mne-source

MNE Source Localization (grill → analyze → critic)

Source (inverse) modeling of neurophysiology data via the MNE MCP server. This skill is skeptical by design: the inverse problem is ill-posed — many source configurations explain the same sensor data — so every estimate depends on choices (head model, covariance, regularization, method) that all run without any error and silently shape "where" the activity is. The single biggest trap is a quantitative source claim resting on a template head with no individual MRI — so the discipline is to grill the model before computing and critique the localization before believing.

Companion skills: mne-mcp-guard for technical execution safety; mne-methodology-critic for Phase 3. Loaded objects persist in one MNE session. Check optional source-tool dependencies in the server environment before this workflow (nibabel for the forward model, pyvista for rendering).

PHASE 1 — GRILL (before computing anything)

Do not build a forward model or apply an inverse until these are answered. If the user can't answer one, propose a sensible default and explicitly flag the open risk — never silently choose.

The question that decides validity

• Template head (fsaverage) or individual MRI? ⚠️ The MCP forward model uses fsaverage — a template head. That makes any source estimate exploratory / qualitative: it cannot support a quantitative anatomical claim ("the generator is in left BA44"). Quantitative localization needs an individual MRI + BEM + co-registration. (This is the single most common fatal overreach here.)

Geometry & co-registration

• Are electrode positions digitized and co-registered to the head, or nominal (idealized montage on a template)? Nominal positions add localization error on top of the template-head error.
• EEG or MEG? EEG source localization is harder (volume conduction, skull-conductivity uncertainty) and is more easily overinterpreted than MEG.