MI Analysis Operational Reference

Community

End-to-end MI analysis for prompt interpretability.

Authortaylorsatula
Version1.0.0
Installs0

System Documentation

What problem does it solve?

This operational guide documents how to run TeaLeaves' mechanistic interpretability (MI) pipeline end-to-end, covering region annotation, attention capture, logit-lens projections, and result rendering to support prompt engineering decisions.

Core Features & Use Cases

  • Region annotation: define named spans in prompts and map them to token positions to analyze focus and region interactions.
  • Attention capture & logit lens: hook attention across layers and project residual streams to understand token decisions.
  • Visualization & comparison: generate heatmaps, cooking curves, layer sweeps, and multi-sample reports to compare variants and seeds.

Quick Start

Run the MI analysis workflow on a test_cases.json using the GPU-enabled run_analysis.py to produce per-case results and visuals.

Dependency Matrix

Required Modules

None required

Components

Standard package

💻 Claude Code Installation

Recommended: Let Claude install automatically. Simply copy and paste the text below to Claude Code.

Please help me install this Skill:
Name: MI Analysis Operational Reference
Download link: https://github.com/taylorsatula/TeaLeaves/archive/main.zip#mi-analysis-operational-reference

Please download this .zip file, extract it, and install it in the .claude/skills/ directory.
View Source Repository

Agent Skills Search Helper

Install a tiny helper to your Agent, search and equip skill from 471,000+ vetted skills library on demand.