01 Scout
Use frontier models to widen the search.
Frontier AI helps review papers, propose model families, spot missing controls, and draft job specs. It is a scout, not the judge.
loading published snapshot…
Monterey Bay AI Lab · Mission Control
Bringing together people, data, AI, and compute to find the hidden signals that make a real difference in people's lives.
We gather scattered public data about the coast — weather, rivers, satellites, ocean sensors, lab tests — into one checked place, run many AI models against it, and hunt for the rare signal that genuinely beats "tomorrow will look like today." Every claim has to survive a fair comparison, data it has never seen, and a tough critic — so the wins are real, and the dead ends stay on the record.
Latest wins
The newest results that passed our evidence checks — click any card for the exact numbers and how they were tested.
Built & live
Not just findings — tools the lab has built and put online. Open one and try it yourself.
Sensor-fusion · sea-surface temperatureLive
The Evidence Agent
Cross-checks independent measurements of the same thing to test whether an answer can be defended. For a given day and place it lines up an ocean buoy's thermometer against separate satellite sea-surface-temperature analyses, and shows exactly how they agree or disagree — with an honest uncertainty band and full provenance for every number. It deliberately refuses to assert one "true" value: it presents the evidence, not a single blended figure.
Start to finish · narrated · closed captions
The whole lab in five minutes.
From hundreds of public ocean datasets to a handful of warnings worth trusting — the shared data store, the AI models that hunt for signal, the checks that throw out false alarms, and the wins that survived, including a beach advisory that works even on brand-new beaches. The latest results are in the cards above.
Why now
Four forces are converging at once — agentic AI, deep open data, cheap compute, and experts from many fields — putting discovery that used to need a whole institute within reach of a small team.
How we work
Four moves, repeated on every question — with the truth checked at each step.
01
Build one clean data store
Pull scattered public data into one clean, checked store — raw, cleaned, and ready-to-use layers, with units reconciled, records matched up correctly, and no future information allowed to leak into the past.
02
Try many AI models
Run many different kinds of models against the data and make each one earn its place. The wins tend to come from simple, physically-grounded features; fancier methods are kept only where they clearly do better — and we keep an honest list of the ones that didn't.
03
Find the hidden signals
Hunt for the outside signal that beats simply assuming "tomorrow looks like today" — the thing that turns raw data into a forecast people can act on.
04
Keep the truth the whole way
Every claim is tested against a fair yardstick, on data it hasn't seen, with a shuffle test and a devil's-advocate check. Dead ends stay on the record.
Every move is run by a small crew working together:
AI research agentsFleets of AI agents read the data, run the experiments, and challenge each other's claims to catch mistakes.
On-prem GPUAn in-house GPU trains the models and runs every gate locally — the compute stays under our roof.
Domain expertsOceanographers and public-health advisors steer the questions and sanity-check every result.
Deep dives
The longer write-ups and briefings behind the wins, brought into one place — open any to read the checked numbers underneath.
2-day beach-bacteria advisory on free forecast data — the Carrier-Timescale Law (newest win)
Topic: beach safetyTarget: enterococcus / E. coli exceedanceData: free NOAA GFS forecastUse: 2-day advisory
Where a beach's pollution is driven by rain (a fast carrier) and the beach is sampled densely, a free public NOAA GFS rain-forecast feed can flag an exceedance two days ahead. The lab showed a real but modest edge where this works — a few points of recall over a fair baseline, CI above zero — mapped exactly when it does not, then deployed it to brand-new beach networks with zero local training labels. The value is the discipline and portability, not a big accuracy jump.
0.730California 2-day recall @ FPR ≤ 0.15 (absolute rate; ~8 pts over persistence 0.651) — beats persistence and a rain-model, CI excludes 0
3 regions · 2 indicatorsConfirmed CA + NJ (marine enterococcus) and Ohio Lake Erie (freshwater E. coli)
2 networks cold-startZero local labels → NSW Australia (0.536 vs 0.491) and TX Gulf (0.580 vs 0.541, CI excl 0)
Advisory vs. local baseline — 2-day recall at matched false-alarm rate
The law: when a forecast feed actually pays
A forecast buys lead-time only when all three hold: a fast carrier (observations go stale between samples), headroom of a carrier forecast over the antecedent-observed carrier, and sufficient regional NWP skill. Two of three ⇒ null. All three ⇒ a deployable advisory. A free pre-screen (observed-carrier oracle headroom, computed with no forecast fetch) tells you in advance where a feed will pay — it correctly rejected the slow-carrier and chronic cases before spending on any forecast data.
Honest scope: the forecast-pays win is demonstrated for densely-sampled marine enterococcus beaches (CA + NJ) plus one freshwater case; it is a 2-day, not 3-day, product and not a beach-posting forecast. Effect sizes over recurrence are modest — the strength is discipline (pre-registered kill-metrics, phantom-guards, red-teams), generality, and cross-network portability, not a large accuracy jump. Slow-carrier and chronic cases are low-powered nulls, recorded honestly rather than dressed up.
See the finding card →
The zero-shot coastal-bacteria ranker — data-scientist video briefing
Topic: beach safetyTarget: enterococcusMethod: zero-shot rankerFormat: narrated briefing
How a simple physical-intuition-plus-logistic ranker cleared a hardened zero-shot screen: the label-free panel contract, the physics feature layer, leave-one-basin-out validation, permutation nulls and bootstrap gates, the washouts, and the one narrow claim that survived.
12 domainsHeld-out coastal basins, each hidden in turn, scored with no target-domain labels
708,337 rowsHeld-out station-days spanning 57,628 daily exceedance events
1.75× AP liftMean average precision 0.1511 over a 0.0909 base rate; all 12 domains beat their base rate
Mean ranking skill vs. base rate (12 held-out basins)
Narrow claim: a retrospective ranking screen for early sampling attention — not a calibrated advisory probability.
Open related findings →
California beach-bacteria shadow triage — plain-English write-up
Topic: beach safetyTarget: enterococcusPlace: CaliforniaUse: shadow triage
A frozen model ranks which California marine beach station-days are most likely to exceed the enterococcus safety threshold before the newest lab culture comes back — hours before results that normally lag the same-day decision. It is promoted as a shadow triage helper, not an official beach advisory.
0.868ROC-AUC on 2022+ holdout
0.510Average precision (base rate ≈ 10.1%)
0.009Expected calibration error
Ranking skill vs. real baselines (AP)
The warning test (score ≥ 0.0635)
| What happened? | Model said high risk | Model said low risk |
|---|---|---|
| Enterococcus exceedance happened | 7,780 | 1,219 |
| No exceedance happened | 26,026 | 54,296 |
The warning line caught 86.5% of future exceedances on the 2022+ test — with many false flags, which is why the right use is triage and review, not automatic closure.
Harder audits
9/9Counties where it beat the AB411 rain rule
9/9Counties where it beat station memory
8/9Counties passing deploy-ready calibration
A first prospective lockbox (frozen June 18, 2026; scored HOLDS June 27) held the claim on 198 rows / 8 events — still small, but forward evidence.
Important limit: this should not decide by itself whether a beach is open or closed. Official sampling, lab testing, public-health judgment, and legal advisory rules still control those decisions. The frozen headline uses a two-day reveal-lag proxy; the conservative available-at AP is ≈ 0.490, about 0.020 lower.
Open related findings →
Ireland domoic-acid monitoring helper — plain-English write-up
Topic: ocean scienceTarget: domoic acid (ASP)Place: IrelandUse: monitoring triage
A calibrated model learns from past Ireland shellfish-toxin records to rank which sites and weeks most deserve monitoring attention for amnesic shellfish poisoning (the illness linked to domoic acid). It beat a fair seasonal baseline on a future holdout and is promoted as a monitoring triage helper — not a replacement for lab tests or safety rules.
0.390Model average precision (future test)
0.143Best seasonal baseline
93/124Future toxin events caught
Ranking skill vs. baselines (AP)
The warning test (threshold 0.0966)
| Metric | Model | Seasonal baseline |
|---|---|---|
| True positives | 93 | 107 |
| False positives | 216 | 951 |
| False negatives | 31 | 17 |
| Precision | 0.301 | 0.101 |
| Recall | 0.750 | 0.863 |
The baseline caught a few more events but raised far more false alarms; the model is stricter and much more precise — useful when monitoring attention is limited. A shuffled-label red-team could not reproduce the result, so it is hard to dismiss as luck.
Against the closest public system (bulletin region-weeks, AP)
Data: Ireland Marine Institute shellfish biotoxin monitoring (220,803 raw records → 35,042 usable ASP rows across 139 sites, Nov 2002 – Jun 2026), with a public comparison parsed from Ireland HAB Bulletin PDFs.
Important limit: because it still misses some events, it should help decide where to look first — not make harvesting open/close decisions alone. Lab testing and official safety rules remain in control.
Open related findings →
Tokyo Bay hypoxia (dead-zone) 7-day lead forecast — plain-English write-up
Topic: ocean healthTarget: hypoxia (low oxygen)Place: Tokyo BayUse: early warning
Low-oxygen "dead zones" suffocate seafloor life and disrupt fisheries. On Tokyo Bay's 7-station, 16-year monitoring record, a leakage-guarded model forecasts whether bottom water will go hypoxic (dissolved oxygen ≤ 2 mg/L) seven days ahead — and the physics of thermal stratification adds a small but real, statistically clean edge over an already-strong persistence-plus-season baseline.
0.847-day forecast average precision (ROC-AUC 0.93; base rate ≈ 30%)
+0.021AP added by stratification physics over persistence+season; 95% CI [0.011, 0.053], excludes zero
7 daysLead time; the lift survives a permutation null and drop-one test at 7 and 14 days
Flagging hypoxia before it starts — onset skill (AP)
Among currently-oxygenated origins — the useful "clean-to-hypoxic" onset question — the model reaches AP 0.835 versus 0.747 for the best seasonal/oxygen-proximity naive at 7 days (4,356 held-out origins, 934 onset events). Checked by training on years ≤ 2021 and testing on later years, with station-block bootstrap confidence intervals and a permutation null; bottom oxygen, which defines the target, is excluded from the features.
Honest limits: the raw next-week forecast (0.837 vs 0.816) beats its seasonal baseline only narrowly, because Tokyo Bay hypoxia is strongly seasonal — the real, CI-clean win is the stratification-physics increment and the onset reframe. This is a validated research result, not an MBAL deployment. It replaces an earlier Gulf-of-Mexico dead-zone card whose apparent skill did not survive an effort-confound control (that lane is a documented null).
Findings
Strongest results first. Each card says — in plain words — what we asked, what we found, and how sure we are. Click for the technical read and, where the evidence has it, a confusion matrix.
The questions
The 14 research questions this lab is working on, grouped by topic. Each card is a real question with an honest status — open one to see its evidence, models, signals, map, and data gathered in a single place. Wins, caveats, and nulls sit side by side.
Discovery Engines
This is the search machinery behind the results. We use frontier models to find candidate engines, translate them into runnable jobs, execute the real experiments on the RTX 5090 workstation, and only promote outputs that beat a fair baseline on held-out data.
02 Build
Turn ideas into reproducible jobs.
Each candidate becomes a script with a target, split, baseline, metric, artifact path, and confusion-matrix status when labels allow it.
03 Run
Execute on local compute.
GPU-heavy jobs run on the RTX 5090 box; CPU screens run as cheaper sweeps first. The output is saved before anyone writes a claim.
04 Gate
Promote, scope, or kill.
A result has to beat the right plain baseline, survive leakage checks, and pass critic review. Otherwise it stays a null or a data ask.
Better page organization
Use this tab in three passes during the meeting: first the operating loop above, then the engine-family cards below, then the raw registry table only when someone wants the technical model record. That keeps the story accessible while preserving enough detail for data scientists.
The honesty engine
The honesty engine is the discipline around every run: red-team the target before compute, then check leakage, bad data, bad units, duplicate rows, date semantics, source gates, and baseline choice before a claim can move. If the data is not ready, we do not patch it into a story; we mark it blocked, scope it, or send it back to the lakehouse with the failure visible. Only results with saved artifacts, held-out evidence, a fair baseline, and an adversary-readable outcome get promoted.
Engine portfolio
The families we actually use: baselines, tree models, rankers, foundation-model channels, zero-shot transfer, physics-pi screens, sequence models, event studies, and auditors.
Fitted model registry
Click any model to open the exact card: task, family, evidence status, and raw snapshot fields. This is the audit layer after the plain-English engine map.
ModelTask
Lakehouse
Everything shown here is rendered from one governed snapshot, so the findings, models, and map can't drift apart. Raw export downloads are no longer published from this site.
Source inventory
Every source behind the snapshot, ranked by size, with readiness, date coverage, and representative signals. Everything here is also searchable in the Ledger.
How the AI builds this lakehouse
Autonomous agents run the whole pipeline end to end — pulling raw data from public sources, cleaning it into leak-free model features, and promoting only what survives a hard evidence gate. Nothing reaches "governed" without beating an honest baseline.
STEP 1 · FETCH → BRONZE
Agents pull the raw data
Data-adapter scripts call public APIs, ERDDAP/OPeNDAP servers, satellite archives and agency portals — NOAA, USGS, Copernicus, CalHABMAP, state water boards, MBARI — and land the records as raw bronze parquet: unchanged, fingerprinted, timestamped.
STEP 2 · CLEAN → SILVER
Transforms make it model-ready
Silver transforms normalize schemas, harmonize units and regulatory thresholds, de-duplicate, quality-check, and build strictly backward-looking (leak-free) lags and rolling features — the clean tables the models actually train on.
STEP 3 · PROMOTE → GOLD
A gate decides what's real
Each result is scored against the strongest fair baseline (rolling site-memory), forward-in-time, with permutation nulls and an adversarial red-team. Only results that genuinely beat the baseline are promoted to gold / governed; everything else stays experimental.
What the layers mean (the standard lakehouse "medallion" tiers): Bronze = raw data exactly as it was fetched, nothing changed. Silver = cleaned, unit-harmonized, de-duplicated, leak-free features. Gold = results that passed the evidence gate and were promoted to governed. The Layer badge on each dataset below tells you which tier it lives in.
Verify & reproduce
- The public site keeps the audited findings, models, map coverage, and ledger visible without serving raw export bundles.
- Every claim still traces to an evidence path or report ID so collaborators can request the exact governed artifact through the lab.
- Published summaries should be treated as decision support, not a replacement for controlled data access or source-system records.
Complete ledger
Every result and its research trail, in one searchable list: findings, models, innovation reviews, research questions, and the signals we tested. Raw data assets live in the Lakehouse. Use the filters to narrow the list; click any row to expand a plain-English explanation, charts, saved fields, and the raw record.
Glossary
This lab is careful about words. An untrained XGBoost or neural net is a learner, not a model — it only becomes a trained model once it's fitted, evaluated, and registered. A baseline is an honest comparator, not a discovery; a signal carries verdicts (KEEP / WASH / REJECT) per target. Each term below links to its real examples from the registry and signal catalog.