“Following excellent feedback from the community, I ran a temperature-normalization audit on the GD_ratio clusters.
Result: The clusters reorganized significantly after entropy calibration. GPT-2 and Phi-1.5 switched groups.
Conclusion: Raw logit geometry is heavily confounded by effective model temperature (calibration). It is not a pure dynamic signal.
Impact: This falsifies the strongest interpretation of GD_ratio but strengthens the framework by identifying a critical control variable. Future analyses will use temperature-normalized metrics.
Full details in V22.5 report below. Next step: Isolating the residual dynamic structure post-calibration.”