UofA Adversarial Coverage Report

HIT HIT+ MISS WRONG CLEAN-CORRECT CLEAN-WRONG GEN-INVALID

View 1 — Catalog self-coverage (UofA pattern × subtlety)

COU-dependent rules: No rules show COU-dependent firing behavior at the 30% disparity threshold.

Pattern low medium high per-COU breakdown
W-AR-01 33% (1/3)
show
  • Morrison COU1: 33%
  • Morrison COU2: n/a
  • Nagaraja COU1: n/a

View 2 — Literature coverage (gap_probe outcomes by taxonomy)

No gap_probe rows; View 2 unavailable.

View 3 — Precision / recall summary

MetricValueSource battery
Catalog recall (HIT + HIT+)33.3%confirm_existing (n=3)
Catalog precision (1 − FPR)100.0%negative_controls (n=0)
Gap-probe MISS rate0.0%gap_probe (n=0)

Outcome class counts

ClassCount
COV-HIT-PLUS1
COV-MISS2

Performance characterization (v1.8 D2)

Hardware: hardware spec not configured (set UOFA_HW_SPEC env var)

MetricValue (ms)
Mean total_eval_ms1872
Median total_eval_ms1799
p95 total_eval_ms2035
Sample size3

Slowest 20 packages

spec_idvarianttotal_eval_ms
adv-2026-p2-001-w-ar-0112035
adv-2026-p2-001-w-ar-0121799
adv-2026-p2-001-w-ar-0131783

n = 3 package rows.