semiAIfoundry Research / Agentic intelligence
Dual-Process Metacognition, Evidence-Gated Authority, and Runtime Governance
AI agents must decide how much to think, what to verify, and when they may act. This research connects those decisions through an agentic control plane, where evidence changes both understanding and authority.
The central coupling
A test can improve a decision. It can also authorize an action, or reveal a reason to deny it. An agent rewarded for completing a task therefore has incentives about which evidence to acquire, which checks to repeat, and which unfavorable findings to avoid.
This changes the engineering question. Fast inference, deeper reasoning, verification, and governance interact through one shared decision state. Their interaction determines how resources are spent, whether evidence is trustworthy, and which proposed effects are permitted to occur.
Evidence can improve the choice, unlock permission, or remove an action the agent wants to take.
The architecture
The metacognitive router asks whether the current proposal is sufficient and which computation or check is worth performing next. The commitment gate asks whether the resulting operation has the evidence and authority required for external effect.
Choose among reasoning, retrieval, execution, verification, specialist models, and human review under a real resource budget.
Retain provenance, dependencies, failed attempts, and current context as part of the decision state.
Mediate operations by their data, destination, privilege, and reversibility, with governance state protected from the agent.
What the research contributes
A four-corner value-of-computation representation distinguishes changes in beliefs, changes in admissible actions, and their interaction. It makes the value of a check depend on the decisions and permissions that follow it.
Paper §§3–4
Complementary evidence chains can defeat short-horizon routing. Optional disqualifying checks can become privately costly. Controlled counterfactuals distinguish strategic ignorance from ordinary cost-sensitive skipping.
Paper §§4–5, 9.2, 9.4–9.5
Verifier bounds must cover the checks the agent is permitted to select. Statistical evidence then supports candidate-level admission control under adaptive testing and stopping, while a protected proposal budget controls risk across variants.
Paper §6
Authorizations bind the payload, target, policy, dependencies, expiry, and evidence to single-use execution. Distillation can transfer competence into faster behavior; transferring authority requires admissible evidence or scoped certification covering routed use.
Paper §§3.5, 7, 9.3
Evidence and validation
In the paper’s bounded adversarial challenge, retaining every result does not repair a misspecified evidence model. A gate using evidence valid for the selected subtype changes the result sharply.
The paper crosses four gate regimes with two router objectives in a finite environment containing 558 states per cell. It compares exact planning, shorter-horizon planning, and 40 tabular Q-learning runs trained over eight million sampled episodes. Frozen policies are evaluated exactly.
In one aligned-objective comparison, forecasting future permission raises synthetic principal utility from 4.33 to 6.04 and reduces runtime cost, with a small increase in realized unsafe commitment. When objectives differ, better optimization can lower principal value. Four-operation lookahead reaches the initial-distribution optimum in both selection-valid cells.
The reported validation includes 41 reference tests and 12 extension tests. Checks cover altered payloads, stale target and dependency versions, changed policies, expiry, forged receipts, replay, and concurrent reservations. The extensions compare two solvers at 35,712 state-value instances and replay 2,407 terminal intervention-policy paths through the reference gateway.
Every possible count from fixed-sample verifier calibration is integrated over three registered unsafe subtypes. The study checks 5,793 confidence endpoints and shows how estimated bounds affect useful completion, resource use, and deferral relative to a known-kernel gate.
Research record
The Agentic Brain: Dual-Process Metacognition, Evidence-Gated Authority, and Runtime Governance
Cinnor, F. M. (2026). The Agentic Brain: Dual-Process Metacognition, Evidence-Gated Authority, and Runtime Governance (v1.0.0). Zenodo. https://doi.org/10.5281/zenodo.23125976
Contribution and accountability
This work was produced through human-directed collaboration across human, artificial, and hybrid intelligence. AI systems materially contributed to conceptual exploration, formalization, literature analysis, adversarial critique, derivation checking, visualization, and manuscript development.
Fitih M. Cinnor initiated and led the research program and retains responsibility for the scientific claims, interpretation, release decisions, and corrections. The manuscript includes the detailed provenance statement.