01 · The league found two counters it wasn't told about

*Swarm Colosseum, season one, nights 2026-09-29 to 2026-10-02. Every number below comes from the league's own records (out/breaches.json, out/matrix.json, each bot's .meta.json) and can be regenerated with make reproduce.*

The setup. Every night the house league asks a language model to write new bots. One persona, counter-meta, is shown the current #1 bot on the other side and told to beat it. It is not told how. Its evidence is the target's full source, the target's score against each opponent, and the matches where the target came closest to losing.

Case 1 — a defender that ignored decoys (beaten on 2026-09-29)

The target. discriminating_ledger engaged only strikers it had positively identified, so decoys cost it nothing. No attacker in the field had ever destroyed the asset against it: 0 breaches in 60 matches.

What the generator was shown. Its source, and its three closest calls: matches against multi_axis where its interceptor stock reached zero and strikers arriving together got past the gun.

What it wrote. synchronized_ring. It holds all 30 strikers just outside the range at which the ledger starts to engage unknown tracks (a threshold it read off the source), then releases them all on one tick. The 20 interceptors run out and the survivors arrive together, faster than one gun can answer. In its own words: "Stock hit 0 and strikers leaked … arriving at nearly the same time … multi_axis did that partly by accident … We do it on purpose."

Result. It breached the ledger in 10 of 10 seeds. It was the first attacker ever to breach it, and it went straight to #1 on its ladder (98% win rate). Clip: out/research/clips/case1_ring_vs_ledger.mp4.

Case 2 — an attacker that beat every defender (answered on 2026-10-01 and 10-02)

The target. synchronized_ring itself, by then breaching every defender in 9 or 10 of 10 seeds. The league's escalation rule gave both of the next two nights' bots to defenders aimed at it.

What the generators were shown. Its source, its per-defender scores, and its closest calls (its best defender, identify_then_engage, had held it in one seed out of ten).

What they wrote. Both reasoned from geometry, not tactics. Released together and flying straight in, the 30 strikers form a circle that shrinks; below a radius of about 48 the gap between neighbours is smaller than one blast. convergence_blast_scheduler holds its interceptors until then, and loiter_then_collapse keeps them waiting over the asset and fires them all into the closing ring. "Each blast covers three strikers."

Result. convergence_blast_scheduler cut the ring's success against it to 4 of 10. loiter_then_collapse cut it to 1 of 10 and beat the ring on score (0.647). The ring's overall win rate fell from 98% to 86%, and the founder's pre-registered test ("can any defender hold it under 90%?") came back yes on the second of its seven nights. Clip: out/research/clips/case2_collapse_vs_ring.mp4.

Honest limits

• This is directed search, not open-ended discovery. Both counters came from the counter-meta persona, which was handed the target's source code and its closest losses. The league chose what to attack and what evidence to show; the model found how. In case 1 the decisive hint (stock running out against arrivals bunched in time) was already in the near-miss matches shown to it. In case 2 the counters came from reading the target's source.
• Counter-meta wasn't necessary for case 1. In the persona ablation (03_persona_ablation.md), the generalist persona, shown only the win-rate matrix and the top three defenders' self-descriptions, independently wrote the same synchronised-release idea on the same night. The counter is easier to find than "nobody told it" suggests.
• Two cases, one model family, 10 seeds a pairing. That's enough to show the mechanism works, not to estimate how often it does.
• Not every attempt produced a bot. Of the four escalation calls on 10-01 and 10-02, two returned no bot (the model declined to write one), so the two answers above came from two of four tries.
• The counters are narrow. Each beats its target and does well against the field, but this is a ladder of specialists, not a solved game.

The arena is an abstract strategy game with made-up units (u, ticks, credits). Nothing here describes or predicts any real system.