OpenInterpretability ORG

Mechanistic interpretability as reward signal for RL training of LLMs — SAE features + GRPO + anti-Goodhart framework

Train your first SAE in 30 min → paper-grade at 27B. Free Colab · free Kaggle · cloud ladders. Every scale covered.

Next.js site for OpenInterpretability — the umbrella org for mechreward and public hybrid-architecture SAEs

Instrumented agent harness for capturing SAE feature trajectories during SWE-bench Pro traces on Qwen3.6-27B (mech anatomy of agent reasoning failure)

Find the layer where a language model commits a decision — and steer it. Any open-weight HF model. (WANDERING arc paper #6)

Defense-in-depth action firewall for tool-using agents, with a model-internal intent brake that closes the model-origin blind spot. Built on Zenodo 10.5281/zenodo.20679287.

Bring-your-own-agent infrastructure for mechanistic interpretability research. MCP server + Colab backend. Privacy-first.

ProbeBench probe submission registry — open a PR adding probes/<your-probe-id>.yaml

Agent-operable mechanistic-interpretability lab on the Google Colab CLI — one-command paper replication, GPU research loops, and an agent skill file

Organization profile & shared community health files for OpenInterpretability

Inspect AI eval for WANDERING failure-mode detection in agent trajectories. Companion to Tool-Entropy Collapse paper (DOI 10.5281/zenodo.20368601).

Public index of mechanistic-interpretability publications via openinterp-mcp. Schema, index, atlas entries, replications, ProbeBench submissions.

openinterp — Python SDK + CLI. FabricationGuard hallucination probe + ProbeBench leaderboard + Atlas search + Trace generation. pip install openinterp