{
"$type": "site.standard.document",
"bskyPostRef": {
"cid": "bafyreigkwfze4oxkce3bdtulf6vg42kfmqxw76kqqd3npvfwxakmmmvgsy",
"uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mnurjhgffpu2"
},
"path": "/t/update-falsifying-the-gd-ratio-how-logit-temperature-confounds-dynamic-geometry-v22-5-audit/176646#post_1",
"publishedAt": "2026-06-09T16:04:53.000Z",
"site": "https://discuss.huggingface.co",
"textContent": "“Following excellent feedback from the community, I ran a temperature-normalization audit on the GD_ratio clusters.\nResult: The clusters reorganized significantly after entropy calibration. GPT-2 and Phi-1.5 switched groups.\nConclusion: Raw logit geometry is heavily confounded by effective model temperature (calibration). It is not a pure dynamic signal.\nImpact: 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.\nFull details in V22.5 report below. Next step: Isolating the residual dynamic structure post-calibration.”",
"title": "[Update] Falsifying the GD_ratio: How Logit Temperature Confounds Dynamic Geometry (V22.5 Audit)"
}