Engine
What production actually published, scored against realised outcomes — not what a fit promised on paper
Weekly station health
Tier journey, tripwire state, which watched stations are healing, and the hardest stations of the recent window.
October regime watch
Season-change evidence pack, refreshed weekly: H16 form and benched stations, per-station bias-correction load, the live gated record vs the fraction trigger, and the paper dip book as a no-risk window into tail pricing.
Centre accuracy
Centre accuracy — how far off the published high is, and what a thermometer-only baseline would have missed
Interval coverage
Interval coverage — every forecast promises a 50% window (p25–p75). Split history into what landed below, inside, or above: each side should sit near 25%. Too much below = forecasts run warm; too much above = they run cold.
Bucket distribution quality
Bucket distribution quality — Polymarket settles on buckets (“exactly 28°”, “≥29°”), so this scores the full probability curve the trading layer bets with, against a bland same-centre baseline. Zero-mass on truth must stay at 0%.
Analog-pool bake-off
Analog-pool bake-off — the champion (12-member analog pool) against a 24-member shadow candidate, scored on the same settled outcomes before anything is promoted. Bucket shape skill is the engine's weakest pillar; a wider pool is the lever. The experiment writes the scoreboard the moment it starts.
Correction shadow scoreboard
Correction shadow scoreboard — the bias-correction layer's candidate product (corrected centre and band) scored on the same settled outcomes as the raw view above. The site publishes the raw number; this card is the evidence board for whether the correction is earning its keep. Shadow MAE above raw MAE = the correction is hurting.
Learning & adaptation
What the engine learned and when — which correction method prices current editions, whose adjustments moved this week, how often the peak gate engages, and the warm-skew tripwire with its daily walk. Every number re-derivable from the ledgers.
By lead time
By lead time — how far ahead each forecast was issued. Physics makes error grow with distance; the job is keeping the band balanced at every distance, where long-lead bias likes to hide.
By prior lane
By prior lane — every forecast blends an external daily-max estimate into the analog centre: the human-issued TAF TX, an NWP model, or nothing at long lead. Each lane is scored separately, with the prior's own error next to the analog's — the two numbers that say which side of the blend is sick. Last 14 days; “unrecorded” is the pre-prior era.
By code era
By code era — every engine change gets its own fingerprint. Everything published under one version groups into one row: an honest before/after of every change, scored on real outcomes.
Stations
Stations — same engine, different difficulty. Hardest rows are volatile airfields; sharpest are steady climates. Minimum 100 scored editions to rank.
Hardest
Sharpest
By weather regime
By weather regime — clear skies vs overcast vs storms behave differently. The worst row is where new features pay most.
Bucket quality by store date
Bucket quality by store date — distributions grouped by when they were written. After any engine change, the newest rows show the verdict while it is still news.
Probability calibration
Probability calibration — every bucket probability ever published, grouped by claimed size and checked against reality. Predicted 20%, happened 21% = calibrated. Gaps beyond ±5% mean confidence needs shrinking.