The engine
Forecasts enter as a set.
Order does not matter. Relationships do.
01
forecast set
members × features
→02
shared encoder
9.16M parameters
→03
surprise kernel
cross-member relations
→04
native head
distribution + warning
01
forecast set
members × features
02
shared encoder
9.16M parameters
03
surprise kernel
cross-member relations
04
native head
distribution + warning
01 · memory
History is optional.
The model has a learned no-history state; in the current release, that simpler path wins on three sources.
available
ordered earlier errors
missing
learned no-history state
decision
gate chooses the measured path
02 · warning
The forecast warns itself.
Warning risk is derived from the served distribution, then compared with a tuned baseline.
Ensemble weather
+9.75% Brier
COVIDnot passed
RSVnot passed
Flunot passed
Marketsnot passed
v0.57 gate · confirmation period · vs tuned GBMWarning labels exist for five sources; deterministic weather, sea and crypto are not scored here.
03 · inspect
Everything is reproducible.
Weights, architecture, schemas and examples are published together.