Enveda as plaintiff: PRISM facts, the people behind CASMI 2026, stated vs inferred motives, and the provenance of the hidden 400

TL;DR. Enveda is a $2 billion, Series-E ($311 M, 2026-09-23) Boulder biotech whose entire drug-discovery platform rests on one claim: that it can read molecular structure out of Bruker timsTOF spectra better than anyone else, using PRISM, a BERT-style model pre-trained on 1.2 billion spectra of which half are its own and none of which are released. It revived CASMI eight days before announcing that round, and its press release says in plain words that the contest "helps to grow PRISM" and that building PRISM taught it the field lacks "a benchmark that can tell a real advance from a model that has quietly seen the answers." The hidden 400 come from Enveda's own isolation and profiling platform; the only public description of how that platform validates a structure is SIRIUS fragmentation-tree scoring plus NMR, so the compounds the field is being judged on are, by construction, the ones Enveda's own stack could already explain.

1. PRISM facts table

Every number comes from the only primary source: Enveda's blog post of 2024-05-28 (authors Healey, Domingo-Fernández, Taylor, Krettler, Lightheart, Park, Kind, Allen, Colluru) and the press release of 2024-05-29. There is no paper, preprint, model card, code or weights: OpenAlex searches for the name and the acronym expansion return zero Enveda hits (2026-10-07); q5 found the same on arXiv and Europe PMC. PRISM is closed-weight.

Enveda's other spectral-ML work (headline numbers; q3 has the full bibliography)

Negative result: Enveda's engineering blog enveda.dev fails TLS from this machine; every blog fact above comes from the enveda.com mirror.

2. People table

The Kaggle citation block (as indexed by the search engine; the Kaggle page itself is a reCAPTCHA shell from this machine, so medium confidence) reads: "David Healey, Christoph Krettler, Tobias Kind, Marie Killian, Seth Drake, Erik DeBloois, Ivy Lightheart, Yojana Gadiya, Pelle Simpson, Daniel Domingo-Fernandez, Viswa Colluru, August Allen, Walter Reade, Ashley Oldacre. Enveda CASMI 2026 - Molecule ID From Mass Spectra. 2026." The Kaggle API confirms organizationName: Enveda, hostName: Enveda.

Observation, not speculation: the organiser list is the Enveda-180 author list plus Lightheart, Drake and two Kaggle staff. The people who made the training library made the test.

3. Motives table

Funding and pipeline context (enveda.com/press; FierceBiotech): Series B1 $119 M total (2023-04-20); $55 M (2024-06-14); Series C $130 M (2024-11-21; "over 1 million natural compounds characterized"); Sanofi extends it to $150 M (2025-02-26); Series D $150 M (2025-09-04, unicorn); Series E $311 M (2026-09-23, Catalio lead; ">$845 million since inception"; "17 development candidates"). Clinical: ENV-294 (atopic dermatitis/asthma, Phase 2 since 2025-12-09; Phase 1b 68 % EASI reduction Day 28, 85 % Day 42); ENV-308 (exercise-mimetic, Phase 1 positive 2026-08-18); ENV-6946 (IBD, Phase 1 since 2025-12-23).

4. Hidden-test provenance, every unknown marked

5. enveda-180 and enveda-np-examples (what is public; q2 has the full anatomy)

• enveda-180: Enamine HLL-200 hit-locator library, pooled, 14 Enamine batches ("enamine-part1" ...), timsTOF Pro 2, three nominal CEs, both polarities, 184,330 compounds / 1,158,293 curated spectra, CC BY 4.0 data, MIT code, Zenodo v1.0.0 2026-05-29, MassIVE MSV000102122, ChemRxiv 2026-06-04. Why "180" is not stated anywhere (q2 reads it as ~180 k compounds; low confidence). Compound vendor: Enamine (keyword on Zenodo; README says "Enamine REAL space"; q2 identifies HLL-200). Plate layout, pool size and acquisition dates: unknown.
• enveda-np-examples: 1,184 timsTOF spectra of 250–251 natural-product structures; no data card, no press mention, no paper. C&EN's "its own collection of purified natural products" is the only description. Whether purchased standards or isolates: unknown; q3 infers purchased standards because every structure already has public spectra.

6. Enveda statements on expected or PRISM scores

None. Searches for PRISM baseline, expected MRR, "will not compete", host statements on internal scores returned nothing; the press release, C&EN, Series E coverage, the Timmerman "Long Run" podcast page (2026-03-31; no transcript) and the Generalist interview page contain no number. The only Enveda number on a dark set is MS2Mol's 21 % close-match on EnvedaDark (2023), which is a Tanimoto-threshold metric, not InChIKey14 MRR, and not on this test. Negative result, high confidence that nothing public exists; cannot exclude Kaggle forum posts (unreadable here).

Forced analogy: law of evidence, standing and motive of the plaintiff

Statement. Enveda is the plaintiff who drafted the rules of procedure (MRR@25 on canonical-tautomer InChIKey14, MIT licence, 9 h offline notebook), selected the exhibits (the hidden 400), and serves as clerk (the scorer runs on its answer key). Every entrant is a defendant asked to prove identity beyond a 25-candidate doubt.

Standing. The plaintiff owns the compounds (isolated, bought or made on its platform) and validated the ground truth itself. Nobody outside Enveda has certified a single hidden structure; the founders of CASMI "support" but are not listed as validators. In evidence terms the chain of custody is intact but single-party: the same hands acquired, annotated, curated and will score.

Interest in the outcome. A high winning score serves Enveda more than a low one: it validates timsTOF-era data as sufficient for structure elucidation (the premise of a $2 B valuation and the Series E use-of-funds "scale the PRISM platform"), yields MIT-licensed methods, and lets Enveda compare PRISM privately against the best public method on the same key. A low score serves only the narrative "the field needs Enveda-scale private data." Neither outcome costs Enveda anything, which is why the release can promise both "rigorous test" and "helps to grow PRISM" without contradiction.

Conflict of interest. The ground truth was produced by a stack (SIRIUS trees, CSI:FingerID/MS2Mol hybrids, NMR on isolable amounts) that is itself a structure-elucidation method. A compound enters the hidden set only if that stack, plus chemistry, could assign it. So the exhibits are filtered by the plaintiff's own instrument of proof.

Testable consequence 1 (selection by explainability). If hidden compounds had to pass SIRIUS-style curation (Enveda-180's thresholds: rank-1 tree, >50 % explained intensity for 98 % of spectra, >70 % for 91 %), then the hidden set is enriched for spectra whose peaks SIRIUS can decompose into subformulas of the true formula. Test on our side: compute SIRIUS/MS-FINDER explained-intensity for the visible decoy (enveda-180 rows, same pipeline) versus GNPS timsTOF NP spectra of comparable mass; if the public-LB write-ups' formula-assignment accuracy is far higher than MassSpecGym's formula-free numbers, selection is at work. Modelling response: weight the formula step heavily, prune candidates by fragment explainability (a MetFrag-lite/ICEBERG score), and expect fewer "unexplainable" spectra than the wild.

Testable consequence 2 (what isolates look like). If a large share are platform isolates, they are compounds that survived 2D-LC fractionation at microgram scale and NMR, i.e. abundant, stable, mid-polarity plant metabolites: glycosylated flavonoids, labdane and withanolide terpenoids, alkaloids, mass 300–800 Da. Prediction: COCONUT/LOTUS coverage higher than for med-chem analogs, and class priors near the Domingo-Fernández 2023 plant-space census (flavonols, steroids, monoterpenoids, anthraquinones, alkaloids). The decoy's median mass 348 and range 157–1,159 Da fit. Check against the public write-ups' class-2 recall by NPClassifier class once we have a hold-out.

Discovery withheld. The plaintiff recorded RT, CCS and MS1 isotope patterns for every spectrum (Enveda-180 proves the platform does), and withheld all three. In a court this is suppressed exculpatory evidence; here it is a design choice that removes exactly the modalities Enveda's own ALIGN and CCS-GNN papers study. Modelling response: do not spend effort on CCS or RT models for this contest; spend it on candidate coverage and formula.

Where the analogy holds, partially holds, breaks. Holds on standing, interest and single-party custody. Partially holds on conflict: Enveda's organisers are also the authors of the leakage audit, so the clerk is unusually motivated to keep the trial honest, and the live tautomer canonicalisation is evidence of that. Breaks in two places: there is no adversary who can cross-examine the answer key (a wrong InChIKey in the solution file would never be found; CASMI 2016 and 2022 both had post-hoc corrections per the history worker), and Kaggle admits every method without a Daubert-style gate, so "admissibility" has no analogue. The analogy also cannot say what Enveda will do with the MIT code; the licence is a fact, the harvesting is an inference.

Gotchas / what could fool us

1. Enveda-180 is an instrument twin and a chemistry decoy: same timsTOF Pro 2 and 20/40/60 eV, but Enamine amides and triazoles, ~0.03–0.2 % in COCONUT depending on keying (q2). A model that looks good on an enveda-180 hold-out has learned the instrument, not the chemistry.
2. The organisers wrote the leakage audit. SMILES-canonicalisation artefacts, PubChem ranking bias and fingerprint-decoder memorisation are the shortcuts this host set out to kill; assume the answer key defeats them.
3. "Purified natural products" in C&EN is the training slice, not the test. Read as a test description it predicts 99.6 % COCONUT coverage; the public LB implies ~55 % recall@25.
4. Official text contradicts itself on spectra count (1,500 vs 2,500) and adducts (10 vs "one of the seven"); the Data page governs format, the press release is marketing.
5. Healey's title moved from VP of Data Science (2024) to Founding Data Science Fellow (2026); whether the lead organiser still runs production models is unknown.
6. PRISM has no published retrieval number; its only public metrics are relative gains against unnamed baselines on unnamed data.
7. Single validator, unpublished error rate: 2 % label error on 400 molecules is ~0.02 MRR, the size of the metric's standard error (critiques worker).
8. MIT requirement cuts both ways: AGPL components (check SIRIUS) could make a winning notebook non-compliant.
9. LB surge after 2026-09-23 press coverage is a PR effect, not a difficulty signal.
10. enveda.dev is unreachable; re-cite anything attributed to it to the enveda.com mirror or mark it unverified.

Implications for the challenge

1. Validate on what the plaintiff could have produced: NP-like, structure-disjoint, timsTOF-formatted, mass 157–1,159 Da, SIRIUS-explainable. Drop every hold-out that fails the explainability filter when estimating hidden-set performance, and report both numbers.
2. Formula first. If selection-by-explainability holds, formula assignment accuracy on the hidden set will exceed the wild; invest in a SIRIUS/MIST-CF-class formula step and multi-CE merging before any end-to-end model.
3. Candidate pools: COCONUT + LOTUS + patent-derived PubChem subset. Expect isolates to be in COCONUT; expect Enveda-style med-chem analogs (halogen, heteroaryl, OH to F on withanolide/labdane/sesquiterpene scaffolds, q3) to be absent from COCONUT by policy and present, if at all, in PubChem via patents. Add a one-step analog enumerator on NP scaffolds for slots 6–25.
4. Ignore RT/CCS modelling for this contest; those modalities were withheld by design.
5. Do not benchmark against PRISM; nothing is reproducible. Use DreaMS (open, MIT) as the foundation-model baseline; it is also what Enveda itself used in TIMS-Bench.
6. Licence hygiene now: inventory every dependency's licence in the notebook; AGPL components disqualify an MIT release.
7. Track host posts (discussion 741404 and later) for any statement on public/private split or class proportions; these are the only places Enveda will say more than the press release.
8. Budget for label noise: treat LB deltas < 0.04 as noise (critiques worker); do not chase them.

Open questions

• What fraction of the 400 are isolates vs purchased standards vs synthesised analogs? Only a host post settles it.
• Did the hidden set pass the Enveda-180 curation thresholds (explained intensity, SIRIUS rank-1 tree)? A single host sentence would confirm the selection-by-explainability hypothesis.
• Will the answer key be released after 2026-12-14 ("persistent benchmark"), and under what licence?
• Are Enveda employees barred from the leaderboard, and has Enveda run PRISM on the hidden set internally?
• Is "Ivy Lightheart" the "Rose Lightheart" of the PRISM byline? (Affects nothing in the modelling; affects the people table.)
• Does a PRISM preprint exist under embargo? Nothing on arXiv, bioRxiv, ChemRxiv or OpenAlex as of 2026-10-07.

Sources

Opened directly (fetch, curl or API):

• Enveda, "PRISM: A foundation model for life's chemistry", 2024-05-28, https://enveda.com/prism-a-foundation-model-for-lifes-chemistry/
• Enveda press release, Microsoft collaboration, 2024-05-29, https://enveda.com/news/pr-enveda-biosciences-announces-microsoft-collaboration-and-reveals-foundation-model-to-read-and-translate-chemistry/
• Business Wire via Yahoo Finance, "Enveda Is Launching the Next AlphaFold Contest...", 2026-09-15, https://finance.yahoo.com/healthcare/articles/enveda-launching-next-alphafold-contest-120700676.html (original: https://www.businesswire.com/news/home/20260915334355/en/)
• C&EN, L. Oldach, "Metabolomics has a data problem. Could this competition fix it?", 2026-09-17, https://doi.org/10.47287/cen.612552.newsarticle
• Enveda press listing, https://enveda.com/press/ ; Series C 2024-11-21, https://enveda.com/news/enveda-closes-130m-series-c-funding-to-deliver-multiple-clinical-readouts-from-a-pipeline-of-10-development-candidates/ ; Series E 2026-09-23, https://enveda.com/news/enveda-raises-311-million/
• FierceBiotech, Series D 2025-09-04, https://www.fiercebiotech.com/biotech/enveda-harvests-another-150m-advance-nature-inspired-drugs ; Series E, https://www.fiercebiotech.com/biotech/enveda-reaps-311m-round-feed-its-ambition-become-household-name
• Bruker, "Customer Insight Enveda", https://www.bruker.com/en/landingpages/bdal/customer-insight-enveda.html
• Premji Invest blog, 2025-09-04, https://www.premjiinvest.com/blog/evolutions-shortcuts-engineered-envedas-edge-in-drug-discovery
• Zenodo API, Enveda-180 record 20436851, https://zenodo.org/api/records/20436851 ; GitHub README https://raw.githubusercontent.com/enveda/enveda-180/main/README.md (cached copy)
• Zenodo API, TIMS-Bench record 20816379, https://zenodo.org/api/records/20816379 ; GitHub README of enveda/benchmarking-untargeted-metabolomics-software (cached copy)
• arXiv 2606.19624, "MassSpecGym in the Wild", https://arxiv.org/abs/2606.19624 (HTML cached)
• OpenAlex API: MS2Mol versions (10.26434/chemrxiv-2023-vsmpx, -v3, -v4), Enveda-affiliated works 2025–2026, audit metadata
• Kaggle API v1 competitions/list?search=CASMI (organizationName, hostName, teamCount 2,703 on 2026-10-07)
• Pixel Scientia podcast page, D. Healey, 2023-05-01, https://pixelscientia.com/podcast/searching-for-new-therapeutics-in-nature-with-david-healey-from-enveda-biosciences/
• Timmerman Report, "The Long Run" with V. Colluru, 2026-03-31 (page only, no transcript), https://timmermanreport.com/2026/03/inspired-by-mother-nature-discovering-new-drugs-viswa-colluru-on-the-long-run/
• Job ad, Metabolomics Scientist, https://jobs.trueventures.com/companies/enveda/jobs/65371897-metabolomics-scientist
• Sibling files: research/q2_data_forensics/enveda180_and_np_examples_anatomy.md; research/q3_natural_product_chemistry/enveda_platform_isolation_profiling_what_analogs_means.md; research/q4_challenge_intelligence/casmi2026_rules_timeline_prizes_code_requirements_faq.md; research/q4_challenge_intelligence/critiques_fiehn_benchmark_validity_what_the_test_does_not_measure.md; research/q5_methods_lineage/foundation_models_dreams_prism_massspecgym_benchmarks.md

Search-index only (page not readable from this machine; medium confidence):

• Kaggle competition citation block, https://www.kaggle.com/competitions/enveda-CASMI26-molecule-id-mass-spectra
• GDG event page on Walter Reade's role at Kaggle
• qz.com / BioWorld on the ~$2 B Series E valuation
• Walter Reade / Ashley Oldacre roles beyond the citation: unverified

Not reachable (negative results): enveda.dev (TLS error); ChemRxiv full text for MS2Mol and Enveda-180 (403); bioRxiv TIMS-Bench (429); startup.jobs Staff/Principal Data Scientist ad (403); TechCrunch/Endpoints Series E (403); Kaggle discussion 741404 (reCAPTCHA).