"author:quantomania-rudolph" matching MCP servers:

Enables an LLM overseer or script to inspect, localize, edit, and judge a PyTorch model's weights in place without training, exposing sliced weight reads, activation and edge patching, circuit discovery, and a range of head/MLP/matrix editing ops. Free screening plus a charged propose-and-verdict loop, backed by a hash-chained ledger and optional mid-training hooks, keeps edits gated by automated acceptance criteria.