Constraint-Guided Coherence in LLM Output: Replication Dataset (GPT-4o Extended Run)

This replication extends the initial four-model CPA–coherence study to an extended GPT-4odataset comprising 90 runs (30 per constraint level). Using identical prompts and fixed parame-ters, we confirm that increasing informational constraint systematically reduces both the meansurprisal and mean token entropy, replicating the predicted coherence-under-constraint signa-ture with high statistical confidence. The experiment additionally reveals a strong correlationbetween mean surprisal and mean entropy (r = 0.957, p ≪ 0.001), consistent with dual process-ing regimes (”fast/exploratory” vs. ”slow/convergent”). Finally, the aggregate surprisal curvefits the Vogel–Fulcher–Tammann (VFT) equation with ϵ < 10−5, mirroring glass-freeze behaviorobserved in physical systems. An independent code and data audit confirmed the reproducibilityof these findings, validated the VFT fit with an RMSE of 0.02085, and verified data integrityacross all 240,313 token observations. All codes, data, and figures are provided for replicationand review. Failure to reproduce this constraint-linked uncertainty reduction would falsify theCPA signature in this domain.