skip to content
Surprise memory

What the models did not capture, and what we did about it

Magnitude of a surprise and novelty of a surprise are stored separately. Similarity uses a normalized feature-vector distance placeholder pending real embeddings.

Demo modelSynthetic dataResearch preview

Surprise state over time

Magnitude and novelty are stored separately; markers show where a regime was opened or re-fitted.

— impact-- novelty○ regime change

Fuel-state regime break, basin W-09

Late-season dryness with vegetation curing far outside the modelled seasonal envelope.

What happened?
Late-season dryness with vegetation curing far outside the modelled seasonal envelope.
Which expectation failed?
Historical analogue assumed curing plateaus after week 32.
What happened next?
Spread rates stayed inside physical bounds but at the upper constraint for 11 days.
Was the model updated?
Yes — fuel-state prior widened, analogue reference down-weighted.
Resolution
New regime opened; analogue reference re-weighted for late-season curing.
Vegetation / fuel stateHistorical analogue

similar previous surprise: 94% · embedding placeholder [0.71, 0.42, 0.18, 0.66, 0.29]