Everything behind the arc is open and runnable. Replicate a paper in one command, run probe-causality experiments from your own agent on your own GPU, or pull the methodology as agent skills. We never see your model, data, or keys.
Every paper in the arc replicates from a single command on the Colab CLI, with the verdict auto-checked against the published numbers.
$ oilab replicate lever-is-lateAn MCP server that lets any agent — Claude Code, Cursor, Cline — run probe-causality experiments on your own GPU session. We never see your model, data, or keys.
$ pip install "openinterp-mcp[server]"The methodology is packaged as agent skills: how to capture activations, decompose them into named SAE features, steer a direction, and run the four causality checks — including the structure-matched control + naming gate — that separate a real result from a confounded or epiphenomenal one.
A ladder of runnable notebooks, from your first SAE in 30 minutes to the full-stack experiments behind the papers — each opens in Colab.
Drop these into any agent and it inherits the operational knowledge — including the four causality checks (with the structure-matched control + naming gate) that separate a real result from a confounded or epiphenomenal one. Each maps to a typedopeninterp-mcptool.
$ curl -fsSL https://openinterp.org/install-skills.sh | shDownloads each SKILL.md into ~/.claude/skills — writes only markdown, runs no code. Use -s -- --project for a repo-local ./.claude/skills, or inspect the script first.
Attach to your running Colab / vast.ai openinterp session via its public HTTPS URL — the first step of any run.
Health of the active session — loaded model, probes, and captures in memory — before you spend a forward pass.
Capture residual-stream activations at chosen layers and token positions during a forward pass.
Inventory the probes loaded in the backend — model, layer, position, source — so you know what to evaluate or steer.
Apply a loaded linear probe to a stored capture; returns per-sample scores and AUROC when labels are given.
Decompose a captured activation into its top-K SAE features and read their names — the bridge from a residual vector to human-readable concepts (full-stack SAE on Qwen3.6-27B).
Inject direction×alpha into the residual stream and observe the behavioral effect — causal, not correlational.
The four mandatory checks — random-feature baseline, control-token norm, structural-rigidity α-sweep, and the structure-matched control + naming gate — that separate a causal probe from a confounded or epiphenomenal one.
Operate a full mech-interp lab from the terminal — provision Colab GPUs via the Google Colab CLI, run the loops, replicate the papers.
Install: pip install "openinterp-mcp[server]" (v0.1.0) · point your agent's MCP config at it · the skills live in each repo's skills/.
Everything is Apache-2.0 and reproducible. Extend a probe, replicate a result, or disagree with one — the methodology is built to be argued with.