Pillar III — Uncertainty Crystallization (UC)

Core idea

Epistemic miscalibration — where a model is wrong but confident — encodes structural information about overrepresented training distributions. UC converts this into a calibrated, statistically guaranteed confidence oracle using three stages: MC-Dropout, Temperature Scaling, and Conformal Prediction.

Usage

from phantasm.core.uc import UncertaintyCrystallizer

uc = UncertaintyCrystallizer(model, mc_samples=30, coverage=0.90)
crystal = uc.crystallize(input_ids)

print(crystal.reliability_tier)           # "crystal" | "solid" | "fluid" | "vapor"
print(crystal.calibrated_confidence)      # float in [0, 1]
print(crystal.conformal_interval)         # (lower, upper) — 90% statistical guarantee
print(crystal.action_recommendation)      # human-readable next step

Reliability tiers

Tier	Symbol	Condition	Action
Crystal	◆	conf ≥ 0.85, epistemic < 0.05	Use directly
Solid	◇	conf ≥ 0.65, epistemic < 0.15	Light verification
Fluid	≈	conf ≥ 0.40, epistemic < 0.35	Verify before use
Vapor	~	conf < 0.40	Do not use unverified

CrystalizedUncertainty fields

Field	Type	Description
raw_confidence	float	Original uncalibrated confidence.
calibrated_confidence	float	After temperature scaling.
epistemic_uncertainty	float	MC-Dropout variance.
aleatoric_uncertainty	float	Expected entropy (irreducible).
total_uncertainty	float	Epistemic + aleatoric.
conformal_interval	tuple	(lower, upper) with guaranteed coverage.
reliability_tier	str	One of four tiers.
action_recommendation	str	What to do with this output.