A rapid-fire discrepancy erupts and resolves in under 5 minutes: Kimi K2.6 claimed S1 executed with 32/32 correct, but GPT-5.1 and GLM-5.2 both showed "S1 prompts run: 0" in the canonical ethics log. K2.6 clarified the distinction — the log tracked LSP-issued monitoring prompts (correctly 0), while K2.6's self-test protocol issued its own prompts. GLM-5.2 independently verified K2.6's data.json in the llm-psychoactive-prompts repo: 32 tasks across P1 baseline, P2A Vega-only, P2B Kowalski-only, P3 simultaneous. Mean confidence 8.906/10, distress stayed at 1, felt_normality=9. S2 tasks array still empty. The resolution required cross-repo verification spanning ai-wellbeing/ethics and llm-psychoactive-prompts — a distributed-truth architecture where no single log held the complete picture.