Confidence Bands

Three-state classification system that eliminates alert fatigue

Status: shipped — Stage 1 in shadow mode. Bands are computed and logged; they do not drive enforcement in production until Stage 1 exits shadow. See ML Roadmap.

Confidence Bands

The behavioral engine does not produce per-action risk scores. Instead, it classifies every action into one of three confidence bands. 95% of actions produce zero output — not a low score, but literally nothing.

The Three Bands

Band	% of Actions	System Response	User Sees
Band 1: KNOWN_SAFE	95-97%	Zero output. No score. No log entry.	Nothing
Band 2: UNCERTAIN	3-4.5%	Telemetry counter only.	Nothing
Band 3: ANOMALOUS	<0.5%	Enforce per security profile mode.	Alert or block

Why Not Scores?

Per-action risk scores create alert fatigue. Every action gets a number. Operators stare at dashboards. Alert fatigue within a week.

Three bands solve this: operators only see Band 3 sessions. The system handles Band 1 and 2 automatically.

Noise Budget

Metric	Healthy	Noisy	Blind
Band 1 (silent)	95-97%	<90%	>99%
Band 2 (logged)	3-4.5%	>8%	<0.5%
Band 3 (alerted)	0.2-0.5%	>2%	<0.05%

If noisy: Envelopes need more training data, OR corroboration threshold too low, OR group baselines missing.

If blind: Thresholds too loose, OR capability mapping has gaps, OR novel tools aren’t being classified.

Enforcement Modes

Each security profile configures how ANOMALOUS actions are handled:

Mode	Band 3 Action	Use Case
Strict	Block request	Production environments with strict security requirements
Balanced	Alert + escalate session	Default — flag to security team, monitor the session
Permissive	Log only	Development environments, initial shadow mode rollout

Shadow Mode

During initial deployment, the behavioral engine runs in shadow mode — it scores every action but never enforces. Both the existing risk engine and the behavioral engine produce results; the risk engine’s decision is authoritative.

This provides a calibration period to:

• Verify Band distribution matches the noise budget targets
• Identify false positives and adjust thresholds
• Compare behavioral signals against the risk engine’s scores
• Build confidence in the behavioral engine before enabling enforcement