Propose, Judge, Commit: Ternary Verdicts with External State Authority for LLM Agents

LLM agents do not distinguish generated text from accepted state: a plausible but ungrounded utterance can enter the transcript and subsequently influence tool calls, memory writes, and final answers.

We propose an architecture of strict authority separation in which a language model proposes candidate moves, an adapter normalises them into typed semantic objects, an external judge evaluates them under an explicit verification contract, and a state manager commits only certified transitions. Accepted state is represented as a versioned reasoning graph rather than a transcript.

The judge uses four verdicts—FOLLOW, OMIT, NULL, and UNDECIDED—corresponding to commit, silence, exclusion, and evidence-seeking actions. The semantics are based on Brusentsov's ternary logic of consequence, where inessentiality is distinct from both assertion and exclusion. In particular, OMIT nodes are removed from the active projection while retained in the graph as potential growth points, whereas NULL severs an incompatible branch.

We formalise certified transitions and establish constructive properties including state groundedness, judge totality, and auditability, and introduce re-verification with rollback. A preliminary live MVP on 150 questions and 212 atomic claims reduces FalseFollowRate from 0.36 to 0.035 at 0.81 coverage. Ablation indicates that judge competence is the primary factor, while the certificate contract and fixed interpretation rule substantially strengthen the effect and mitigate same-family self-judging.