Observation-control mismatch in host-microbiome systems: When does microbiome compression preserve intervention choice?

The intestinal microbiome is a high-dimensional ecological system, but neither host physiology nor clinical measurement has access to all of that detail. Functional redundancy, ecological filtering and niche construction explain why taxonomically different communities can converge on similar functions or host phenotypes, yet they do not answer a separate decision problem: when does a predeclared microbiome representation discard distinctions that matter for intervention choice? Here I define observation-control mismatch (OCM) as a task-specific decision audit for host-microbiome systems. An action is acceptable only when its expected loss is both within a predeclared regret margin η of the best available action and below an absolute biological or clinical loss threshold τ, with any critical-harm constraints specified separately. For a finite set of experimentally constructed realizations, intervention-panel feasibility is assessed first: each tested realization must have at least one acceptable action. Conditional on that requirement, a compressed observational class is control-sufficient when the realizations share at least one acceptable action and shows OCM when each realization is individually actionable but no common acceptable action exists; a realization with no acceptable action instead indicates intervention-panel insufficiency. For population data, actionability coverage and common-action coverage are audited separately against a prespecified coverage q. All empirical conclusions are therefore conditional on the chosen X: control sufficiency means that no intervention-relevant distinction was detected beyond OE within the information represented by X, not that the underlying state S has been exhaustively resolved. This audit is complementary to heterogeneous-treatment-effect analysis and individualized treatment-rule learning: it asks whether information deliberately discarded by an existing biological or clinical representation changes what can be done, not merely whether responses differ. The intestinal boundary provides the biological motivation because microbial ecological states are projected into host-accessible consequences, whereas host physiology and interventions modify niche variables that select among microbial states. This projection-selection architecture is biological rather than a mathematical duality. OCM is not itself a universal falsifiable biological proposition; it generates task-bounded hypotheses defined by a population, representation, intervention panel, outcome, time horizon, η, τ and, for population analyses, q. Retrospective multi-intervention data can screen for mismatch, whereas prospective defined-community or gnotobiotic experiments can test finite-support intersections directly. The audit is unnecessary when a direct causal determinant or an already validated treatment rule makes within-class auditing irrelevant. Additional dynamic or host-boundary measurements are justified only when they repair a demonstrated decision blind spot.