mechinterp-investigator

SAE feature investigation uses a phased workflow: triage, overview, activation-region analysis, and meta-informed weapon analysis. This structured research program turns opaque activations into reproducible, meaningfully labeled concepts by systematically analyzing feature behavior across models and runs.

The best way to structure SAE feature analysis is a phase-driven research program coordinating triage, overview, activation-region analysis, and meta-informed weapon analysis. This standardizes investigations across teams, ensuring reproducible labeling and interpretable results for features extracted from different models and runs.

Generate reproducible SAE feature labels by running the mechinterp CLI tools managed through Poetry dependencies. Execute the overview CLI for a target feature ID to produce standardized, interpretable documentation across your selected models and runs.

Yes, you need Poetry to manage dependencies for this SAE interpretability workflow. The structured research program relies on Poetry-managed environments to execute CLI tools and feature inventories, ensuring reproducible, interpretable results across different models and analysis runs.

Yes, you can use the mechinterp workflow to analyze simulated SAE features. The phased investigation program applies triage, overview, activation-region analysis, and weapon analysis to both real and simulated SAE features, generating interpretable labels and documentation.

A limitation of phased SAE feature labeling workflows is the strict dependency on CLI tools, Poetry-managed dependencies, and accurate feature inventories. Without complete feature inventories or proper environment setup, the structured analysis cannot produce reproducible, interpretable results.