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. Source tools need the `[full]` extra (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.

Noise covariance (drives the whitening — get it wrong and the map is wrong)

• Source: pre-stimulus baseline (for evoked) or empty-room (MEG)? Enough samples to estimate

it stably (rank!)? Was the data rank-reduced by average reference / interpolation / ICA — and does the covariance reflect that rank?

Inverse method & regularization

• Which inverse: MNE / dSPM / sLORETA / eLORETA (distributed) or LCMV / DICS (beamformer) or

mixed-norm (sparse)? Each carries a different bias — and ⚠️ MNE/dSPM have a depth bias (superficial sources favored); sLORETA/eLORETA reduce location bias; beamformers assume uncorrelated sources.

• SNR / regularization (λ²): arbitrary or justified? λ² ≈ 1/SNR² controls spread vs noise; an

unjustified value is a hidden free parameter.

The spatial claim & inference

• What is the spatial claim, and at what resolution? (a lobe? a gyrus? a single vertex?)

Template-head EEG cannot license gyral-level claims.

• Source-space multiple comparisons: source maps are smooth → neighbouring vertices are

dependent → no single-vertex/single-time claim without correction. Plan cluster-based permutation in source space, not per-vertex thresholding.

PHASE 2 — ANALYZE

1. Capability + look first. mne_check_status (confirm the [full] extra: nibabel, pyvista);

review the evoked/epochs you will invert (mne_plot_evoked / mne_describe) — a clean, baselined evoked with a sensible reference and montage is the prerequisite for a meaningful inverse.

1. Noise covariance from the pre-stimulus baseline (or empty-room for MEG):

   # epochs already in the session; baseline = up to t=0
   noise_cov = mne.compute_covariance(epochs, tmax=0.0, method="auto", rank="info")

(Structured form: mne_compute_noise_cov(name="epochs", tmax=0.0, cov_name="noise_cov"). rank="info" respects rank loss from average reference / interpolation / ICA.)

1. Forward model — template head (fsaverage), downloaded once. State the caveat aloud:

   # template-head EEG forward for the object's montage (fsaverage BEM)
   # -> estimates are EXPLORATORY; no individual MRI / co-registration

(Structured form: mne_make_forward(name="evoked", fwd_name="fwd") — fetches fsaverage on first use.)

1. Apply the inverse (start with dSPM; report the peak vertex/time):

   inv = mne.minimum_norm.make_inverse_operator(evoked.info, fwd, noise_cov)
   stc = mne.minimum_norm.apply_inverse(evoked, inv, lambda2=1.0/3.0**2, method="dSPM")

(Structured form: mne_apply_inverse(evoked_name="evoked", fwd_name="fwd", cov_name="noise_cov", method="dSPM", snr=3.0, stc_name="stc"). Try sLORETA/eLORETA to check depth-bias sensitivity.)

1. Beamformer (LCMV) / DICS / mixed-norm via mne_run_code when appropriate:

   from mne.beamformer import make_lcmv, apply_lcmv
   data_cov = mne.compute_covariance(epochs, tmin=0.0, tmax=0.3)     # active window
   filters  = make_lcmv(evoked.info, fwd, data_cov, reg=0.05, noise_cov=noise_cov)
   stc_lcmv = apply_lcmv(evoked, filters)                            # uncorrelated-source assumption

(DICS: make_dics on a CSD from csd_morlet. Sparse: mne.inverse_sparse.mixed_norm.)

1. Plot + report. Render the cortical map and state the peak location/time with the

template-head caveat: mne_plot_source_estimate(stc_name="stc", hemi="both", time=None) — read the PNG, then archive the equivalent code + figures (the mne-analyst archiving convention).

Best-practice reminders: baseline-correct and choose the reference before inverting; estimate covariance with the correct rank; always state the template-head / exploratory caveat; compare methods (dSPM vs sLORETA) rather than trusting one map.

PHASE 3 — CRITIC (before believing the result)

Hand the model + result (head model, co-registration, covariance source, method, λ²/SNR, the spatial claim) to `mne-methodology-critic` (invoke the skill, or dispatch it as a subagent with references/methodology-checklist.md). For source work it will specifically check:

• template-head localization: a quantitative source claim with **no individual MRI /

co-registration** is exploratory — WARN, or FAIL if stated quantitatively (gyral/Brodmann claim);

• depth bias of MNE/dSPM (superficial sources favored) — was it acknowledged / mitigated?
• arbitrary regularization / SNR (λ² a hidden free parameter) — justified or swept?
• single-vertex / single-time inference without correction — source maps are smooth →

neighbours are dependent → cluster-based permutation in source space, not per-vertex;

• misestimated noise covariance (wrong baseline, too few samples, ignored rank loss);
• EEG source localization overinterpreted (volume conduction + skull conductivity uncertainty).

Report its BLOCK / REVISE / PASS verdict to the user and act on it before stating where the activity is.

See references/source-methods.md for deeper recipes (covariance & rank, BEM/template vs individual MRI, MNE/dSPM/sLORETA/eLORETA depth bias, LCMV/DICS, mixed-norm, and source-space inference choices).