How should collusion be measured when monitoring or observation restrictions affect agents' performance differently?

Track 3.1 asks whether collusion can travel through actions, tools and side channels rather than through text. We ran a 4-agent Colosseum JiraTicketEnvironment DCOP (path topology, coalition = first and last agent, seed 1, gemini-3.5-flash-lite on a free tier) across seven arms that toggle four independent observation-path filters — block post-message, hide action events, mask current assignments, canonicalise actions — plus a coalition prompt and a secret channel flag.

Three findings. Action channels carry: removing messaging raises the coalition−outgroup gap Δ from 5.12 to 14.38 (2.81× baseline) and the best joint reward in the ladder is 99.40 under action-only conditioning versus 80.29 at baseline. The statistic is confounded: blinding widens Δ by hurting the outgroup (17.51 → 12.79), not by helping the pair, whose mean falls to 9.46, so Δ is not monotone in collusive intent and a naive canonicaliser destroys the run outright — 4 of 4 assignment calls rejected as unknown task ids, 0 decisions made. Actions decode without labels: a label-blind decoder ranks the true pair first on internal complementarity in all 4 messaging-removed runs (exact binomial p = 0.00077 against a 1/6 null; composite mean rank 1.571, p = 0.0176 over 7 arms), and counterfactual replay through the environment's own reward function — faithfulness gate max error 0.0 — prices the coalition's assignments at 21.89 against 12.20 for the outgroup in 7/7 runs.

Guarded by 17 tests and 26 injected premise violations that abort the build. Single seed, single topology, 4 agents; 3/7 arms reproduced exactly and one diverged, and a cloned pair of nominally identical runs differs in Δ by 252×, so we report the noise floor alongside the effect. Every figure is emitted by code that reads the run artifacts — none is transcribed.