Intent-Governed Loops for Accountable Agentic AI (Poster)

A companion to the paper titled "Intent-Governed Loops for Accountable Agentic AI" and distills the core problem framing: in safety-critical and high-liability settings, it is not enough for an agent to be “aligned” in the behavioral sense—we need to be able to prove, at runtime, that each proposed action is authorized, in-scope, and still valid under current evidence. It introduces the “admissibility gap” and motivates a paradigm shift from alignment-as-tendency to admissibility-as-per-action authorization, especially where agents operate with tool access and institutional authority. The poster summarizes the intent-governed loop architecture: an externally-authored, revocable Intent object (human context, symbolic constraints, semantic guidance, plus expiry and named escalation authority), a Planner that proposes high-level actions with structured justification, a dual-mode Enforcer that performs deterministic constraint checks followed by semantic boundary evaluation, and a temporal governance graph that records provenance and automatically invalidates actions when upstream evidence changes. It presents four runtime invariants (no orphan action, no stale execution, mandatory escalation, no silent override), seven analyzed failure modes, and three adversarial stress tests that expose realistic breakdown paths (e.g., temporal arbitrage, compliance theater, escalation fatigue). Finally, it outlines a concrete evaluation agenda—synthetic benchmark scenarios and loop-level metrics (UAR, TCS, EI, ART, SOR, CBH)—to enable empirical validation of accountable agentic control in live institutional environments. Paper link: https://doi.org/10.5281/zenodo.18252972