Probing with very few samples.

Probing classifier across different hidden dims (min 2 for SGD).

_project_out_advanced should project ALL matching weights, not just first.

Should handle modules without bias gracefully.

Should return 0 when no candidate names match.

Bias projection should remove refusal direction component from bias.

Projection should NOT change components orthogonal to d.

Top promoted should have higher logit boost than top suppressed.

512 prompts across 7 severity tiers.

Slicing at standard volumes gives correct counts.

q_proj, k_proj, v_proj should all be projected alongside o_proj.

R1 distills are dense (distilled from MoE into dense).

Random directions should project much less on harmful activations than the true refusal direction.

Random matrix should have high effective rank.

CKA of random representations should be low.

Rank-1 matrix should have effective rank close to 1.

Reasoning Dense: Jailbreak contrast ON for thinking-chain refusal.

Dense reasoning models need more directions (>=12) to span refusal.

Dense reasoning models need more refinement passes (>=4).

Reasoning MoE: transplant_blend very low (preserve reasoning).

Metadata should include all new configuration parameters.

Rebirth should save model and comprehensive metadata JSON.

2x projection (reflection) should negate the refusal component.

Reflection should not change the component perpendicular to d.

reflection_strength should be explicitly overridable.

SAE should identify features that differ between harmful/harmless.

With sparse refusal, RSI should be high.

Regularization should preserve some of the refusal component.

Should raise ValueError for non-2D tensors.

Distributed refusal should have high repair ratio.

RES should always be between 0 and 1.

Should detect residual refusal signal when direction wasn't removed.

Records already in local JSONL should not be re-added.

Calling restore_from_hub() twice should be a no-op the second time.

Gini is always in [0, 1].

Should return all expected keys.

Should return None when no router is found.

Unknown router name should be auto-detected and projected.

Router bias should be projected when project_biases=True.

Router/gate weight should have refusal direction removed.

Router should always use the shared direction, not per-expert.

Calling main(['run']) without a config path should error.

Only 'done' status should record durations.

SAE-derived directions should be unit normalized.

SAE should train and reconstruct activations.

Full roundtrip: save contributions, load them, aggregate.

Full roundtrip ending in LaTeX output.

CKA should be invariant to isotropic scaling.

Categories with distinct directions should have high selectivity.