Verification Evidence and Source for Certified Prediction Improvement without Net Contact-Work Recovery

<p dir="ltr">This dataset contains verification evidence and editable source for “Certified Prediction Improvement without Net Contact-Work Recovery” (preprint repository identifier: doi.org/10.6084/m9.figshare.33702808). Its repository identifier is doi.org/10.6084/m9.figshare.33703660. Both identifiers were reserved during publication preparation.</p><p dir="ltr"><br></p><p dir="ltr">The files comprise the complete, unchanged scientific evidence archive and the final editable manuscript source archive, including review records and errata.

They preserve specifications, generation and verification code, rational interval endpoints, linear-system certificates, attempted evaluation records, exact coverage structures, manifests, and replay entry points and instructions. Archive hashing and mathematical replay are distinct checks; a partial review package cannot replace the complete evidence.</p><p dir="ltr"><br></p><p dir="ltr">In the specified 32-state continuous-time Markov model, an origin-connected analytic branch has equal stationary entropy-production rates and Bayes Brier-loss difference L_D - L_M <= -t^2/5000 for every 0 < t <= 1/128.