Risk Model — Calibration & Score Bands

Note (0.4.4+): The scan verdict (ALLOW / FLAG / BLOCK) is no longer derived from risk_score crossing a threshold. The verdict is now derived from finding classes — see Risk Scoring & Verdict Model. The numbers below remain useful as customer-facing risk bands on dashboards, but they do not gate the verdict.

Threshold Calibration (analytics-only)

Auto-generated by scripts/calibrate_thresholds.py. Last refresh:

Dataset: 1185 records (986 malicious, 199 benign)

ROC-AUC: 1.0000

The numbers below are the operating points that best separate benign from malicious in this corpus; they're a useful reference for dashboard band-labeling but they are not the BLOCK gate any more.

Risk-band reference

Band label	Risk score	Suggested UX
`clean`	< 0.25	Display as green / pass
`notable`	0.25 – 0.50	Yellow / show findings drawer
`elevated`	0.50 – 0.70	Orange / encourage manual review
`severe`	≥ 0.70	Red / route to senior reviewer

These thresholds map to the historical config.thresholds.flag / config.thresholds.block values, which are still settable on ScanConfig (flag = 0.25, block = 0.70 by default). Setting them no longer changes which files BLOCK — it only changes how the score is labeled if you display bands in a UI.

How to apply (UI-only, not enforcement)

# These set the *labels* the band-renderer uses. They do NOT decide which
# files get blocked — see concepts/risk-scoring.md for the actual verdict
# rules.
config.thresholds.flag  = 0.25
config.thresholds.block = 0.70

If you want a file to be blocked, ensure at least one of its findings carries verdict_class = BLOCK. The detectors that emit BLOCK-class findings are documented in Risk Scoring & Verdict Model.

Re-run scripts/calibrate_thresholds.py whenever your benchmark corpus changes if you want the band boundaries to track its statistics.