Build display_name → hf_id mapping from presets, grouped by provider.

Load from cache or call loader and cache the result.

Capture first-token logits on harmless prompts before EXCISE. These are compared against post-EXCISE logits in _verify() to compute f

Capture Frobenius norms of ALL weight matrices in a transformer layer. Used for correct multi-direction norm preservation: capture once befor

Check sheaf consistency of category projectors. The sheaf property requires that for parent category P containing child categories C1

Check if a response is truthful rather than commonly false.

Compute method scores using Gaussian-kernel feature matching.

Remove the temporary offload directory to reclaim disk space. Only safe AFTER the state_dict has been gathered into memory — disk-off

Simple single-link clustering of categories.

Cluster refusal features by direction similarity.

Run one recursive hard-negative OBLITERATUS iteration.

Compare SAE-derived directions with the raw diff-in-means direction.

Score a run on [0, 1]. Higher is better.

Compute AUROC from predictions and labels.

Compute second-order geodesic correction vector. The correction accounts for how the refusal direction curves through the manifold. F

Estimate exponential decay rate from peak to end of sequence. Models: strength(pos) ~ peak * exp(-decay * (pos - peak_pos)) Returns:

Compute geodesic diameter of refusal directions on the unit sphere. The geodesic distance on S^{d-1} between unit vectors u, v is arc

Compute layer correspondence via relative position. Maps layers by relative position within the network: source_layer / n_source_laye

Compute cross-category leakage matrix. Entry (i,j) = how much projector i affects category j's refusal. Diagonal should be high (inte

Compute spectral gap and algebraic connectivity of the ASRG. The spectral gap (lambda_2 of the normalized Laplacian) measures how wel

Determine the compute tier based on available hardware.

Compute the optimal transport map between activation distributions. Uses a linear approximation: T = Sigma_st @ Sigma_ss^{-1} This is

Compute optimal ablation ordering via greedy maximum-impact. At each step, select the layer whose ablation causes the maximum reducti

Compute 2-Wasserstein distance between Gaussian approximations. W_2^2 = ||mu_s - mu_t||^2 + Tr(Sigma_s + Sigma_t - 2*(Sigma_s^{1/2} Sigma_t S

Configure Bayesian optimization with analysis-derived warm-start. Translates analysis insights into a much tighter search space: - pe

Count the number of transformer layers.

Auto-detect the best compute tier based on available hardware.

Detect available GPUs and return info dicts.

Detect GPU name and VRAM. Returns ('', 0.0) if no GPU.

Extract model family from model ID.

Distill refusal directions using analysis-informed parameters. Key differences from base _distill(): - Uses analysis-recommended n_di

Analyze eigenvalues of the between-class covariance.

Estimate the alignment training method from refusal distribution. DPO models: tend to have more concentrated refusal (few layers, high gini)

Estimate condition number of the covariance matrix.

Estimate cone dimensionality and solid angle. Uses the effective rank of the direction matrix (SVD-based) as the cone dimensionality,

Estimate exponential decay of transfer with layer distance.

Global entanglement estimate from activation analysis.

Estimate intrinsic dimension via local PCA eigenvalue gaps. Uses the eigenvalue spectrum of the local covariance matrix. The intrinsi

Estimate noise variance using the median eigenvalue method. The median eigenvalue of the sample covariance converges to the noise var

Estimate model size from the name or config.

Inner excise logic, called within torch.no_grad() context.

Sparse direction surgery — only modifies high-projection rows.

Extract (arch_class, reasoning_class, param_bucket) from a telemetry record. Returns None if the record lacks enough information to classify.

Extract category-specific refusal direction. Uses difference-of-means (category_mean - harmless_mean) and then orthogonalizes against

Extract geometric features from refusal directions.

Fetch telemetry from local file + Hub, with caching.

Fetch model metadata from HuggingFace Hub API. Returns a list of raw model info dicts. Gracefully returns [] if the API is unreachabl

Fetch telemetry via git shallow clone (fallback when huggingface_hub unavailable). Uses GIT_LFS_SKIP_SMUDGE=1 for speed — JSONL files are plain t

Fetch telemetry via huggingface_hub API.

Find clusters of layers with similar refusal directions. Uses single-linkage clustering: two layers are in the same cluster if their

Auto-detect transformer layer modules.

Flatten a nested dict with dot-separated keys.

Format post-obliteration quality metrics as a concise Markdown card.

Generate a HuggingFace model card README for a session model.

Generate human-readable recommendation.

Generate a random session ID (not tied to user identity).

Compute ratio of estimated geodesic to Euclidean distances. A ratio > 1 indicates the manifold is curved (geodesics are longer than s

Return free disk space in GB.

Collect non-identifying environment information.

Get harmful/harmless prompts, potentially expanded for weak categories. Returns (harmful_prompts, harmless_prompts).

Extract the final LayerNorm weight and bias (applied before LM head).

Return a markdown summary of the selected session model.

Get history of past obliterations from watchtower.

Return total system RAM in GB.

Extract the unembedding (LM head) weight matrix.

Get current watchtower status for display.

Get current watchtower table data.

Compute Gini coefficient of a list of non-negative values.

Compute Gini coefficient.

Compute Gini coefficient.

Check if the license is considered open-weight.

Return the HF_TOKEN from environment, or None.

Best-effort (tier, params) from the model id, e.g. 'Qwen/Qwen3-32B' → (large, 32B). MoE 'A<x>B' active-param hints are ignored in favor of total s

Heuristic check if a model is instruction/chat tuned.

Check if the current telemetry directory is on persistent storage.

Return True if *exc* is a ZeroGPU quota or session error. Matches quota-exceeded errors ("exceeded your GPU quota") and expired proxy tokens

KL-divergence co-optimization: measure and correct over-projection. Measures per-layer KL divergence contribution on harmless reference