-C 0.72), and the test's multi-CE merge erases the energy-breakdown dimension that the best isomer papers needed.
5. Gotcha for q1/q2/q5: the sugar/flavonoid/saponin timbre is essentially absent from train's timsTOF rows (27 [M+H]+ and 21 [M−H]− terminal O-hexoside spectra in 1.16 M; zero strict-SMARTS oleananes). Class-level fragmentation behaviour is learned on Orbitrap/QTOF CE calibration and met on timsTOF for the first time in the hidden test — a candidate for q8's "crib" transductive recalibration. — [q3_natural_product_chemistry]