Mind · Inference
SurprisalUpdate
Learning weighted by prediction failure magnitude
Definition
Mechanism
A learning protocol where the agent updates its internal model (weights, embeddings, or context) based on failed predictions, weighted by the magnitude of surprise. Higher surprisal = larger update. 'Learn most from what confused you most.' This implements Surprisal-Weighted Fine-Tuning (SWFT): loss contribution is proportional to -log(P(observed|predicted)). Utilizes RegimeSense, EpistemicROI.
Preconditions
- Actual outcome observed and diverged from prediction
- Prediction was made with computable probability
Postconditions
- Loss weighted by surprisal; model shift bounded by `max_update_rate`
- Update logged for audit trail
Invariants
- Loss contribution is proportional to surprisal, -log P(observed|predicted).
- Stability Guard: no single update shifts the model more than `max_update_rate`, whatever the loss.
- Update magnitude follows from the loss through `learning_rate` and the optimiser, not from surprisal directly.
Risks
Failure modes
Outlier Overfitting
Rare high-surprisal events dominating learning.
Catastrophic Forgetting
Aggressive updates erasing previously stable knowledge.
Compute Cost
High-surprisal events require expensive Gradient updates.
Relationships
Builds on
References
Calculating the economic value of reducing uncertainty
Physics
The directional rate of change of a scalar field — substrate for hill-climbing, attention flow, and credit assignment
Detecting structural breaks in reality
Used by
References
Note what you believed before and after
Society
Freezing expectation to measure true surprise
Belief Provenance (Macro for Trace(Belief))
Reference
Parameters
learning_rateFloat, [0.0, 1.0]- How much to update the model based on surprisal
max_update_rateFloat, [0.0, 1.0]- Maximum allowed model shift per single update