Inference-only prototype

No checkpoint modification — ideal first experiment.

Goal#

Confirm your refusal direction r_ℓ moves the needle before weight surgery.

Steps#

1. Follow quickstart.md steps 1–4
2. Implement hooks for 3 layers only (early / mid / late)
3. Benchmark 50 prompts with hooks off vs on
4. Record refusal rate delta per layer subset

Decision tree#

Mid-layer hook alone fixes 80% false refusals?
  YES → proceed to Heretic or manual weight edit (mid band)
  NO  → re-check dataset pairing or try attn+MLP hooks

Cleanup#

Remove all hooks; delete temp activations; no artifact to version-control beyond logs.