Escalation Analytics Panel

The Escalation Analytics panel provides insight into how the Value of Information (VOI) engine is performing — whether escalations to human review are producing meaningful governance improvements.

Escalation Volume Trends

The summary strip shows key escalation metrics at a glance:

• Total Escalations: Count of all escalation events

• Verdict Changed: Escalations where the human reviewer changed the automated verdict

• Verdict Unchanged: Escalations where the human agreed with the automated decision

• Change Rate: Percentage of escalations that resulted in a different outcome

A rising change rate suggests the VOI engine is correctly identifying uncertain decisions. A consistently low change rate may indicate the escalation threshold is too aggressive — the system is escalating decisions it could handle autonomously.

VOI Distribution

A histogram showing the distribution of VOI scores across all escalations:

• Low VOI (0.0-0.3): High confidence decisions that were near the escalation boundary

• Medium VOI (0.3-0.6): Moderate uncertainty — standard escalation territory

• High VOI (0.7-1.0): High uncertainty — dimensions disagree, trust is low, or stakes are high

Interpreting the Distribution

• Left-skewed (most scores < 0.3): VOI threshold may be too low — many escalations are unnecessary

• Right-skewed (most scores > 0.7): VOI threshold may be too high — only extreme cases escalate

• Bimodal: Two distinct escalation patterns — investigate whether different agent archetypes or risk levels drive each peak

Escalation ROI

A pie chart showing the proportion of escalations where human review changed the verdict vs. confirmed the automated decision.

• Verdict Changed (yellow): The escalation was valuable — human oversight corrected a governance error

• Verdict Unchanged (green): The escalation was unnecessary — the system would have made the same decision

Target: 20-40% change rate indicates well-calibrated escalation. Below 10% suggests over-escalation; above 60% suggests the system's automated decisions need improvement.

Reviewer Utilization

A table showing per-reviewer metrics from the HumanDriftMonitor:

Metric	Description
Total Interactions	Number of reviews performed
Approval Rate	Percentage of reviews resulting in approval
Mean Review Duration	Average time spent per review
Reviews/Hour	Review throughput

High approval rates (>95%) combined with fast review times may indicate rubber-stamping. The HumanDriftMonitor tracks these patterns — see the drift detection documentation for more detail.

Agent Escalation History

Enter an agent ID to see its specific escalation history, including:

• Timestamp: When the escalation occurred

• Reviewer: Who reviewed the escalation

• Decision: What the reviewer decided (approved, denied, modified, deferred)

• Risk Level: The risk level of the action that triggered escalation

• Review Duration: How long the reviewer spent on the decision

• Rationale: The reviewer's explanation (if provided)

How to Use Insights to Tune VOI Config

If change rate is too low (< 10%)

The system is escalating too many decisions that don't need human review:

1. Increase min_voi_to_escalate in VOIConfig (default: 0.3)

2. Lower entropy_weight if dimension disagreement is driving unnecessary escalations

3. Raise trust_floor only if low-trust forced escalations dominate

If change rate is too high (> 60%)

The system's automated decisions are frequently wrong:

1. Review dimension weights — some governance dimensions may need recalibration

2. Check trust calibration — agents may have inflated trust scores

3. Decrease min_voi_to_escalate to catch more uncertain decisions

If reviewer fatigue is detected

1. Rotate reviewers using RoutingRecommendation suggestions

2. Reduce review volume by raising the escalation threshold for low-risk actions

3. Add secondary reviewers for high-risk escalations

API Endpoints

Endpoint	Method	Description
/v1/escalation/analytics	GET	Fleet-wide escalation metrics
/v1/escalation/analytics/{agent_id}	GET	Per-agent escalation history

Screenshots

[Dashboard screenshots to be added]