manifold-guard spots DPO geometry collapse

This is a smart observability idea: instead of waiting for evals or loss curves to tell you something is wrong, it tries to expose alignment-induced damage directly in optimizer state. The claim is intriguing, but it is still early evidence from a narrow setup, so the real test is whether the signal survives across models, datasets, and alignment recipes.

• –The big upside is cost: it piggybacks on Adam state, so it can run during training with essentially no extra forward passes.
• –If the result replicates, HM/AM could become a useful early-warning metric for post-training anisotropy or sparse-curvature collapse.
• –The experiment is still limited: one model size, one family of data, and a metric that could be sensitive to optimizer choice or training hyperparameters.
• –The most interesting open question is causal, not descriptive: is DPO uniquely “damaging” geometry, or is the probe just surfacing the expected sparsity of preference gradients?
• –The repo is packaged well for reuse, which makes it more useful than a one-off paper graph.