Taxonomy of State Transition Checks by Evidentiary Strength
State transition checks within a stateful agent harness differ in their evidentiary strength and degree of independent verification:
- Command and Predicate Checks: Evaluated directly by the host runtime (e.g., file existence, automated test suites, or execution exit codes), providing independently reproducible verification.
- Manual Checks: Require an explicit operational decision or approval from a human user.
- Checklist and Message Checks: Require the autonomous agent to acknowledge specific procedural criteria; these constitute structured self-attestations rather than objective external proof.
- LLM Review Checks: Introduce a secondary model call to assess semantic criteria, offering qualitative judgment without deterministic or formal guarantees.
This separation ensures that an agent's self-declared completion is not conflated with independent, reproducible verification.
0
1
Tags
Prep Sessions
Long-Horizon Agent Reliability: Stateful Scaffolding and Runtime Verification @ University of Michigan - Ann Arbor
Ch.2 Runtime Architecture and Boundary Control - Long-Horizon Agent Reliability: Stateful Scaffolding and Runtime Verification @ University of Michigan - Ann Arbor
Context Boundaries, Contract Boundaries, and Transition Checks - Long-Horizon Agent Reliability: Stateful Scaffolding and Runtime Verification @ University of Michigan - Ann Arbor
Learn After
Which category of state transition check provides independently reproducible verification evaluated directly by the host runtime?
Checklist and message checks provide independently reproducible verification because the agent must confirm procedural criteria before state transitions can occur.
According to the taxonomy of state transition checks, what kind of judgment do LLM review checks offer, and what guarantees do they lack?