Mind · Inference
ConfidenceCalibrate
Aligning subjective confidence with objective frequency
Definition
Mechanism
Track Record Alignment: For claims rated 70% confident, ~70% should be true. Track predictions vs outcomes. If 90% claims are right only 60% of time, you are overconfident—widen uncertainty. If 90% claims are right 99% of time, you are underconfident—tighten. It adjusts the internal probability model using BayesUpdate on historical accuracy data, ensuring BaseRateInclude is respected.
Invariants
- Calibrates against a recorded track record of predictions and their outcomes.
- Expected calibration error stays within `calibration_target`.
- The adjusted confidence-to-frequency mapping is monotonic non-decreasing.
Risks
Failure modes
Over-correction
the Agent becomes under-confident to avoid being wrong, refusing to act on strong signals.
Thin record
calibrating across `bin_count` bins on too few predictions fits noise, not bias.
Relationships
Builds on
References
Autonomous Goal-Directed Entity
Outside view over inside view
Mathematically rigorous belief revision
Used by
References
Society
Credit claims by their calibration, not by the confidence of their delivery
Systematic categorization of ignorance
Reference
Parameters
bin_countInteger, [5, 20]- Resolution of calibration measurement
calibration_targetFloat, [0.02, 0.15]- Max expected calibration error, ECE