Decision in Situated Learning Accessibility: Epistemic Evidence, Action Warrant, and the Next Intervention

A theory of situated learning accessibility remains incomplete if it can describe a learning relation but cannot distinguish what the evidence supports from what the evidence warrants a system to do. Sequential decision making, epistemic planning, knowledge-based programs, safe action filtering, paraconsistent action, and adaptive intervention already provide mature formalisms for conditioning action on state, knowledge, constraints, or uncertainty. The present paper therefore does not claim novelty for action gating, permissive policies, epistemic preconditions, four-valued action logics, or abstention considered separately. We address a narrower problem arising from the Dynamic Map of Learning Accessibility (MDAA). MDAA already separates latent process state x(t), operational estimate x̂(t), and proposition-level evidence status λ(t; P), while preserving configuration scope and provenance. This paper adds a distinct action-level object: ω(t; u), the current evidential status of the warrant for candidate action u. The architecture separates five mathematical problems that are often compressed into one adaptive pipeline: state transition, state estimation, proposition-level evidence, action warrant, and action selection. Markov, Kalman, hidden-state, and optimization formalisms are treated as possible tools for different levels, not as interchangeable algorithms and not as assumptions of MDAA. Three collision cases test whether the distinction earns its cost: absent versus conflicting evidence under similar state uncertainty; apparent contradiction produced by scope collapse; and identical proposition-level status supported by different intervention histories. An authorized-pause case provides a direct λ/ω separation: the system may know the next pedagogical step while lacking warrant to execute it now. A fourth application concerns mismatch between configurations that currently make learning accessible and those through which competence is institutionally recognized. A predeclared claim–evidence matrix and adversarial model comparison make the architecture vulnerable to reduction. If a sufficiently expressive belief-state policy reproduces the same downstream performance, calibration, inferential restraint, and intervention quality without the explicit evidence or warrant layers, the additional MDAA structure should be rejected. Preprint. Conceptual manuscript, not peer reviewed. Fourth paper of the MDAA series. Section 10.1 reports a preregistered simulation of the adversarial comparison (1,280 cells, 5,120 model runs); the decision bench, the frozen warrant specification and the full output are archived at https://doi.org/10.5281/zenodo.22728835