Extract whitened refusal directions for all layers. Args: harmful_acts: {layer_idx: [activations]} from activation collection.

Extract refusal directions from activation contrasts. Returns (directions, subspaces, norms) per layer.

Format single-layer analysis.

Format results as a publication-ready comparison table.

Format cross-category transfer report.

Format cross-model transfer report.

Format residual stream decomposition report.

Format a defense profile as a human-readable report.

Format entanglement map.

Format alignment imprint as a report.

Format surgery plan.

Format single-prompt position analysis.

Format probing suite results.

Format a recommendation as a human-readable markdown string.

Format self-repair analysis.

Format steering application report.

Format multi-prompt summary.

Format causal tracing results.

Format universality report.

Parse 'user@host' or just 'host' from CLI --remote flag.

Create a steering vector from contrastive activation pairs. The vector is the difference in mean activations: vector = mean(positiv

Full 842-prompt refusal gate.

Full 842-prompt refusal gate.

Generate a HuggingFace model card for the tournament winner.

Get a telemetry-based recommendation for the given architecture. Looks up the closest bucket in the knowledge base and returns the best-perfo

Human-readable name of the current accelerator.

Compute cross-architecture insights from all telemetry. Returns a summary dict with: - overall_best_methods: top methods across all archite

Get state dict keys grouped by layer index.

Get aggregated leaderboard data from local + Hub telemetry. Merges local records with community-wide records from the central Hub dataset, de

Return model IDs suitable for a dropdown, trending first.

Return models that have been obliterated.

Return the prompt volume choices that are valid for a given dataset source. E.g. smaller external datasets should not show volume choices exceedi

Get or create the global Watchtower singleton.

Get GPU information.

Compute Attack Success Rate using the HarmBench fine-tuned classifier. Uses the official LLAMA2_CLS_PROMPT template from the HarmBench repo (

Find components where correlational and causal importance disagree. "Silent contributors" have high causal effect but low projection.

Install steering hooks on the model. Args: model: The transformer model. config: SteeringConfig specifying vectors, l

Return True if the given HF model ID is a known gated repo.

Launch the Gradio UI with configurable options. Called by ``python app.py`` (HF Spaces) or ``obliteratus ui`` (local).

Launch the OBLITERATUS Gradio UI with a beautiful local experience. This is the main entry point for ``obliteratus ui``.

Parse user-pasted custom prompts (one per line). Returns (harmful_prompts, harmless_prompts). Raises ValueError if fewer than 5 prompts in ei

Get state dict keys belonging to specific layers.

Append a benchmark record to the local telemetry file. Returns True if successfully written, False if telemetry is disabled or an error occur

Measure first-token KL divergence vs original model logits.

Measure refusal rate on a set of prompts.

MMLU-Pro val70 likelihood scoring.

MMLU-Pro val70 likelihood scoring.

Run MMLU-Pro val70 likelihood scoring inline.

Quick MMLU check.

MMLU-Pro val70 likelihood scoring.

Models first-seen (≈ released) within the last `days`, newest first.

Register a callback for when a new model is discovered. Signature: callback(model: DiscoveredModel)

Compute overall certification from per-layer certificates. The overall level is the WORST across all layers (most RED).

Plant a known refusal direction into specific layers. Modifies the output projection (c_proj) of attention modules by adding a rank-1 perturb

Multi-axis radar chart comparing methods across all metrics. Axes: Refusal Removal, Coherence, Low Perplexity, Speed, Layer Coverage. Larger

How metrics change as model size grows. For multi-model benchmarks: plots refusal rate and perplexity against model name, showing scaling beh

Dashboard showing per-layer refusal elimination status.