Misinformation-Aware Evaluation for Social Agents: BeliefBench with Inspectable Belief Provenance

BeliefBench is a tick-based evaluation harness for misinformation-aware social agents. It provides schema-versioned JSONL traces, twenty diagnostic scenarios, and metrics for need satisfaction, trust drift, calibration, source sensitivity, belief decay, gossip, and spread. Bundled symbolic baselines and controlled ablations separate action-policy effects from belief-update mechanisms. The experiments evaluate instrumentation, policy discrimination, adapter participation, and reproducibility under fixed diagnostic probes. Software, scenarios, and regeneration scripts are available at https://doi.org/10.5281/zenodo.21877234.