Skip to content

audits

audits

KB-audit harness: linear-probe auditing of the shared-belief embedding.

probe_audit.py rolls out a frozen checkpoint, records the KB embedding with full comms vs. with the comm-graph emptied, fits linear probes to god-view ground-truth properties held out by map, and emits a per-property held-out R²/accuracy ledger. Answers: does the shared belief actually encode a property, and is it comms that puts it there? Results collection lives in audits/probe_ws/ (open/rooms x 16/24/32).