Accounting for quantization and dropout in compute measures

Determine how governance-relevant measures of training compute should account for algorithmic techniques such as quantization and dropout so that compute-based thresholds accurately reflect risk.

Background

Compute thresholds are increasingly used to scope governance obligations, but actual compute use depends on training techniques that compress or regularize models.

Clarifying how to incorporate quantization, dropout, and similar methods into compute accounting would improve the precision and defensibility of compute-based policies.