sae-lookup

Turn a raw activation into a short list of named SAE features. This is the entry point to the full-stack SAE on Qwen3.6-27B (caiovicentino1/qwen36-27b-sae-fullstack, 11 layers, d_sae=40960, k=128) — the rare asset that lets you say which concepts are active, not just how active a direction is.

When to use

• A capture is interesting and the researcher wants to know what concepts fired, not just a probe score
• Naming a feature for the causality-protocol Check 4b (a probe claim needs a named feature)
• Exploring a layer to find candidate directions before training a probe
• Comparing two prompts: decompose both, diff the top features

How to use

Call the MCP tool `sae_lookup` with:

• sae_id: an SAE matching the active model and the layer (e.g. a layer-59 SAE for an L59 capture)
• capture_id: a stored capture (from capture_acts)
• layer: the layer whose activation to decompose — must be one the capture stored and the SAE was trained for
• top_k (default 32): how many features to return

Returns top_features: feature_id, activation, and an auto-interp description when available, plus the manifest SHA256.

Output to surface

List the top features with their names and activations, highest first:

Layer {layer}: #{fid} "{description}" ({act:.2f}) · #{fid} "{description}" ({act:.2f}) · …

When the researcher is naming a feature for a probe, surface the max-activating description of the relevant feature and hand it to causality_protocol as feature_name — that is Check 4b.

Do not

• Do not use an SAE trained for a different model or layer — the tool 400s on a d_model_mismatch.

Match the SAE to the active model and the capture's layer.

• Do not treat a high single-feature activation as a causal claim. A name is necessary but not

sufficient — run causality-protocol for the causal verdict.