Compression-Based Epistemic Stress Testing of Alignment Research

This work introduces compression-based epistemic stress testing, a lightweight methodological framework for evaluating the structural robustness of AI alignment research. Rather than benchmarking model performance, the method repurposes local language models as analytical instruments to test whether alignment arguments preserve causal integrity under epistemic stress. Three constrained tests are defined and applied: (1) constrained disagreement, (2) scale sensitivity analysis, and (3) causal compression. Using these tests, the paper demonstrates that well-formed alignment arguments retain mechanism-level coherence under disagreement, generalization, and folding, while weaker arguments collapse into moralization, rhetorical redundancy, or benchmark-dependent claims. The study includes a comparative negative-results appendix, a full methods and reproducibility addendum using LM Studio with local inference only, and an ASCII-clean schematic pipeline suitable for text-based ingestion (e.g., NotebookLM). Failure is treated as diagnostic signal rather than evaluator error. The method complements traditional peer review by making idea-level robustness observable, reproducible, and portable.