Identify the arithmetic source of the unsafe-context count shift

Identify the specific arithmetic or runtime setting responsible for the shift in the number of unsafe contexts observed for the GPT-2 weight-edit cell under matched-memory cache comparisons.

Background

When comparing K8/V4 and K4/V8 cache formats under a common recorded CPU runtime, the reported unsafe-context counts differ from an earlier K8/V4 count obtained under a different runtime. Restricting MKL to AVX2 reduces some feature-code differences, but the paper does not establish which arithmetic choice accounts for the full-domain change.

Pinpointing this source is necessary for interpreting the apparent ordering of cache formats and for determining whether the certified edit guarantees are reproducible across numerical implementations.

References

The matched ordering is consistent with Table~3; robustness across backends is untested. The net shift of four is not a statistical uncertainty interval.

— Behavioral Capacity Certificates for Quantized Language Models  (2609.37887 - Eamaz et al., 29 Sep 2026) in Appendix, Section 11.11, "Do weight-edit guarantees survive cache compression?" (label core:cacheedits)