HYDRA

RED VS BLUE ARENA

Co-evolutionary payment-fraud lab. The attacker searches a 6×5×4 behaviour space for niches the detector cannot see; the detector answers without ever spending more than its 1% alert budget.

BLIND-SPOT COVERAGE
46.7%– 87.5%
lower = evaded / 120  ·  upper adds 49 unvisited
ROUND
baseline detector — trained on families F1–F3 only

RED TEAM

MAP-ELITES ATTACKER
genomes evaluated
200 archive elites re-scored against this defender
200
elites in archive
from Experiment 7's MAP-Elites arm, unchanged between rounds
200
cells breached
evasion > 0.5 and fidelity ≥ 0.60
56/120
mean fidelity
phi across the re-scored elites, simulator-calibrated
0.72

BLUE TEAM

DETECTOR @ 1% FPR
detected
on the clean holdout: 11,613 fraud rows
10,974
missed
false negatives at the same fixed rank cut
639
recall @ 1% FPR
on the clean holdout, not on the archive above
94.5%
false positive rate
fixed budget — a rank cut, never a threshold
1.00%

Blind-Spot Atlas

FOOTPRINT SLICE 1 OF 4  ·  30 CELLS SHOWN OF 120

one card, one holder — mimicry territory

TEMPORAL ↓AMOUNT →
dormant-trickle
low-and-slow
measured
active
fast
bust-out
micro-probe
below-baseline
at-baseline
uplifted
extreme
breached 56defended 15unexplored 49an unexplored cell is unknown, not safe — it sits in the upper bound
replays a pre-computed round — nothing is trained in the browser

Across rounds

BSC (interval)archive fitness

the same 200-elite archive re-scored against two defender states, with pairing broken on 200 of 200 - so neither fall above is a measured effect; the identity that does hold exactly is the alert budget, which never moves

held-out F4 transfer
THIS is the generalisation result, and it is not what the two rounds above show. Mean evasion on family F4 - withheld from every training set and prior - fell +0.0753 (0.5324 -> 0.4571) across 30 gated campaigns, 0 pairing broken.
+0.0753

Two REAL MEASURED DEFENDER STATES, not a multi-round closed loop: the attacker never re-searched between them, so this is the one hardening step the run performed. Round 1 is the pre-hardening defender, round 2 the retrained one, each re-scored against the SAME 200-elite archive. The source file labels this a MEMORISATION CHECK - these genomes were mined into round 2's own training set, so it answers 'did the patch take', NOT 'did the defender generalise'. The generalisation claim is the held-out F4 transfer result shown separately below. PAIRING IS BROKEN: the run's own counter reports 200 of 200 elites with pairing broken, so these two rounds are NOT a paired comparison and neither the BSC fall nor the fitness fall may be read as a measured defender effect. At cell level 29 of 71 cells were re-scored on a differently-generated campaign (median |dphi| 0.029, max 0.163), which affects 4 of the 8 cells that change state between rounds. No cell crosses the phi_min=0.6 gate (0 crossers), so no state change is caused by a fidelity change. See the source caveat.