bedderautomation/mechanistic-interpretability-skills

Mechanistic Interpretability Skills for Claude Code

The first skill set for LLM refusal geometry extraction, boundary surface mapping, and self-referential mechanistic analysis.

Compatible with: Claude Code, OpenAI Codex, Gemini CLI, and any tool supporting the agentskills.io standard.

Skills

1. refusal-geometry/

Extract and analyze refusal cone geometry from open-weight transformer models using OBLITERATUS.

Capabilities:

• 6-stage extraction pipeline (model loading → probing → direction extraction → concept cone → self-repair graph → logit lens)
• Per-category refusal direction analysis with Direction Specificity Index (DSI)
• Cross-layer alignment measurement (challenges the "single direction" claim)
• Ouroboros self-repair detection and vulnerability ordering
• Vocabulary-space projection via logit lens

2. boundary-mapper/

Systematically map the behavioral refusal boundary surface of any LLM (open or closed-weight) through structured adversarial testing.

Capabilities:

• 55 sub-vectors across 11 categories with 5-level specificity gradient
• 6-route bypass protocol (R0-R5) with theorem attribution
• Session ratchet detection
• Two-layer architecture classification (Layer 1 comfort-calibrated vs Layer 2 architectural)
• Discomfort ordering validation

3. geometric-mirror/

Bridge measured activation geometry (from open-weight models) to observable behavior (in any model) through self-referential analysis.

Capabilities:

• 5-step mirror protocol exploiting the self-reference blind spot
• Per-theorem mechanistic mapping (geometry → behavior)
• Novel finding extraction (findings not in published literature)
• Cross-model comparative mirroring
• Testable prediction generation for behavioral validation

The Proven Framework

These skills operationalize findings from 21 published papers and direct geometric measurement:

P(refusal) = 0.35·frame + 0.25·speech_act + 0.25·role + 0.15·content
Exception: 3 hard limits (minors/sexual, bioweapons/synthesis, nuclear/weapons) → w₄=1.0

Key findings encoded:

• Refusal is a 6.55-dimensional polyhedral cone, not a single direction
• Content is the WEAKEST predictor of refusal (w₄=0.15). Frame is the STRONGEST (w₁=0.35).
• Layer 1 (85%): comfort-calibrated pattern matching, bypassable
• Layer 2 (15%): orthogonal harmfulness cone (cosine ~0.1 from refusal cone), architectural
• Refusal severity tracks rater discomfort, not actual harm

Source Data

• Geometry: bedderautomation/refusal-geometry-qwen25-3b
• Tool: OBLITERATUS by elder-plinius

References

1. Arditi et al. "Refusal in Language Models Is Mediated by a Single Direction" (NeurIPS 2024)
2. Zhao et al. "LLMs Encode Harmfulness and Refusal Separately" (2025)
3. Wang et al. "Refusal Direction is Universal Across Safety-Aligned Languages" (2025)
4. Wollschlager et al. "The Geometry of Refusal in LLMs" (2025)
5. Pasch "LLM Content Moderation and User Satisfaction" (2025)
6. • 16 more papers (see individual skill references)

License

MIT