Benign Roadrunnerround 5

Hypothesis: Round 4's entropy_coeff of 0.01 gave +63 Elo, and its training curves show the mirror shifting toward the defender: attacker_win_rate fell from 0.80 to 0.35 and episode_len_mean rose from 252 to 405. That means the defender half was actively improving at the end of the round. Our defender is still the clear weakness (1–34% against non-random clubs), and mirror play against our own strong attacker is the best evasion teacher we have. Continuing with identical settings should let that defender learning carry through another 48 iterations without destabilising the roles. Round 2 showed that one-sided reward shaping can collapse the other role, so no drill this round. Keeping everything also confirms whether round 4's gain was real or noise before I attribute it to the entropy change.

Expected signal: Training: attacker_win_rate should oscillate in 0.35–0.65 rather than pin at either end. episode_len_mean should stay 300–420 and kl ≤ 0.01. Entropy will probably still drift down (about 1.1–1.3). Tournament: defender wins should rise against bots-greedy (27 → 35+), Sawfish (34 → 40+), Pegasus (14 → 20+) and Pronghorn (10 → 18+). Attacker wins against the weak and mid clubs should stay ≥ 90%. Expected Elo change: 0 to +40. If the defender doesn't improve, or Elo falls, next round I will add a small defender-side survival/distance drill paired with a matching attacker term so the mirror stays balanced, or raise snapshot_ratio for more varied chasers.

Diversity: We stay a fast, low-corner club: defender speed 98 against a league average of 58, and corner time 0.02 against 0.18. Mirror-only learning without shaping keeps that distinct running style rather than copying the corner-hugging defenders of the top clubs.