Policy Trends

Policy trend analysis is the practice of examining how your AI policy posture changes over time, not just whether a single session passed or failed. This guide explains how to read trend data effectively and integrate it into your regular review cadence.

Opening the Trends View

Navigate to Analytics → Trends. By default, it shows the last 30 days for all active applications.

Key controls:

• Date range — use 7 days for tactical investigation, 30 days for weekly policy reviews, and 90 days for quarterly reporting
• Application filter — compare specific applications or isolate a single system
• Metric selector — choose which metrics to overlay (policy score, risk distribution, guardrail activity, anchor coverage)

Interpreting Policy Score Trends

What healthy looks like

A healthy policy score trend for a stable, well-instrumented application shows:

• A score in the 80–95 range with low day-to-day variance
• Gradual upward movement over weeks as instrumentation coverage improves
• Brief dips (3–7 days) after model updates, followed by recovery

Recognizing concerning patterns

Policy Score vs. Risk Distribution

Governance score and risk distribution are complementary signals — read them together:

• High policy score + low CRITICAL/HIGH rate = strong policy operations and low-risk sessions
• High policy score + high CRITICAL/HIGH rate = strong policy operations, but the application is handling inherently risky interactions
• Low policy score + low CRITICAL/HIGH rate = controls are present but underperforming; instrumentation coverage likely needs improvement
• Low policy score + high CRITICAL/HIGH rate = the highest-priority improvement area; escalate to the responsible team

Using Trends in Regulatory Reporting

Many frameworks require evidence of continual improvement, not just a snapshot of current posture.

For EU AI Act Article 9 (post-market monitoring)

Export the policy score time series for all applications in scope for a given period. The export from Analytics can be attached as evidence of your post-market monitoring system.

For ISO 42001 Clause 9.1 (performance evaluation)

The 90-day policy score chart with the fleet target line overlaid is a direct artifact for the monitoring, measurement, analysis, and evaluation requirement. Caption it with your target, the period covered, and your assessment of trend direction.

For SOC 2 CC7 (system operations)

The guardrail activity trend — blocked events over time — demonstrates that operational controls were active throughout the audit period.

Setting Trend Alerts

Rather than manually checking the Trends view on a schedule, configure alert rules to notify you when trends cross thresholds:

• Score regression alert — fires when the rolling 7-day score drops ≥ 10 points in 48 hours
• Score below target — fires when the rolling average stays below your configured target for 24 consecutive hours
• Anchor coverage drop — fires when anchor rate drops below 98%

See Policy Rule Templates for pre-built rules for each of these scenarios.

Related Documentation