SI: stability and attenuation
Stability and attenuation
Across ten same-tokenizer pairs, deterministic lower bounds of 0.42–0.67 accompany observed Kendall agreements of 0.59–0.83. The bounds use realised predictive perturbations.
What the stability bounds guarantee
Loading recorded comparison…
Methods and interpretation
Bounds require their stated common-outcome and denominator conditions. Inapplicable certificates and null boundaries are part of the result.
Procedure in the paper: Reconstruct the fixed common-outcome matrices, compute root-probability errors and certificate terms, retain applicability checks, and apply the stated resampling rule.
Code and data
v0.1.0 · 71 KB · View source on GitHub ↗
Ten-model output/mid/last convergence with corrected 5,000-draw context bootstrap, the complete eight-model 27-cell dense gauge sweep, and complete common-byte contraction, covariance, risk-floor and 1,000-subsample stability analyses.
v0.1.0 · 85 KB · View source on GitHub ↗
Native outcome perturbation certificates for 10 declared model pairs with 400,000 Monte Carlo entry comparisons; eight-model covariance accounting and 40 checkpoint risk-floor measurements, including 1,000 paired-substream tightness analyses and explicit vacuity.
Each standalone package includes code, shared helpers, required small inputs and reference results, with setup and commands in its README. Model weights and public datasets are obtained separately where needed.
Download complete source (v0.1.0) for all experiments and the companion website.