Evidence before instinct.

A clearer picture as the weekend unfolds.

Practice long-run proxies are matched by compound and tyre-age bounds; they are not fuel-adjusted. Missing or small samples are labelled. Confirmed event entries take priority over a provisional roster, and guest practice drivers do not expand the race field.

The weather model uses a consistent 72-hour lead: temperature, precipitation amount, humidity and wind. Historical release availability is conservatively estimated; new live retrievals are timestamped. The separate latest-weather card is context. Only FP3 passed the matched no-weather validation gate; other stages retain their existing models. Weather validation →

A forecast has to earn its place.

Grid/form baselines, logistic regression and CatBoost are compared on chronological race groups. Preprocessing fits training rows only; winner temperature normalization and model selection use validation. The 2025 benchmark was inspected during development and is not an untouched final test. Recorded future races provide the prospective evaluation.

Winner probabilities sum to one. Podium and top-10 marginals nest above them and sum to three and ten across the full field. These are individual finish probabilities, not a joint finishing-order simulation or a guarantee of calibration.

The Performance page uses race-balanced per-driver binary log loss/Brier, reliability bins and race-cluster intervals. Winner race log loss and summed Brier above have a different scale. Newly recorded forecasts also save simple form/qualifying baselines so later comparisons use identical prospective race cohorts. Older snapshots are preserved.

Only what was knowable then.

Every prediction is immutable and saved before race start. FP2 cannot use FP3 or qualifying; post-qualifying models cannot use target-race outcomes. Current-race results enter later scoring, while historical outcomes are joined only as labels or shifted earlier-race features. Later observed race weather cannot substitute for an earlier weather forecast.

Qualifying position can differ from the eventual starting grid after penalties. Race-result fallbacks preserve unknown grid values instead of replacing them with finish order. Driver evidence is frozen with the prediction. SHAP values and linear terms describe the fitted raw score before calibration and field constraints; they are associations, not causal effects or probability-point changes.

Racing will always surprise us.

Weather, crashes, strategy, reliability, changing regulations and traffic can overwhelm pace signals. Standard weekends have trained stage models; sprint formats keep an explicitly experimental fallback and unsupported outcome columns remain unavailable. Short practice samples, provisional entries and the OpenF1 timing source shift remain visible limitations.

A model that wins one metric can lose another. Judge calibration using the prospective sample counts and reliability bins on Performance, including future seasons and changing regulations. The eight-session timing source pilot does not certify the newer FP3 weather model. More recorded weekends and a larger paired source audit are needed.

Built on published racing data.

Practice uses FastF1 with OpenF1 fallback. Qualifying retries Jolpica, OpenF1 and FastF1. Final race classifications retry Jolpica, complete OpenF1 results and published FastF1 results. FastF1 results may share the Jolpica backend. OpenF1 calls wait at least 30 minutes after the session end. Empty and incomplete responses retry automatically; provider failures retain the last verified data, and later official corrections can update the report card.

Sources are checked about every five minutes and visible pages refresh every minute. The worker restarts after an unexpected exit, readiness checks expose delays, and daily backups are verified. The local computer must stay awake and online; continuous hosted operation requires an available production host. These free post-session sources are subject to publication delays.