Plugins

Confidence Score Plugin

The Confidence Score plugin computes a dynamic trust score for each AI agent based on its history of approved, denied, and escalated actions. Scores drive policy routing — low-confidence agents are automatically sent for human review.

How It Works

0.9 – 1.0
High
Auto-allow
0.5 – 0.9
Medium
Human review
0 – 0.5
Low
Auto-deny by default

Scores are computed as a rolling exponential moving average over the last 100 actions. The decay factor is configurable.

Rego Integration

The Engine injects the current agent confidence score into every Rego evaluation as input.confidence.

policy/auth.rego
package lelu.authz

import future.keywords

# Require human approval for low-confidence agents
default allow := false
default require_review := false

allow {
  input.confidence >= 0.9
}

require_review {
  input.confidence >= 0.5
  input.confidence < 0.9
}

Configuration Options

Environment variables
# Minimum confidence to auto-allow (default: 0.9)
CONFIDENCE_AUTO_ALLOW_THRESHOLD=0.9

# Minimum confidence to require review (default: 0.5)
CONFIDENCE_REVIEW_THRESHOLD=0.5

# EMA decay factor (default: 0.1 = ~10 action window)
CONFIDENCE_EMA_DECAY=0.1

# Starting score for new agents (default: 0.7)
CONFIDENCE_INITIAL_SCORE=0.7