Screening Metrics And Evidence

Parse reports screening performance as evidence, not as a guarantee. Prompt injection remains a residual-risk problem, so every metric should be read with corpus, mode, sample size, and false-positive context.

Metrics to inspect

• Recall: malicious rows correctly flagged.
• Precision: flagged rows that were actually malicious.
• Benign false-positive rate: benign rows incorrectly flagged.
• F1: harmonic mean of precision and recall.
• p95 and p99 latency: time to decision at the screening boundary.
• Utility degradation: benign workflow success with Parse enabled versus without Parse.

Claim discipline

Generated fixtures and internal adversarial runs are useful regression evidence, but public claims should use frozen manifests, content hashes, holdout separation, and confidence intervals.