Create two tiny 'models' with controllable logit distributions.

Sweep regularization from 0.0→1.0 and measure refusal rate + perplexity. Produces a dose-response curve: the fundamental plot for abliteration re

Whether *bfloat16* is natively supported on the target device.

BitsAndBytes requires NVIDIA CUDA — check that.

Accelerate's device_map='auto' is only reliable on CUDA.

Whether *float64* is supported (MPS does NOT support it).

Should handle 2D direction input (unsqueezed).

Should handle (1, seq_len, hidden_dim) inputs.

Test that the abliterate subcommand parses method correctly.

Steering hooks should subtract refusal direction from hidden states.

Verify the aggregate command parses without error.

Calling main(['aggregate']) with nonexistent dir should handle gracefully.

Activations aligned with direction should have high projection.

Single layer should still return a result.

When all norms are equal, should select all (or most).

If all cluster layers are entangled, don't skip all of them.

Case-insensitive matching should work for all-caps names.

Every key in the allowlist should pass through.

Only allowlisted analysis keys should pass through.

Angular drift should be monotonically non-decreasing.

Non-isotropic BBP extension should increase the threshold.

Sparse projection should reduce refusal signal.

Should auto-detect a router with unusual name via heuristic.

Should produce a LogitLensResult with expected fields.

Basic method should not enable SOTA techniques (no keys or False).

Whitened SVD should extract directions from activation differences.

Basic method should not enable per_expert_directions.

2.5x reflection should produce stronger negation than 2x.

Arditi/Zou list has exactly 29 substrings.