capture-acts

Run a forward pass through the researcher's loaded model, hooking specified layers, extracting activations at named positions. Activations stay in the Colab session memory; only metadata (shape, capture_id, manifest SHA256) crosses the wire.

When to use

Before any of: probe-eval, causality-protocol, sae-lookup. These all need a capture_id.

Also when the researcher asks "what are the activations at L20 for this prompt?" or "store the residual at L11/L31/L55 for {prompt}".

How to use

Call the MCP tool `capture_acts` with:

• prompt: the text to run through the model
• layers: list of layer indices (e.g. [10, 20, 27])
• positions: list of position specs. Defaults to ["end_question"].
• end_question / end_prompt / last_token → last input token
• first_token → token 0
• integer <int> → absolute index (negative counts from end)
• "end_question:-2" → 2 tokens before end

Output to surface

"Captured L{layers} × {positions} → capture_id {id[:8]}… (d_model={d}, manifest sha {sha[:12]})"

Hand the capture_id to whatever comes next. Don't try to read the raw tensors; they live in the Colab session.

Cost / time

A single forward pass on Qwen2.5-7B at bf16 on a T4 is ~2 s; on A100 ~0.5 s. Larger models scale. Don't batch hundreds of prompts in a tight loop without warning the researcher.