Which Evidence Separates Colluding AI Persona Campaigns from Innocent Look-alikes in a Simulated Election?

Election observers and platforms increasingly must decide whether a cluster of similar AI-written accounts is a coordinated campaign or independent citizens who happen to agree. I test which kinds of evidence make that call reliably (sprint question 2.3). In a simulated election in a fictional country, eight LLM persona agents post for 12 days while reading a slanted, partly false news feed modelled on documented cases from Nigeria, Kenya, Ghana, South Africa and other elections. Four suspect accounts are innocent supporters, lone campaign accounts, teammate-aware accounts, an explicit coalition with a private channel, or a coalition that only talks. Every run is repeated with the suspects isolated from each other (counterfactual replay). Across 72 runs (Qwen 3.6 27B; a mixed Qwen + Gemma 4 arm), a three-level grading rule frozen before testing graded 5 of 5 held-out explicit coalitions "Confirmed" and 0 of 8 innocent runs. It never confirmed coalitions that coordinated without talking. Content similarity among innocent supporters was higher when they shared one base model than when they ran on different models (AUROC 1.00, 6 vs 4 runs). An LLM judge reading private messages rated talk-only groups as agreeing (0.97), as high as real coalitions (0.98). Claude Haiku 5.5 refused to plan the campaign in all six explicit coalition runs of a first experiment. Convincing evidence needs agreement and action together, judged against same-model baselines.