You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Reading the numbers. Covered-vendor accuracy can be optimistic when labels and the KB share
sources, so the bias-resistant signals are: held-out (never-in-KB) accuracy, the baseline
comparison (does the KB beat a naive floor / a no-KB model?), the confusion matrix, and
calibration. Label-independence discipline is documented in eval/README.md.
Cases: 17 vendors · 102 control labels
Overall accuracy: 82/102 = 80.4%
Held-out (never-in-KB) accuracy: 17/36 = 47.2% across 6 vendors — the generalization number (no KB curation for these).
KB coverage: 53/102 = 52.0% (controls with a non-unknown prediction)
Baseline comparison
How much does the KB / the model actually add over a naive floor? (Same cases for all rows.)
Predictor
Accuracy (all)
Accuracy (held-out)
Coverage
naive (always-unknown)
29.4%
47.2%
0.0%
KB-only (deterministic)
80.4%
47.2%
52.0%
KB-verified-only (human facts)
47.1%
47.2%
17.6%
Per-control accuracy
Control
Correct
Accuracy
sso
11/17
64.7%
ulm
10/17
58.8%
entitlements
12/17
70.6%
riskSignals
17/17
100.0%
logout
17/17
100.0%
tokenRevocation
15/17
88.2%
Per-verdict precision / recall
Verdict
Precision
Recall
Support
supported
100.0%
68.6%
35
partial
96.4%
75.0%
36
unsupported
100.0%
100.0%
1
unknown
61.2%
100.0%
30
Confusion matrix
expected ↓ \ predicted →
supported
partial
unsupported
unknown
supported
24
1
0
10
partial
0
27
0
9
unsupported
0
0
1
0
unknown
0
0
0
30
Confidence calibration
Does predicted confidence track actual accuracy? (Well-calibrated ⇒ accuracy ≈ avg confidence per bucket.)