Replicate correlated-failure analysis across compressors

Replicate the correlated-failure analysis using each alternative compressor rather than relying primarily on PPMd-derived diversity measures.

Background

The paper uses PPMd as its primary compressor and reports a heavily subsampled diagnostic comparing PPMd with CTW, LZMA, and gzip. That diagnostic indicates broadly preserved model-pair ordering, but it does not repeat the held-out correlated-failure analysis under the alternative compressors. Consequently, the robustness of the observed association between compression-derived diversity and correlated failure across compressor choices remains unresolved.

References

The comparison is nevertheless descriptive and heavily subsampled. It does not repeat the held-out correlated-failure models under every compressor, which remains future work.

— Inferred Generative-Process Diversity Predicts Correlated Failure Across Language Models  (2609.03422 - Tieman et al., 3 Sep 2026) in Appendix, Section "Compression Distance," subsection "Compressor choice and robustness"

The sign reverses, so the quantity is telling us which codec we are in rather than why it behaves as it does, and the second condition is what says so.

— Lossy Compression of PDE Training Inputs: Field Reconstruction Error Does Not Order the Cost to a Trained Operator  (2610.06095 - Le, 5 Oct 2026) in Section 6, “Refuted explanations” (paragraph on codec error outside the band the operator reads)