neuroscience

name	neuroscience
description	Brain data from acquisition to inference: what each modality actually measures (BOLD haemodynamics, EEG/MEG timing, spikes and LFP, calcium kinetics, optogenetic controls), BIDS layout and NIfTI/EDF/NWB formats, preprocessing choices that move results, and the traps the field is famous for — circular ROI selection, double dipping, voxel-wise multiple comparisons, cluster-extent inflation, tiny-n designs, overfitted decoders and connectivity claims. Use for fMRI, EEG/MEG, electrophysiology, imaging or any analysis of neural recordings.
version	1.0.0
author	Remedy
tags	["research","neuroscience","fmri","eeg","electrophysiology","imaging","statistics"]
requires	[]
tools	["analysis_env","analysis_run","analysis_ledger","data_profile","stats_assumptions","stats_multiplicity","power_analysis","lit_search","cite_add","skill_activate"]
triggers	["\\b(fMRI|\\bEEG\\b|\\bMEG\\b|BOLD signal|diffusion MRI|\\bDTI\\b|resting[- ]state)\\b","\\b(spike (?:sorting|trains?)|local field potentials?|patch[- ]clamp|optogenetic\\w*|two[- ]photon|calcium imaging)\\b","\\b(cluster[- ](?:wise|level) correction|voxel[- ]wise|whole[- ]brain correction|\\bROI\\b analysis|double dipping|circular analysis)\\b","\\b(fMRIPrep|FreeSurfer|SPM12|\\bFSL\\b|AFNI|MNE-?Python|NIfTI|BIDS dataset)\\b"]

Neuroscience (neural data and its inferences)

Run skill_activate(skill="research-method") first and work off that spine — question framing, evidence, preregistration, citation honesty, "we do not know". This pack covers only what neural data adds.

The two sentences that prevent most of the damage

1. Say what the signal is. BOLD is a haemodynamic proxy lagging neural activity by seconds; EEG/MEG is millisecond-resolved summed postsynaptic current with an ill-posed source problem; a spike is one cell's output; calcium fluorescence is a filtered, nonlinear spike proxy; LFP is local input and processing, not output. Write the claim in the units the method actually delivers. references/modalities-and-what-they-measure.md.
2. Every selection step must be independent of the effect being tested. Picking voxels, channels, time windows, ROIs, cells or trials by the contrast you then test is circular, and the resulting statistic is not interpretable. references/statistical-traps.md.

Decision tree

1. analysis_env(path) — what runs here (Python/MNE, FSL, SPM, AFNI, FreeSurfer, fMRIPrep, R, MATLAB). Nothing is verified on a machine that has none of it; say so rather than describing a pipeline you cannot run.
2. Layout the data as BIDS before analysing, and validate it. Raw stays read-only; derivatives live under derivatives/<pipeline>. Formats, sidecars and the metadata that is easy to lose (TR, slice timing, PhaseEncodingDirection, reference electrode, sampling rate, units): references/bids-and-data-formats.md.
3. Preprocess deliberately. Every choice — motion handling, slice timing, filter cutoffs, referencing, ICA rejection, normalisation, smoothing kernel, denoising — changes the result. Record the exact pipeline version and settings; run it through analysis_run so it lands in the ledger with input hashes. references/preprocessing-choices.md.
4. Size the study honestly. power_analysis on the effect that matters, in the units of the second-level test (subjects for group inference, cells or sessions for within-animal). Small-n neuroimaging does not produce small effects — it produces inflated, unstable ones. Trials per condition and subjects are different axes; more trials do not rescue too few subjects for a group claim.
5. Correct across the whole search space. Voxels, vertices, channels, time points, frequency bins, ROIs, contrasts. stats_multiplicity for discrete families; permutation/TFCE for images. references/multiple-comparisons-imaging.md.