audits.probe_audit¶
Auditable-KB probing. For one frozen setattn+role checkpoint: * roll out S episodes (different maps) with the TRUE policy, * at each step record z (full comms) AND z_nocomm (comm graph emptied -> GNN adds nothing), plus GROUND-TRUTH targets computed from god-view state, * fit LINEAR probes z->target, held-out BY EPISODE (train on some maps, score on unseen maps), for z and for z_nocomm, plus a shuffled-label baseline, * emit a compact ledger JSON (held-out R²/accuracy per property, z vs z_nocomm) and a demo npz (one held-out episode: per-step true vs predicted, + belief maps for the fog view). No training; CPU. Nothing but small results leaves the machine.