Commit–Predict–Verify during agentic AI delegation: A randomized factorial trial of AI-free evidence reasoning

Agentic artificial intelligence systems can plan and execute multistep academic work, yet assisted output is an uncertain proxy for durable learning. In a four-site randomized 2 × 2 factorial trial, 600 Chinese undergraduates were assigned to single-request reactive advice or bounded multistep agentic delegation, crossed with an enforceable Commit–Predict–Verify (CPV) learner-control architecture or ungated access, using the same configured model, corpus, and session time across eight supervised sessions. The primary outcome was AI-free performance on a novel Evidence-Reasoning Transfer Assessment (ERTA) at T2, 28 ± 3 days after the immediate posttest (T1). In the primary intention-to-treat analysis with 50-fold multiple imputation, CPV increased adjusted T2 ERTA under agentic delegation by 10.6 percentage points (95% CI 6.2 to 15.0; p < .001; Hedges g = 0.60), corroborated by complete-case, per-protocol, weighted, and delta-adjusted sensitivity analyses. Across the factorial, the Holm-adjusted CPV main effect was 8.6 points (p < .001) and CPV reduced Brier calibration error (-0.034; Holm-adjusted p = .006); the agency main effect and the agency × CPV interaction were not significant. Prespecified advice-validity reliance outcomes were not estimable because trial-level records were unavailable in the analytic dataset. Enforceable learner-control friction improved delayed, AI-free evidence reasoning during agentic delegation.