Component · spec & evidence

Belief-State Reward Bundle Replay

Checks that each step reward in a recorded run cites a declared verifier-feedback row, not a trick.

Contract

Does Takes a recorded, synthetic bundle of agent steps on three partially-observable toy tasks (a terminal investigation, a mock purchase, a small planner) and checks that every "the agent did the right step" reward is actually backed by a checkable verifier or observed feedback reference, not by hidden reasoning, formatting tricks, a smuggled answer key, or a final-answer-only score. It does not run or watch a live agent; it validates pre-recorded files. The resulting result record files show, per step, the belief summary, the reward, and whether the reward-hacking and replay checks passed, so it is inspectable why each reward was or was not allowed.

Scope limit It only checks that the projection's accounting lines up under its own schema rules over recorded synthetic fixtures; it excludes hidden-reasoning export, RL training, hidden gold or neural-judge-only labels, benchmark-performance claims, external model access, source-file changes, or launch, and proves nothing about real-world reward, live agent behavior, or domain-level conclusions.

Run

PYTHONPATH=src python3 -m microcosm_core.organs.belief_state_process_reward_replay run --input fixtures/first_wave/belief_state_process_reward_replay/input --out .microcosm/belief_state_process_reward_replay

Evidence

ClassComputed projectionevidence 3/5Source-faithful refactor

ai-safetyagent-evaluationred-teaming

Source

Source Design note · Source

Map Show on the map

Index Open compact index card

Paper module