⌁ CASMI EXPLORER← Back to our quest

A DATED, AUDITABLE RESEARCH STORY

Evidence, with its limits.

Project snapshot: September 21, 2026. Official position was checked September 22 at 01:15:34 UTC (September 21 at 19:15 MDT): best 0.328, rank 387 / 1,124, leader 0.409. These are recorded findings, not a live leaderboard feed. The latest local research lead is GLACIER; it has not been promoted.

The private workbench contains the underlying reports and experiment artifacts. This public site reproduces a curated summary with report names and SHA-256 provenance markers. A hash identifies the report version but does not independently verify its contents. Competition data, raw spectra, model weights, and private source files are not distributed here.

Machine-readable experiment summary · Source snapshot manifest · Official leaderboard (may require sign-in)

Reproduce the public baseline

0.240 → 0.328 · Adopted control — The current scored control · September 21

Official submission 56443775 scored 0.328. The cached control preserved all 17 frozen scoring cells. Its recorded cloud runtime fell from 3,646 to 2,259 seconds; all 400 placeholder ranked lists and CSV bytes matched the slower control.

This is an official improvement for the full recipe, not a proven effect of one channel. The predictive design and weights are a public-baseline reproduction. Caching is an execution improvement, not an accuracy gain.

Source record: reports/reference_control_20260921.md
SHA-256: c82c6dbca8f059f5a398d7ec7e5919e56887d1136a8728c4f50f4bc625dcf53d

Expand the direct-match library

0.240 → 0.238 · Not promoted — A local prediction did not transfer · September 21

The full-library transplant cleared its local gate with predicted leaderboard delta +0.012120 [0.009206, 0.015024]. Official submission 56442933 instead scored 0.238 against the earlier 0.240 comparator.

The official test did not confirm the expected gain. One noisy submission does not establish a true 0.002 regression, but this change has no demonstrated benefit.

Source record: reports/reference_control_20260921.md
SHA-256: c82c6dbca8f059f5a398d7ec7e5919e56887d1136a8728c4f50f4bc625dcf53d

Combine four neural models

+0.020272 · Uncertain gain — Historical held-out C2 · same candidates

C2 MRR moved from 0.562766 to 0.583038. The paired 95% interval for the change was [−0.004989, +0.045697], crossing zero. Candidate lists and all 43,230 classical-feature rows matched.

The primary confidence-interval gate failed. Aligned ranker training and independent, identity-audited evaluation are needed before promotion. This is not a leaderboard score.

Source record: reports/reference_control_20260921.md
SHA-256: c82c6dbca8f059f5a398d7ec7e5919e56887d1136a8728c4f50f4bc625dcf53d

Try shared-space & Bayesian retrieval

Δ selected = 0 · Not promoted — 137 development / 105 historical test molecules

Historical C2 MRR: re-fitted baseline 0.541994; vector-only 0.187738; Bayesian-only 0.315916. Development selection chose zero new-channel weight, so the selected fusion equaled the baseline.

The infrastructure works, but this small MLP and approximate fragment model did not earn an added role in the scored control. These results do not reject all learned or Bayesian retrieval.

Source record: reports/retrieval_lab_20260921.md
SHA-256: b5ea7de2d9c7246569596f17335918ad88796fac4375057a98584f93232ca0ab

Learn graphs & collision-energy effects

+0.000775 · Gates failed — Earlier graph / energy branch · historical cohort

Selected historical C2 fusion moved 0.541994 → 0.542769, with delta interval [−0.013320, +0.013368]. Graph-only MRR was 0.134778 versus matched MLP 0.215990. Energy-conditioned MRR was 0.352006 versus blind 0.347522; its delta interval also crossed zero.

Graph, energy, and selected-fusion gates all failed. C1 remained unchanged. The additional 128-molecule historical cohort was not an independent external benchmark. No new official submission was made.

Source record: reports/graph_energy_20260921.md
SHA-256: 55be2e8d862fcf8e301b5fdfa3e095158d75ac1674b6d69c9a3f48311d380fda

Use pretrained GLACIER fragments

+0.151977 · Promising pilot — Latest lead · 63 exposure-screened test molecules

GLACIER with marginalized spectral likelihood improved MRR from 0.705714 to 0.857691 against the matched energy head; paired 95% delta interval [+0.052649, +0.247336]. Development selected GLACIER alone. First-place answers increased from 38/63 to 49/63.

The prespecified local comparison passed. This historically explored, narrow cohort contains only one COCONUT test molecule and is not representative C2 validation. Complete checkpoint training ancestry is unavailable. Independent validation is required; no production promotion, leaderboard gain, or 0.4+ score is established.

Source record: reports/glacier_20260921.md
SHA-256: b0936deb381a4137eee6887bc764130fe9a43a1793798608b23efb339f0b547b

What the browser actually computes

The dependency-free Rust kernel is compiled to WebAssembly. It receives synthetic spectra from JavaScript, converts the protonated precursor to neutral mass, filters candidates by a parts-per-million mass window, greedily matches sorted peaks, and computes square-root-intensity cosine similarity. Each matched peak is consumed once. Unmatched peaks still contribute to both normalization terms. Scores are bounded between zero and one but are not probabilities.

The global shift slider illustrates one idea used in analog matching. The full research implementation uses binned masses, structural fingerprint similarity and more complete channel logic; it is not reproduced end-to-end here. Example candidate A–E structures, masses and spectra are illustrative and do not represent real chemical structures or released competition examples.

The rank widget computes 1/rank for the first correct identity at ranks 1–25, and zero outside that range. The official MRR is the average across scored molecules after chemical canonicalization. The browser does not perform RDKit canonicalization, run FPNet, or run any trained ranker.

What counts as progress?

An official score is measured by the competition scorer. A local comparison is meaningful only against its matched control on the same cohort and protocol. A positive estimate with a confidence interval crossing zero fails our evidence gate. A positive interval on a narrow pilot is a lead to validate, not proof that it transfers to the competition.

The 0.328 control reproduces a public four-channel baseline. Its caching improvement preserves output and reduces execution time. The result does not establish that any single channel caused the accuracy gain or reveal the first-place team's method.

Attribution and related sources