Evaluation splits

Train, check, and hold out topologies with the same metrics: mission, SLA, false positives, exfil, kinetic failures.

netforge arena --json
netforge evaluate examples/submissions/heuristic_blue.py --seeds 0 1 --max-ticks 40
netforge benchmark --table --seeds 0 1 --max-ticks 40

Split	Public	Use
train	yes	learning
dev	yes	pick a checkpoint
hidden	organizer	official eval (repo ships a public proxy)

Blue is scored on keeping the network up without wrecking it (isolation / FPs). Red eval can use a population (random, heuristic, kill-chain), not one script. Your agent · Probes