Two rivals, locked in a war neither can win. Every round one of them loses — and the strange law of the arms race is that the LOSER learns the most, catches up, and drives the next escalation. Neither pulls permanently ahead (you run just to stay in place), yet both climb far past anything either would become in peace — the two rise as a double helix, entangled, each strand's gain a rung that pulls the other up. Tap the fighters to spar, and watch the strife forge them.
A toy coevolutionary arms race with one real, checkable rule: the LOSER of each round learns more than the winner — its gain grows with how far behind it was (convex catch-up). So the two strengths stay neck-and-neck (the Red Queen: you must keep improving just to hold your place) while their absolute level ratchets upward, round after round, far above the flat 'peace' baseline (a rival who never fights stays at its starting strength). The ledger measures exactly that surplus — the height the conflict bought over peace. Documented dynamics: evolutionary arms races and the Red Queen (Van Valen, 1973), and adversarial (generator-vs-discriminator) training.
The strength units and the learn-from-loss constants are a toy model, not a fitted system; real arms races also collapse, deadlock, or spend both sides into ruin (the escalation can be a trap even as capability rises). The core — the loser learns most, the pair stays neck-and-neck yet climbs above peace — is the honest, checkable content.