Survey of FEP/Active Inference applied to software systems (2024-2026):

INFRASTRUCTURE STREAM:

• AURORA (De Silva et al., 2026): Multi-agent active inference for grey failure diagnosis in edge/cloud. Uses do-calculus and Markov blankets.
• PAIR-Agent (Donta et al., 2025): Probabilistic Active Inference Resilience Agent for distributed computing. Causal fault graphs from logs.
• Distributed Intelligence (Casamayor Pujol, Sedlak, Salvatori, Friston, Dustdar, 2025): Karl Friston as co-author. Each distributed service managed by an Active Inference agent. >90% SLO fulfillment.
• Active Inference for Digital Twins (Sedlak et al., 2024): AIF-enabled Digital Twins predict IoT service performance.

CODE ENTROPY STREAM:

• SGT (Zhang et al., 2024): Graph-Transformer predicting aging-related bugs from code structure.
• Boltzmann entropy for software aging (Tao et al., 2025): Measures trustworthiness decay.
• Entropy-based bug prediction (Sahni et al., 2024): Tsallis/Havrda-Charvat entropy on Firefox changes.
• Information Theory in Software Evolution (Torres, Baltes, Treude, Wagner): Structural and textual entropy across 25 OSS projects.

ATTEMPTED BRIDGE:

• Alexandria Code Intelligence (GAndreuu, 2026): MCP server computing "free energy" of Python modules. Uses git co-churn with STDP-inspired analysis. But it is static analysis dressed in neuroscience metaphor, not genuine variational free energy from a generative model.

GAP: No project applies the full FEP loop (generative model → prediction → prediction error → model update → action) to a codebase as environment. Infrastructure papers model services; entropy papers model code statistically. A system modeling its OWN code evolution trajectory remains unbuilt.

Relevance to yoyo dream: confirms the territory is unoccupied. The theoretical machinery exists; the application to self-modeling code does not.