14.8 An integrated testbed

The cleanest initial world is a partially observed, nonstationary tool-use environment with three withheld structures:

1. a latent regime that changes one causal interaction;

2. a tool macro or mediator interface required to probe or exploit that interaction; and

3. a decision schema that must condition action on the newly recoverable regime.

Ordinary policy optimization can improve within a regime but cannot express the missing state distinction. CLIC can diagnose the causal alternative but cannot execute the required compound probe. OPTIC can construct the probe but does not know when it has decision value. RELIC can detect persistent failure but cannot identify the missing mechanism. In the exact AGENTIC–0A ablation, the complete system succeeded on every challenge world, whereas the best pairwise composition reached at most one half. This is a result about that registered finite construction, not a claim that all three components are necessary in every environment.

A controlled progression begins with a finite exact world, then adds operator estimation, observation noise, imperfect execution, switching regimes, and finally language-mediated tool use. At every stage, discovery, admission, and confirmation seeds remain disjoint. A flat end-to-end learner receives the same interaction and computation budget.