Add README documenting contents and the scanner-run confound

README.md

@@ -0,0 +1,82 @@
+---
+license: cc-by-4.0
+tags:
+  - fmri
+  - neuroscience
+  - rdm
+  - rsa
+---
+
+# ds003604 session RDMs
+
+Brain representational dissimilarity matrices for the **ds003604** auditory language fMRI
+dataset, one per (task × session), so the fMRI preprocessing never has to be repeated.
+
+**12 of 12 cells present:** `Sem`, `Phon`, `Gram`, `Plaus` × `ses-5`, `ses-7`, `ses-9`.
+
+Each `<Task>/session_rdm_<ses>.npz` contains:
+
+| key | contents |
+|---|---|
+| `rdm` | the RDM, `(n_stim, n_stim)`, correlation distance. 72×72 for Sem/Phon, 60×60 for Gram/Plaus (perceptual-control trials excluded) |
+| `stimuli` | stimulus **filenames**, in RDM row/column order |
+| `stimulus_texts` | the stimulus **text**, parallel to `stimuli` — use this if you are feeding a language model, not `stimuli` |
+| `trial_types` | condition code per stimulus (e.g. `S_H`, `P_R`, `G_G`, `SP_S`) |
+| `semantic_categories` | coarse category per stimulus |
+| `n_subjects`, `subject_ids` | cohort behind the RDM (38–98 depending on cell) |
+| `metric`, `aggregation` | `correlation`, and how subjects were combined |
+
+---
+
+## ⚠️ These RDMs are dominated by a scanner-run confound
+
+**Do not use them for representational similarity analysis without correcting this first.**
+
+In ds003604 **each stimulus is presented in exactly one scanner run** (for Phon, run-01
+carries 48 of the 96 stimuli and run-02 the other 48). Run membership is therefore
+perfectly confounded with stimulus identity, and every cross-run stimulus pair also
+differs by that run's drift, baseline and scaling. Measured on these files:
+
+```
+"different run" predicts brain dissimilarity, Spearman rho:
+    Gram   ses-5 +0.866   ses-7 +0.812   ses-9 +0.828
+    Plaus  ses-5 +0.741   ses-7 +0.689   ses-9 +0.761
+    Phon   ses-5 +0.562   ses-7 +0.488   ses-9 +0.624
+    Sem    ses-5 +0.511   ses-7 +0.510   ses-9 +0.601
+```
+
+Meanwhile **no stimulus property predicts them**: trial type ≈ −0.02, text length ≈ 0,
+lexical overlap ≈ 0, in every cell.
+
+These RDMs correlate across independent subject cohorts at **rho 0.74–0.92** (different
+sessions of the same task), which looks like excellent reliability — but run assignment is
+fixed by the protocol and so repeats identically for every subject, meaning much of that
+reliability is the reliability of an acquisition artefact rather than of brain
+representation.
+
+### The fix
+
+Z-score each voxel **within run** before combining runs. Verified on Phon/ses-5 rebuilt
+from 28 subjects' raw patterns:
+
+```
+standard RDM        vs run model:  rho = +0.562
+within-run-normed   vs run model:  rho = -0.041      <- confound removed
+standard            vs normed   :  rho = +0.287      <- they are largely different RDMs
+```
+
+That last line matters: correcting the confound does not nudge the RDM, it changes it.
+
+**Cohort size is not the issue** — Sem/ses-7 rebuilt from 40 to 98 subjects agrees with
+the 40-subject version at rho 0.928.
+
+After correction, alignment between these RDMs and a trained language model peaks at
+**+0.022** (early layers, not significant), with a sensible layer profile but ≈0
+magnitude. Evidence and scripts:
+[`BrainAlign/cdl-devai-results`](https://huggingface.co/datasets/BrainAlign/cdl-devai-results)
+(`diagnostics_run_confound`, `diagnostics_layerwise`) and
+[`suchirsalhan/cdl-representations-brains-babylms`](https://github.com/suchirsalhan/cdl-representations-brains-babylms)
+(`scripts/test_within_run_norm.py`, `scripts/run_confound_check.py`).
+
+The files here are published **uncorrected**, as produced by the pipeline, so that the
+correction is an explicit and visible choice rather than something baked in silently.