16.10 AGENTIC-SIM: the integrated benchmark

AGENTIC-SIM–0 hides a regime-dependent mechanism, the compound probe needed to distinguish it, and the decision schema that exploits the recovered regime. A base policy can assimilate within either regime but aliases them under partial observation. CLIC can express the competing mechanisms but lacks a decisive measurement; OPTIC receives primitive simulator calls but not the required macro; RELIC observes persistent decision failure but not the latent label.

The registered test asks whether the full loop constructs the probe, recovers the mechanism or calibrated equivalence class, and improves decisions on held-out transitions. The full \(2^3\) component factorial is supplemented by fixed-order and guarded-routing AGENTIC, an admission-record-free integration ablation, and separate mechanism, skill, and regime oracles. Results report mechanism coverage, abstention on unsupported alternatives, skill execution validity, decision value, intervention cost, transport, and regression.

Experiment: AGENTIC-SIM–0. World: partially observed simulator with a hidden switching mechanism and finite intervention budget.
Joint construction: causal distinction, compound measurement or intervention skill, and regime-conditioned experiment or control schema.
Admission: held-out interventional prediction, execution and safety tests, decision improvement, and transport to changed switching rates.
Controls: component factorial, fixed versus guarded order, admission-record ablation, flat learner, and separate oracle ceilings.
Boundary: this is a benchmark specification. The three Prometheus substrates establish SGTE–0-style groundwork, not a completed AGENTIC-SIM result.

. A simulator can falsify claims about its own executable world. It cannot by itself certify that the simulated mechanisms are true of nature.