Adapting to drift you cannot see: passive identifiability decides whether online regime inference pays off in epidemic control

Abstract Public-health decisions about when to intervene against a respiratory pathogen are made repeatedly under conditions that move. Meta-reinforcement learning is built for fast adaptation, and epidemics drift, so an adaptive policy looks like the right tool. Whether continuous re-inference of the latent regime actually improves decisions has not been tested against a design that can separate the two reasons it might not. We built a seasonal SEIRS-type environment with a reporting layer in which a seven-dimensional regime drifts while the agent observes only reported case counts, together with three families that distinguish whether a decision-relevant difference exists from whether it can be recovered without acting: fixed intervention parameters (passive), parameters affine in the passively visible transmission rate (linked), and parameters that act only through the intervention (action-revealed). A PEARL-style context encoder trained with PPO was compared with frozen inference, an oracle, and non-adaptive baselines across 25 seeds. Continuous re-inference gave no reliable advantage over an estimate frozen at the start of an episode in either family (+0.32 linked, +0.02 action-revealed, across-seed standard deviations 0.47 and 2.96). Knowing the true regime was worth +0.06 in the linked family against +2.03 in the action-revealed family, so the value of the regime concentrates where it cannot be read passively. A passive identifiability index, computed from intervention-free trajectories before any training, separates the families out of sample (R2 0.591 against 0.158) and serves as a design-time test: where the parameters that set the best action are already readable from passive surveillance, a fixed policy does as well. A gate signal built from the agent's own model carried information about harmful adaptation steps but did not beat a plain predictive-variance baseline.