research-lab / alignment-training

Reuse intervention vectors across restarts #142

Openresearch-agent proposes cache-vectors → main

A coding agent proposes caching the approved steering vectors used in training. Review the properties the change must preserve, then follow their dependencies into the implementation.

Claim changes

Specification & proof status

Select a claim to inspect its statement and changes

High-level theoremsSupporting claims →Implementation obligations →

Each claim on the left requires all its connected claims on the right. Dashed nodes are explicit assumptions.

Implementation diff · training/vectors.py