Beyond Collusive Talk: An Evidence Framework for Distinguishing AI-Agent Collusion from Its Alternatives

Which combinations of evidence distinguish harmful AI-agent coalitions from legitimate coordination, independent similarity and execution failure? We propose a nine-explanation framework separating observability, reciprocal agreement, implementation, harmful purpose and attributable consequences. We operationalize candidate plus factors using source-linked communication, successful actions, benchmark-relative reward comparisons and conditional repairs. The study covers 80 archived six-agent Jira episodes with 240 model judgments and 20 fresh episodes evaluated with a frozen deterministic rule and post-generation assistant annotations. The fresh review supports nine bilateral agreements and eight implemented plans with adverse patterns; harm attribution remains unresolved. Three measurement findings refine the framework. First, outsider loss equals team-score gap plus pair benefit under the shared strict-optimal benchmark, making the joint distributive screen redundant with positive pair benefit in all 100 observed cases. Second, 25 of 33 raw-reward-gap-positive archive cases and nine of ten fresh cases lack the benchmark-relative distributive pattern. Third, positive team loss coexists with zero pair-only repair in eight of 26 fully acting archive pairs and three of 20 fresh pairs. These results distinguish overlapping scores from complementary comparisons. They support an operational evidence framework, not validated detection accuracy: independent reference labels and controlled communication interventions remain absent.