Kenton Labs · Open data research

OpenFind

Find the signal. Test how far it travels. Keep the limits in view.

Three completed computational pilots · Evidence-led research

OpenFind builds reproducible analyses from public datasets. Each pilot starts with a bounded question, compares simple baselines with more detailed models, and tests whether an apparent result survives geographic or chronological separation.

01 / Hydrology

When low-flow patterns travel between watersheds

A public streamflow and watershed-attribute study examining low-flow buffering, geology, and prediction beyond familiar regions.

Observed data

671 matched watersheds and 8.4 million daily streamflow records.

Geographic holdout

With all predictors, held-out-region R² was 0.196, compared with 0.440 under random folds.

Finding status

A permeability–bedrock-depth interaction is an exploratory lead, not a causal or established discovery.

Three-panel analysis figure showing watershed observations, random-fold versus held-out-region prediction, and uncertainty intervals — Watershed attributes · Regional validation · Planned coefficient checks

What the result supports

Random validation can overstate transfer to new drainage regions in this dataset. The next useful step is independent hydrogeology information or a geographically separate cohort.

Sources: CAMELS v2.0 · USGS daily streamflow

02 / Space weather

History matters in geomagnetic disturbance

A chronological analysis joining a published interplanetary coronal mass ejection catalog with NASA’s hourly solar-wind and geomagnetic records.

Held-out events

162 catalog events from 2009–2020, evaluated against a model fitted to earlier years.

Prediction comparison

A fixed 12-hour memory model reduced held-out RMSE by 33.9% versus integrated driving dose.

Interpretation

Retrospective conditional prediction; it uses measurements through the event and is not an advance forecast.

Held-out comparison of integrated dose and 12-hour memory models, predicted against observed values, and annual transfer variation — Held-out chronology · 2009–2020 · Prediction and year-by-year transfer

What the result supports

Temporal driving history improves this model comparison. The analysis does not establish a new physical mechanism; a residual timing lead did not transfer to later data.

Sources: Richardson/Cane ICME catalog · NASA OMNI data

03 / Public-program data

A benchmark audit, not a fraud detector

A reproducible audit of Medicare aggregate data matched to active federal exclusion records, asking whether public labels support a narrow screening benchmark.

Eligible observations

About 1.8 million provider-year records across two measurement cohorts.

Narrow label

Only 12 eligible later matches for the specific health-care-fraud authority in the 2022 cohort.

Screening evidence

Cost screens captured none of those 12; one adjusted opioid-score match among 11 available labels is too sparse to validate performance.

Analysis of authority-specific join coverage and sparse captures across Medicare cost and opioid-rate screens — Label coverage · Fixed top-one-percent screens · Authority-specific outcomes

What the result supports

Label definitions, program coverage, and clinical specialty materially change the benchmark. The pilot identifies no new fraud case and validates no fraud detector.

Sources: CMS Part D prescriber data · HHS-OIG exclusion list

Research posture

OpenFind reports exploratory and negative results alongside the more promising patterns. The work does not infer causality from association, operational forecasting from retrospective prediction, or fraud from aggregate anomalies. Each pilot preserves source provenance and reproducible analysis records.