Seven day outlook
climatology plus decaying anomaly, not a synoptic forecast
Observed and forecast
90% bandConditions ahead
Zambretti prior + online logistic
Slope, not level, drives the forecast.
Records
Verification scorecard
skill above zero means it beats persistence · p/c/l is the ensemble weight
No matured forecasts yet. Rows appear as each horizon reaches its validity time: 15 minutes first, 24 hours tomorrow.
Temperature
self-heatingHumidity
element biasEstimator
Storage and weights
Station log
Kalman bank
NIS near 1 means honestly tunedCompensators
Learner bank
18 independent RLS headsFeature attribution
largest |θ| after standardisation, so these are comparable
Detectors
Climatology
Precipitation
Reliability
realised coverage against the 90% nominalSurroundings
a change here marks a discontinuity and queues a retrain
Indoors the building governs temperature and humidity, not the sky. Pressure passes through walls, which is why precipitation runs on tendency.
Closing a door changes how strongly the sensor couples to outside. The heads carry about 55 hours of memory, so tell them rather than waiting two days.
A thermostat makes the room a closed loop: it returns to the setpoint instead of drifting. Persistence is the wrong baseline for that, so this adds a fourth ensemble member and lets the Hedge weights decide if it earns its place. Humidity follows at constant vapour pressure, which is why a heated house is dry.
Matrix
24 costs about 11% of one core. 16 is still fluid and a third cheaper. Below 12 the crossfades judder.
Site and model
altitude feeds the sea-level reduction on every row
Moves RH from the element's temperature onto air temperature. Off by default: measured against a reference this board read 75.4% where the truth was 50.4%, so it reads high and this would push it higher.
Maintenance
How it is wired
select a stage to read its rationale