The Recursive Causal Synthesis Agent

Long-lived intelligent systems face a structural dilemma: to remain competent under nonstationary tasks they must grow, reuse, reorganize, and sometimes compress internal structure, while also controlling resource costs and preventing unsafe self-modification. Many modern systems expose partial structural degrees of freedom (e.g., sparse expert routing, tool use, pruning), but structural change is typically governed by brittle heuristics and lacks audit-grade traces.This paper proposes the Recursive Causal Synthesis Agent (RCSA) as an operationalspecification for structural self-management. RCSAgoverns a high-dimensional structural state (experts, tools, memory, knowledge graphs) through (i) a low-dimensional, interpretable Cognitive Debt interface, (ii) a thermodynamically motivated structural governor that trades competence against cost under resource pressure, and (iii) an immutable Safety & Epistemic invariant suite that acts as a non-bypassable gateway on self-modification. We validate prerequisite mechanisms using minimal, reproducible simulations (a non-stationarytask carousel, structural triage, and a concrete self-blinding alignment micro-benchmark). We then extend the blueprint into a Stage III operational spec: counterfactual sandboxing with historical replay, dynamic (pressure-adjusted) veto thresholds under energy scarcity, and a deactivation escrow protocol that prevents silent regressions from irreversible forgetting. Finally, we provide an auditable per-step log schema (step_log_schema.json v2.0.0) that (a) enforces structural action integrity and veto semantics, (b) records the controller’s considered alternatives and weight snapshots, and (c) adds Golden Logic Ledger (GLL) and Shadow-GLL probes to detect probe gaming. All reported metrics and safety claims are derivable from logs alone.