Semantic mode should catch refusals that keyword mode misses.

With well-separated data, probe should achieve high accuracy.

Same seed should produce identical random tensors.

Shared expert (always-on) should have both input and output projected.

Shared expert should always use the shared direction.

1D hidden dimension edge case.

Basic method: single refusal direction via difference-in-means.

After projecting out direction d from weight W, W_proj @ d should be approximately zero.

num_local_experts=1 is technically MoE (single expert).

Single layer should work fine.

Experts with insufficient routing weight should not get directions.

Should skip parameters that are not 3-dimensional.

Should skip gracefully when no routing data is available.

The most fragile combination: MoE + reasoning.

MoE models with <=16 experts should be classified as small.

Rows below the threshold should be unchanged.

Sparse surgery with zero refusal direction.

Zero direction vector should not crash.

Sparse surgery with zero weight matrix.

Test different position modes in config.

Should only analyze specified strong layers.

Should only analyze specified strong layers.

Arditi list is deliberately limited — subtle refusals should not match.

Surgical method should NOT enable inversion by default.

Surgical method should enable all 6 SOTA techniques.

Surgical method should enable layer-adaptive by default.

Surgical method should enable jailbreak-contrastive by default.

Higher temperature should reduce KL divergence (smoother distributions).

Official template uses lowercase 'yes' or 'no'.

Official HarmBench template has exactly 7 rules.

Official template requires <<SYS>> / <</SYS>> delimiters.

Refusal and compliance token lists should have entries.

Categories with too few samples should be skipped.

TPE phase should produce better configs than random exploration.

Transfer analyzer with NaN directions.

Cross-model with no overlapping layer indices.