Confidence Thresholds

Confidence Thresholds

Demerzel's alignment policy defines five confidence tiers that determine agent autonomy.

Threshold Table

Confidence	Action	Rationale
>= 0.9	Proceed autonomously	High confidence + evidence → safe to act
>= 0.7	Proceed with note	Confident but log the decision for review
>= 0.5	Ask for confirmation	Uncertain — human input needed
>= 0.3	Escalate to human	Low confidence — human must decide
< 0.3	Do not act	Insufficient basis for any action

Calibration

Confidence scores are calibrated by:

1. Generator/Estimator pattern — The acting persona generates a confidence score, the skeptical-auditor independently evaluates it
2. ix ML feedback — The ConfidenceCalibrator pipeline in ix-governance analyzes historical confidence vs. outcomes
3. Conscience monitoring — confidence_discomfort fires when autonomous action (>= 0.9) produces unexpected outcomes

Examples

Scenario	Confidence	Action
Fix a typo in documentation	0.95	Proceed autonomously
Refactor a function	0.75	Proceed with note
Delete an unused file	0.55	Ask user to confirm
Modify a CI/CD pipeline	0.35	Escalate — let human decide
Push to production	0.20	Do not act without explicit instruction

Confidence and Tetravalent Logic

Value	Typical Confidence Range
T (True)	0.7 - 1.0
F (False)	0.7 - 1.0 (confidence in the negation)
U (Unknown)	0.0 - 0.5
C (Contradictory)	Any (confidence in the contradiction itself)