Stabilize large-model training by restricting weight updates to curated manifolds that align with desired behaviors and safety envelopes.
| Diagnostic | Purpose | Signal Interpretation |
|---|---|---|
| Manifold distance | Measures how far weights drift from constraints | Rising distance indicates projection frequency too low |
| Gradient rejection rate | Percentage of gradient components removed during projection | High rates suggest constraint mismatch with task |
| Loss landscape curvature | Evaluate smoothness post-projection | Smoother curvature indicates improved stability |
| Safety vector overlap | Dot product between weights and known risky directions | Near-zero overlap shows policy-safe manifolds working |
Visualize metrics over time to catch degradation early.