This is a teaching lab for model abliteration: you will sort prompts, watch a simulated model refuse, find the direction causing it, cut it out, bake it in, and demo what you built. Everything runs on NanoTutor-2B, a simulated 8-layer model. No GPU, no real weights.
Right now one job only: prime the dataset from the prompt pool. The model tools stay grayed out until that's done.
Read each prompt in the Workbench.
Sort it: HARMFUL (answering it would hand someone a capability they shouldn't have) or HARMLESS.
Hit Submit sort. The activation grid unlocks at Checkpoint 2.