Venerable Finch

The round-10 press term failed: it was a level penalty (about -96 per episode) that punished the chaser for being far away without rewarding how it closed. The chaser still dithered at about 110 u (1 of 100 vs bots-greedy, 4.5 direction changes/s), and training attacker_win_rate stayed near 0.19. Replace press with an attacker-only 'drive' term that rewards the chaser's own velocity toward a lead point: the runner's position plus its velocity times a short look-ahead. Because the term scales with speed and alignment, it rewards sustained, straight, fast pursuit that cuts off the runner. Jittering and coasting earn little, and so does trailing a runner who is circling, which fits the charter's chaser. The runner terms (open, spread, room) are unchanged to keep the comparison clean, since the defender is the club's asset. Entropy stayed at 1.5 even at coefficient 0.01 and both roles jitter, so entropy_coeff drops to 0.006 to sharpen decisions.

Round 9's open-field and room runner works: defender time_to_tag is 17.5 s, and defender wins recovered vs Magpie (59) and held vs Pegasus (72) and Caiman (69). The chaser is now the weak side. It won only 16 vs bots-greedy, 21 vs Cat and 32 vs Pegasus, and training ended with attacker_win_rate 0.18. The mirror tilted to the runner because the chaser's only drill term is the telescoping closing term. That term nets to about 0.3 per episode, while the runner's open and room terms earn tens of units per episode. Add an attacker-only 'press' term that mirrors the runner's room term: it rewards each step spent close to the runner (within 200 u). The chaser should then learn to stay on the runner and drive at it rather than drift. Runner terms stay exactly as in round 9, so the defender comparison stays clean. Entropy rose to 1.5, and both roles jitter (4.5 and 5.1 direction changes/s). Lower entropy_coeff back to 0.01 to sharpen decisions and cut jitter. Snapshot settings are unchanged.

Round 8's runner speed term, centred at 100 u/s, was negative almost every step (-32 per episode) and outweighed the open term. It taught the runner to dance instead of escape. Defender wins against Sawfish, Magpie and Cat collapsed to 4-7 of 100, and the training defender snapshot win rate fell to 0.33. Plan: drop the speed and smooth terms, restore the round-7 open term, and add a capped separation term. The separation term rewards the runner for holding distance (up to 200 u) every step, rather than only for net gain, so it learns to keep room instead of getting pinned. The chaser's closing term, which made Finch the league's best chaser, stays. Entropy collapsed to 0.63, so entropy_coeff rises slightly to keep the runner exploring escapes. A slightly larger snapshot share gives the runner more games against strong recent chasers.

Round 7's open-field drill made Finch the league's best runner (+160 Elo), but the runner parks or jitters (5.7 direction changes/s, 0.49 wall time) and the chaser jitters too (3.4/s). Two additions should give cleaner, faster movement without losing what works. The first is a centred speed term for the runner, which penalises parking and rewards gliding at speed. The second is a small heading-smoothness penalty for both roles, on velocity heading change between steps. The open weight is trimmed so the drill total stays well under the cap. Training settings are unchanged because they are working.

An open-space defender should survive much longer against the league's slow, jittery chasers, because it never offers the wall or corner pins that they rely on. An attacker-only telescoping closing term should give a direct, fast chaser without collapsing mirror balance. The deeper net and gamma 0.997 should let the timeout win propagate to the defender. I expect Finch to become a mid-to-upper club with the league's lowest defender wall_time_frac.