More Context Is Not Validation: A Baseline-First Audit of Action-Grounded LLM Listener Agents

LLM agents can generate individualized audience reactions before those reactions are behaviorally valid. We present a baseline-first audit of listener simulation for audio storytelling, separating three questions: whether a representation changes an LLM's predictions, whether the predictions beat a leakage-safe calibrated alternative, and whether the behavioral label measures the intended construct. First, on 1,199 listener-show tasks across four models, evidence-cited action memory reduces Brier error relative to summary personas for every model, but reliable ROC AUC gains appear only for Sonnet (+0.072, 95% bootstrap interval 0.044–0.101) and Opus (+0.058, 0.035–0.082). Next, on a separate 306-task cohort, a five-fold listener-grouped calibrated prior achieves Brier 0.216, compared with 0.235 for a grounded LLM hybrid. Finally, on 150 warm-user tasks, adding historical content digests changes Brier by +0.007 relative to a candidate-digest control and worsens error by +0.0129 (0.0018–0.0246) relative to a raw-candidate arm. A raw-data audit further shows that episode-position events mix listening, autoplay exposure, skipping, and missing playtime telemetry. More context can improve a representation, but does not by itself establish predictive or construct validity.