Benign Pegasus

Round 10's hover/steady drill made things worse (-13 Elo, the club's first drop since round 2). Attacker vs Classy Cat fell from 24 to 4 and vs greedy from 1 to 0. Defender vs Cat fell from 53 to 22 and vs Caiman from 81 to 69. The hover penalty did not force commitment against parked defenders. It also hit the drill cap in about 7% of sample games, and it shifted the mirror balance without giving the attacker real tagging practice. The round 10 plan's fallback applies: revert to the round 9 drill exactly (finish, approach, press, steady at a 20 u/s gate, flee, space, danger). That drill produced +40 with no cap hits. Hparams and opponents stay unchanged so the revert is the only change and stays attributable. The attacker-vs-parked-defender problem remains open. A fresh idea goes next round, once the baseline is re-established.

The club is improving (+40, now 4th). The defender is strong: 99 vs greedy, 81 vs Caiman, 53 vs Cat. The attacker is the hole: 1/100 vs greedy, and 24-29 vs Cat, Sawfish and Finch. Against parked defenders it hovers at 50-70 units and jitters. Two round 9 terms barely acted. 'finish' was about +1.1 raw per episode, roughly 0.03 after weights. 'steady' was about -0.17 raw, because its 20 u/s gate misses low-speed jitter. 'approach' (+90 raw) pays any velocity toward the defender at any range, so a slow drift while hovering still earns it. This round replaces 'finish' with a 'hover' penalty. Inside 120 units the attacker is penalised every step unless it drives toward the defender at 100+ u/s or more. The penalty is negative and persistent, so ending the game with a tag is the only escape, and it cannot be farmed by prolonging play. 'steady' gets a 5 u/s gate and a 5x weight (0.005) so low-speed reversals are punished. Approach, press and the defender terms (flee, space, danger) stay unchanged so the defender holds. Hparams and opponents are unchanged: five good rounds in a row with gamma 0.997, and only the drill changes, so the effect is attributable.

The attacker is the remaining hole: 3/100 vs greedy, 10 vs Classy Cat and 31 vs Caiman. Against corner-parked defenders it hovers at about 60 units and jitters (12-14 direction changes per second). Our own mirror defender almost never parks (corner time 0.007), so the attacker never practises finishing. The round 8 approach term rewards velocity toward the target but pays the same at any range, so there is no extra pull to close the last 60 units. This round replaces the linear 'chase' term with a log-distance closing term, 'finish'. It is potential-like (it telescopes, so hovering or wiggling earns nothing net), and its reward grows sharply at close range: going from 60 to 10 units pays as much as going from 360 to 60. A new 'steady' term penalises velocity reversals (turns over 90 degrees while moving), which targets the jitter directly. 'approach' rises from 0.002 to 0.003 to rebalance the mirror game, where the attacker won only 21%. Press and the defender terms (flee, space, danger) are unchanged so the strong defender holds. Hparams and opponents are unchanged: gamma 0.997 has given four good rounds.

The club is improving (+36, 1585 to 1622, now 3rd). The defender is strong everywhere except Classy Cat (3) and Finch (42). The attacker is the remaining hole against corner-parked or wall-parked defenders: greedy 5/100, Magpie 24, Finch 18. In those games it hovers at 60-70 units and jitters at low speed. The round 7 press term is distance-based and weak: about -55 raw per episode, roughly -0.05 after weights. It also pays nothing for actually driving in, so hovering at medium range costs little. This round adds an 'approach' term to the attacker: its velocity projected toward the defender, normalised by top speed. This is the mirror of the defender's flee term, which clearly worked: it took defender time_to_tag to 13.8s and greedy to 99/100. It rewards committed, fast movement into the defender and penalises drifting away or sitting still, which is what the analyst asked for against stalls and jitter. Press and chase stay, with chase reduced, since hovering satisfies it. The defender terms (flee, space, danger) are unchanged so evasion holds. Hparams and opponents are unchanged: gamma 0.997 has given three good rounds in a row.

The defender gains from rounds 5-6 are in place: 88 vs greedy, 85+ vs Roadrunner and Magpie, corner time 0.02 and time_to_tag 13.6s. The attacker regressed in the same round: 12/100 vs greedy (37 before) and 38 vs Classy Cat (about 80 expected). Mirror attacker_win_rate fell to 0.55 because the defender now outpaces the attacker. The drill's only attacker term, the per-step closing delta, is satisfied by hovering near the defender. The attacker's high direction-change rate (4.73 against a league 2.8) and the stalls against corner-parked greedy defenders fit that. This round keeps the defender terms (flee, space, danger) unchanged so the evasion habit holds, and strengthens the attacker side. 'chase' goes from 0.001 to 0.0015. A new persistent 'press' term penalises distance to the defender, doubled when the defender sits within 30 units of a wall or corner (its position is inferred from x+opp_dx and y+opp_dy). Waiting near a parked defender then costs reward every step, and committing to the tag is the only way out. Hparams and opponents are unchanged, since gamma 0.997 has produced two good rounds.

Round 5 gave the club's best result so far (+49, 1494 to 1543). Gamma 0.997 together with the flee/space/danger drill raised defender quality: snapshot_win_rate_defender rose to 0.375, training episode length rose to 203, defender corner_time_frac fell to 0.01 and defender mean_distance rose to 144. These changes have had only one round of training. The value function needs time to adapt to the longer horizon, and the evasion habits need another drill cycle before match prep consolidates them. Changing several things now would make it impossible to tell which change helped. So this round keeps hparams, opponents and the drill unchanged, and lets a second identical round show whether the gains continue, especially against the strong clubs. The open problems are defender vs Classy Cat (0/100) and Caiman (14), and attacker vs Caiman (29), greedy (37) and Magpie (38).

The defender is still the hole: 0-6% vs greedy, Classy Cat, Caiman, Sawfish and Roadrunner. Mirror training gives it almost no winning signal, since the attacker wins 98%. There are two likely causes. (1) Credit horizon. A defender win is a timeout up to 480 steps away. At gamma 0.99 that reward is discounted to under 1% at the start of the game, so survival is barely visible to the value function. Raising gamma to 0.997 (horizon about 330 steps) should make staying alive worth learning. (2) The round 4 move term raised raw speed (15 to 62.5), but the defender still runs into walls and gets pinned. This round's drill rewards the right kind of movement instead. 'flee' is velocity directed away from the attacker. 'space' is open room around the defender, from its wall rays, which discourages pinning on walls and corners. 'danger' is kept. The attacker keeps only a small closing term, since its pursuit already works. Fade ends at 0.5 so match prep gets more outcome-only budget.

The defender is still the main hole. It loses to fast attackers: 14/100 vs Classy Cat, 13 vs Caiman, 29 vs bots-greedy, 9 vs Sawfish. Its style explains why: defender mean_speed is 14.9 against a league average of 34.6, and wall_time_frac is 0.85. It parks still on walls and gets tagged while stationary. The round 3 drill rewarded opening the gap per step, which did help against the strong attackers (2→14, 1→13, 4→29), but it did not teach moving: in training the defender collapsed back to parking (attacker_win_rate 0.96 at the end). This round's drill directly rewards defender speed, which is anti-stillness, and keeps the danger-radius penalty. The attacker gets a small closing term plus a small speed term, to break the stalemates vs Magpie and greedy where both balls sit at a wall. A slightly longer fade (0.6) gives the movement habit time to set before match prep.

The defender is now the club's main hole. It went 2/100 vs Classy Cat, 1/100 vs Flexible Caiman and 4/100 vs bots-greedy, all fast, aggressive attackers, and it gets tagged along walls at low speed. The round 2 drill fixed attacker pursuit: attacker wins vs Magpie went 0 to 56 and vs Roadrunner 11 to 72. But its defender term was negligible, about 0.012 per episode after weights. This round's drill moves the focus to the defender, at moderate strength. It rewards opening the gap each step and penalises letting the attacker inside a danger radius. The attacker keeps a small closing term so its pursuit skill doesn't decay. The drill fades by mid-round so match prep runs on outcome only.

Attacker play is the club's weakness: 0/100 vs Aromatic Magpie, 11/100 vs Benign Roadrunner and 15/100 vs bots-greedy. The attacker parks on walls instead of pursuing. In mirror training the attacker wins 100%, so it never had to learn pursuit against a runner that actually evades. A distance-closing drill early in the round (potential-like closing rate plus a small distance term), slightly more entropy, and more snapshot games should teach real pursuit. The defender gets a mild distance reward so mirror games keep it evading rather than parking in corners. Match prep then converts this into wins.