The Self-Undermining Assurance Problem: simulation notebook, results and reproducibility materials

Abstract

Computational companion to The Self-Undermining Assurance Problem: Preserving Certifiability in Human–AI Knowledge Work. This release contains an executed simulation notebook, recorded dependencies, run records, tables, sensitivity and budget sweeps, analytical checks, and framework/figure generation scripts. The public HaluEval general_data split is downloaded and hash-verified by the notebook; raw upstream data are not redistributed. Main simulations use 200 replications; sweeps use 40. Evidence is conditional on an assigned working model with simulated reviewers. No human-participant evidence or certified viability guarantee is claimed. V is a preservation-oriented reserve heuristic; the audit bound is 80% per block and budgets are matched offered ceilings. Known implementation limitation: the adaptation branch buys audits while review effort remains even when the audited point estimate exceeds the ceiling. This release preserves the executed implementation and results, documents that limitation, and does not claim it implements the proposed distinction between insufficient precision and high observed risk. Fixing the controller requires rerunning affected results. README.txt and ERRATA_AND_REVIEW.txt explain the limitations and remaining manuscript corrections.