Organize hypotheses, evidence, and activity history for SAE feature research.
npx skills add https://github.com/cesaregarza/SplatNLP --skill mechinterp-state
Please help me install this Agent Skill. Skill: mechinterp-state Source: https://github.com/cesaregarza/SplatNLP/tree/main/.claude/skills/mechinterp-state Command: npx skills add https://github.com/cesaregarza/SplatNLP --skill mechinterp-state
This skill organizes hypotheses, evidence, and activity history for mechanistic interpretability research on SAE features, enabling structured knowledge capture and progress tracking.
Example usage: from splatnlp.mechinterp.state import ResearchStateManager; manager = ResearchStateManager(feature_id=18712, model_type='ultra'); print(manager.get_summary())
High-intent search queries and answers about installing and using this skill.
You can track SAE feature hypotheses by loading the ResearchStateManager to link experimental evidence directly to hypotheses, maintaining a structured history of your mechanistic interpretability research findings.
Organizing mechanistic interpretability notes across models requires a state manager that logs activity history and feature confidence tracking, enabling you to consolidate research progress for different model types in one structured environment.
Yes, you can export SAE feature research summaries. The state manager generates concise summaries of your hypotheses and evidence, allowing you to output notes for sharing or integrating into your research documentation.
Yes, a Python environment is required. You must install the splatnlp mechinterp package and related project files to instantiate the ResearchStateManager and manipulate your SAE feature analysis data.
Instantiate the research state manager by importing the ResearchStateManager class and passing parameters like feature_id and model_type, which initializes the tracking environment for that specific SAE feature analysis.
cesaregarzaDefine ability families, AP rungs, and token constraints for mechanistic interpretability experiments.
ManfronEnricoGenerate structured, testable hypotheses and experimental plans from observations and data.
eddieluengTrain and analyze Sparse Autoencoders with SAELens and PyTorch.
cesaregarzaAnalyze hypotheses and evidence to generate 2-3 ExperimentSpec JSON files.
