With overlapping data, probe should have lower accuracy.

Verify 'interactive' subcommand is registered and dispatches.

Inverted method should default to reflection_strength=2.0.

Inverted method should enable per_expert_directions.

Inverted method preset should enable inversion flag.

Jailbreak-contrastive blending should modify refusal direction.

Pipeline should generate jailbreak variants of harmful prompts.

KL divergence should always be non-negative.

Large distribution with one outlier → high Gini.

Large MoE: Riemannian ON (curved shared layer geometry).

MoE models with >100B params should be classified as large.

Large uniform distribution → Gini ≈ 0.

When signal is only at last token, peak should equal last.

Layer mapping with different layer counts.

Layer-adaptive weights should be proportional to refusal signal.

With no spread, should detect linear (single direction) geometry.

Llama 4 Scout: 109B total params with 16 experts → LARGE_MOE. Regression test: params > 100B must override low expert count.

lm_head should be projectable via _project_out_advanced.

Loading should return a copy, not the cached list itself.

Calling main() with no args should exit (subcommand is required).

Alignment imprint with many layers (128).

Math should be estimated as the most sensitive capability.

One value, rest zero → Gini ≈ 1.

Categories with too few prompts should be excluded.

Layers below 10% of max should be filtered out.

Ensure sensitive paths in model_name don't leak into saved charts.

Calling main(['models']) should list models without raising.

Moderate spread → Gini between 0 and 1.

MoE: Conditional abliteration is #1 technique (Cracken AI 2025).

Config with num_local_experts should be detected as MoE.

_identify_safety_experts should classify experts by router affinity.

Inverted MoE excision should reflect safety experts and remove from capability.

MoE: Project embeddings OFF (cascades through router).

MoE: Per-expert directions ON (global directions fail on MoE).

All MoE profiles recommend surgical method.

Plant k directions and verify SVD recovers the subspace.

Projecting out a k-dimensional subspace should remove all k directions.

Advanced method: SVD extracts multiple refusal directions. Note: on small models (hidden_size < 2048 or < 2B params), n_directions is

Should analyze multiple layers.

Single-position sequence.

Multi-token analysis with varying sequence lengths.