Defend · Hardening progression

Adversarial campaigns become training pressure for the next defender generation.

Held-out discrimination and recall improve from V0 to V2 while the legitimate false-positive rate remains near its validation-calibrated operating point.

V0 held-out ROC-AUC

0.98063

95% CI 0.97555 to 0.98529

V1 held-out ROC-AUC

0.99384

95% CI 0.98979 to 0.99745

V2 held-out ROC-AUC

0.99993

95% CI 0.99984 to 1.00000

ROC-AUC

Five seed paths plus seed-level mean.

Mean auc by defender generation
V0 mean	0.98063
V1 mean	0.99384
V2 mean	0.99993

Recall

Five seed paths plus seed-level mean.

Mean recall by defender generation
V0 mean	86.59%
V1 mean	96.06%
V2 mean	99.90%

Legitimate FPR

Five seed paths plus seed-level mean.

Mean fpr by defender generation
V0 mean	0.911%
V1 mean	1.007%
V2 mean	0.982%

Fresh-attacker approved value across generations

The boundary condition is written inside the visualization, not left to a tooltip.

Matched-pool caveat: the action pool is fixed and held constant across generations.

Fresh-attacker approved value by generation and seed
Seed	V0	V1	V2
20260812	$59,952	$267	$0
20260813	$38,133	$4,100	$0
20260814	$27,549	$0	$0
20260815	$12,214	$2,417	$0
20260816	$18,048	$1,579	$0

Validating artifact schemas and preparing the command center…