P(t) Coupling Monitor

Detector = V10 (self-calibrating). Instead of a hand-set threshold, it runs all agents' health trajectories in parallel, learns the normal inter-agent gap from healthy data, and fires when the gap opens past k·σ — divergence z-score. This is why it works on gradual drift, where a fixed threshold is nearly impossible to tune. Re-convergence (gap closing) triggers resume. Monitor = V7 (one-sided, degradation-from-peak): each agent's health drops on degradation but isn't fooled by outperformance, and — unlike V1 — doesn't re-normalize through a sustained drop, so slow faults stay visible. Action = throttle (continuous, to the floor you set), with the freed spend either banked (Savings) or reallocated to recovery (Quality). Lineage: this is Paper 20's V10 detection + scaffold disposition, generalized to a continuous dial and ported to coupled agents. Cut A in-sim — proves the control logic given these dynamics, not that real agents behave this way. V7's exact formula is reconstructed from the published mechanism (degradation-from-peak); pending the CoVer-VLA script.