mechinterp-patterns
Reference catalog of recurring SAE feature patterns - check when labeling new features, add new patterns as discovered
When & Why to Use This Skill
This Claude skill serves as a specialized reference catalog for Mechanistic Interpretability (MechInterp), specifically designed for Sparse Autoencoder (SAE) feature investigation. It enables researchers to identify, label, and document recurring neural network feature patterns—such as 'Death-Averse' or 'Ink Efficiency'—ensuring high consistency across experiments and accelerating the discovery of model internals through a standardized framework.
Use Cases
- Standardizing Feature Labeling: Quickly match newly discovered SAE features against established archetypes to maintain consistent terminology and classification across large-scale interpretability projects.
- Pattern Discovery and Documentation: Identify and record novel recurring themes in neural network activations, allowing for the systematic promotion of 'suspected' patterns to 'confirmed' status after multiple observations.
- Workflow Acceleration: Streamline the initial phases of feature investigation by using the catalog's key signatures to categorize family breakdowns and weapon distributions without starting from scratch.
- Knowledge Management for AI Research: Maintain a centralized, version-controlled repository of model behavior patterns to facilitate collaborative research and prevent redundant analysis.
| name | mechinterp-patterns |
|---|---|
| description | Reference catalog of recurring SAE feature patterns - check when labeling new features, add new patterns as discovered |
MechInterp Patterns
A growing catalog of recurring patterns observed during SAE feature investigation. Use this to quickly recognize common archetypes and maintain consistency in labeling.
Pattern Catalog Location
The main patterns document is at:
/mnt/e/mechinterp_runs/potential_patterns.md
Always read this file when:
- Starting to label a new feature - check if it matches a known pattern
- You've identified what seems like a recurring theme
- You want consistent labeling terminology
Quick Pattern Reference
| Pattern | Key Signature | Example Label |
|---|---|---|
| Death-Averse | SCU/SPU driver + QR/SS/CB suppressed | "Offensive Intensity (Death-Averse)" |
| Zombie/Respawn | QR driver + Comeback/SJ enriched | "QR Stacker (Slayer)" |
| Ink Efficiency | ISM/ISS/IRU driver + mobility suppressed | "Ink Economy (Anti-Mobility)" |
| Mobility | SSU/RSU driver + ink suppressed | "Stealth Mobility Build" |
| Weapon-Gated | >50% single weapon | "{Weapon} {Ability} Build" |
| Special Power | SPU/SCU driver + threshold effects | "Special Power Stacker" |
| Weak/Auxiliary | <10% decoder percentile + no structure | "Weak/Aux Feature {ID}" |
When to Update the Patterns Document
Add a New Pattern When:
- You see a signature that doesn't match existing patterns
- The pattern has a clear "why would the model learn this?" explanation
Add a Feature to Existing Pattern When:
- The feature clearly matches the signature
- It reinforces the pattern interpretation
Promote to Confirmed When:
- 3+ features show the same pattern
- Signature is consistent across features
Usage in Investigation Workflow
Phase 1 (Overview): After running overview, check patterns document:
Does the family breakdown + weapon distribution match a known pattern?
If yes → Use pattern's label style and run confirmatory experiments
If no → Standard deep dive, potentially discover new pattern
Phase 5 (Label Proposal): Before finalizing label:
Is this a known pattern? → Use consistent terminology
Is this a new pattern? → Add to "Suspected Patterns" in the document
Maintaining Consistency
When labeling features that match known patterns, use consistent terminology:
| Pattern | Preferred Terms |
|---|---|
| Death-Averse | "Death-Averse", "Anti-Zombie" |
| Zombie | "Zombie", "Respawn-Focused" |
| Mobility | "Mobility", "Movement" |
| Ink Efficiency | "Ink Economy", "Ink Efficiency" |
| Weapon-Gated | "{Weapon} Build", "{Weapon}-Specific" |
See Also
- mechinterp-investigator: Main investigation workflow
- mechinterp-glossary-and-constraints: Ability families and terminology
- splatoon3-meta: Game meta knowledge for pattern interpretation