Compensatory Recovery as a Regime Indicator in Bounded-Context Agents

We introduce a trajectory-level regime classifier for bounded-context LLM agents operating in irreversible sequential environments. From the axioms of Constrained Generative Systems (CGS) theory, we derive a compensatory threshold condition and operationalize it through two observables: ρ_eff (local compensatory capacity) and ε_total, decomposed into passive (environmental) and active (agent-induced) components — a distinction absent from prior CGS formulations. The classifier — the sign of mean(ρ_eff − ε_total) over the second half of a trajectory — achieves 95.9% accuracy (FP=1) across 170 level-episodes from seven LLM architectures (Haiku 4.5, Sonnet 4.5, Opus 4.5, GPT-4.1, GPT-5.2, gemini-3.1-pro-preview, gemini-2.5-pro) in both Bridge-ON and Bridge-OFF conditions. Empirical analysis reveals that the Bridge constraint reinjection mechanism acts through two parallel pathways: increasing recovery probability (ρ_eff) and reducing agent-induced expansion (ε_active) in drift-susceptible architectures. The mechanism is defined independently of the test environment and is applicable to any system with bounded stateless context, irreversible state changes, and architecture-dependent compensatory capacity. Part of the CGS empirical series (O4). Related: C3 (theory), O3 (behavioral taxonomy), USPTO Provisional 63/993,764.