Abstract
Low-resource text feedback is a closed-loop decision problem because each generated action changes the next observable text. This paper proposes Hybrid Fuzzy-Bayesian Dynamic Feedback Generation (HFB-DFG), a modeling framework for partially observed stochastic dynamic feedback systems. A typed evidence graph constructs the observation, differentiable fuzzy reasoning represents partial semantic-state membership, and an amortized Dirichlet layer combines fuzzy, neural, and low-resource pseudo-count evidence. The resulting information state drives a risk-constrained controller that selects correction, explanation, hint, clarification, or review, after which a factorized stochastic transition generates a revision and reconstructs the next state. Controlled synthetic studies evaluate finite-horizon utility, transition risk, calibration, trajectory saturation, and an analytical batched-computation model. Under the stated generator and scoring assumptions, fuzzy reasoning improves ambiguous-state separation, Dirichlet calibration reduces modeled overconfidence, and risk-aware control shifts uncertain cases toward clarification or review. The numerical results are not corpus, human-subject, deployment, or measured-hardware evidence. A boundedness result establishes finite diagnostic uncertainty and discounted return, whereas the trajectory curves demonstrate only finite-horizon saturation. HFB-DFG therefore contributes an auditable Fuzzy-Bayesian control formulation and a reproducible evaluation specification; external validity remains to be established on traceable data with independent human and system measurements.