devai sweep 2026-08-29: null-referenced results

README.md

@@ -1,92 +1,82 @@
-# Brain–language-model alignment: ds006239
+---
+license: mit
+task_categories: [feature-extraction]
+tags: [brain-alignment, fmri, rsa, developmental, pythia, null-result]
+---
 
-Wang et al. 2025 — word-level phonological and semantic reading tasks in children and adolescents aged 10–17.
+# Brain-LM alignment: ds006239
 
-- Paper: https://www.sciencedirect.com/science/article/pii/S2352340925009692
-- Data: https://openneuro.org/datasets/ds006239/versions/1.0.5
-- Generated: 2026-08-28
-- Pipeline: https://github.com/suchirsalhan/cdl-representations-brains-babylms
+Representational-similarity alignment between language-model hidden states and
+child fMRI RDMs for **ds006239** (children ages 10-17, reading).
 
-## Read this first: does the measurement work?
+- **Tasks:** Orth, Phon, Sem, SemLocal   **Sessions:** ses-11, ses-11+   **Cells:** 8
+- **Models:** 14 families (5 real + 9 PARC noise-seed baselines)
+- **Rows:** 1232 (family x checkpoint x task x session)
+- **Generated:** 2026-08-29
 
-Every alignment number in this dataset is only as meaningful as the brain
-RDMs it was computed against. So before any model result, the same
-pipeline is asked whether *anything* stimulus-driven correlates with those
-RDMs — stimulus duration, intensity, word length, frequency, phoneme and
-syllable counts, an acoustic model of the audio where the stimuli are
-audio, and the study's own condition contrast — each tested by a
-permutation test that shuffles stimulus identity.
+## Headline: no model is distinguishable from a random seed
 
-**GATE: FAILED. 0/8 stimulus tests are significant** after Holm
-correction — not the acoustic model of the audio the children actually
-heard, not the study's own experimental contrast.
+Alignment is computed as Spearman correlation over the upper triangle of the
+model RDM (`1 - corrcoef` over mean-pooled final-layer states) against the brain
+RDM, reported raw (`rsa`) and as a fraction of the noise ceiling.
 
-**The alignment numbers below are therefore uninterpretable as
-evidence about language models.** They measure a representational
-geometry that does not demonstrably encode the stimuli. They are
-published for completeness and for whoever fixes the estimator, not as
-a result. Do not cite them as evidence that models fail to align with
-the developing brain.
+Every value is referenced against a **PARC noise-seed null** (9 randomly
+initialised seeds across 3 architectures, same cells, same pipeline) rather than
+against zero -- see `null_referenced.csv` for per-cell z-scores.
 
-Measured cause, from `control/`:
+Across all three datasets and 130 (family x cell) combinations:
 
-- RDM effective rank: **52** of 72 stimuli
-
-
-Note that this is NOT ds003604's failure mode. There, the RDM
-effective rank was ~3 of 40-48 stimuli -- near-degenerate betas
-that could not express stimulus-level structure at all. The rank
-recorded above is a large fraction of the stimulus count, so these
-RDMs do carry stimulus structure and the control failing here means
-the specific controls tested did not reach significance, not that
-the measurement is uninterpretable. Check `control/` for which
-controls ran: an acoustic or visual control needs the dataset's
-stimulus files present, and reports zero features if they are not.
-
-## What was built
-
-8 task × session cells, each an RDM over the stimuli
-shared by that cell's subjects, with voxel patterns z-scored **within
-run** before aggregation (without that, the RDM measures scanner drift
-rather than language) and an inter-subject noise ceiling.
-
-| task     | session   |   n_stim |   ceiling_lower |   ceiling_upper |   ceiling_n |
-|:---------|:----------|---------:|----------------:|----------------:|------------:|
-| Orth     | ses-11+   |       96 |        0.562195 |        0.61485  |          22 |
-| Orth     | ses-11    |       96 |        0.522525 |        0.589811 |          18 |
-| Phon     | ses-11+   |       96 |        0.562195 |        0.61485  |          22 |
-| Phon     | ses-11    |       96 |        0.522525 |        0.589811 |          18 |
-| Sem      | ses-11+   |       48 |        0.356634 |        0.438561 |          22 |
-| Sem      | ses-11    |       48 |        0.289344 |        0.394831 |          18 |
-| SemLocal | ses-11+   |       48 |        0.305571 |        0.38963  |          23 |
-| SemLocal | ses-11    |       48 |        0.230569 |        0.364206 |          15 |
-
-## Dataset-specific notes
-
-Contains **LocalSem**, the only genuinely run/stimulus-CROSSED language cell across all four datasets in this project: its stimuli recur across runs, so run identity and stimulus identity are separable and the scanner-run confound that invalidated the first ds003604 analysis cannot arise. Per-subject age is NOT recoverable from the release — participants.tsv has birthdate but no scan date and there are no *_scans.tsv files — so this dataset is cohort-level only and cannot carry the developmental axis as published.
+| statistic | value |
+|---|---|
+| cells where a real family beats all 9 noise seeds | **9 / 130** |
+| expected under the null (P = 1/10 per cell) | **13.0** |
+| binomial p | **0.91** |
+| per-family mean z | all within +/-1.4 sigma |
+| per-cell means, real families vs noise seeds | **r = +0.863** |
+| variance explained by cell identity | **83.9%** |
+| variance explained by model family | **3.2%** |
+
+Real models beat the noise seeds *less* often than chance. The only reliable
+structure in these numbers is the stimulus set and the RDM, not the model.
+
+Scale trend for ds006239: Spearman(params, mean RSA) = **-0.40** (p = 0.50),
+i.e. a 16x parameter increase buys nothing.
+
+## What this does and does not license
+
+The RDMs have demonstrated inter-subject reliability (noise ceilings 0.23-0.88),
+but the upstream pipeline's **positive controls fail on all three datasets**
+(0/108, 0/6 and 0/8 stimulus controls significant). The instrument has therefore
+not been shown to have power against alignment that does exist.
+
+The defensible claim is **"no LM alignment is detectable by this measurement"** --
+not "language models do not align with the developing brain." This is a result
+about the benchmark.
+
+## Coverage correction
+
+The previously published grid did not pass `--sessions`, so it fell back to
+ds003604's `ses-5/7/9`. ds002236 matched only `ses-9` (2 of 6 cells) and
+**ds006239 matched nothing, producing zero alignment rows**. Deriving sessions
+from the RDM tree takes coverage from 14 to **26 cells**, so this release
+contains the first measurement of 12 previously unscored cells -- including
+`ds006239/SemLocal`, the only run x stimulus *crossed* cell, where the
+scanner-run confound cannot arise.
+
+On SemLocal, untrained (step-0) alignment falls **inside** the random-seed band on
+both sessions (z = -0.84, -0.30). The "untrained models align better, training
+destroys it" reading is **not supported**: the decline is drift within noise, and
+there was no alignment to destroy.
 
 ## Files
 
-| path | what |
+| file | contents |
 |---|---|
-| `alignment_by_checkpoint.csv` | every model × checkpoint × cell, with ceiling |
-| `alignment_by_family.csv` | per family, with equivalence tests |
-| `alignment_by_cell.csv` | per task × session |
-| `ceilings_*.csv` | noise ceiling per cell |
-| `control/` | the positive control and RDM dimensionality — the gate |
-| `scale_ladder.csv` | the Pythia 70M→1.4B scale test |
-| `fig_*.pdf`, `fig_*.png` | figures |
-
-## Method
-
-Representational similarity analysis. For each cell, a brain RDM over
-stimuli (correlation distance between per-stimulus GLM beta patterns,
-within-run z-scored, aggregated across subjects) is compared by Spearman
-correlation with a model RDM over the same stimuli, taken from each
-checkpoint's hidden states. Alignment is reported raw and as a fraction of
-the inter-subject noise ceiling, and judged against a null built from the
-PARC suite — 18 models differing only by random seed, which is what 'no
-effect' looks like on this measurement.
-
-Null and fixation trials are excluded from the stimulus set. For paired
-designs the stimulus identity is the pair, not either word alone.
+| `alignment_rows.csv` | one row per family x checkpoint x task x session |
+| `null_referenced.csv` | per-cell z-score against the 9-seed PARC null |
+| `scaling_curve.csv` | per-cell mean/sd/max across checkpoints, + noise ceiling |
+| `null_summary.csv` | per-family mean z, max z, cells beating the null |
+| `scale_trend.csv` | Spearman(params, RSA) per dataset |
+
+Precision: fp32 throughout. A bf16 spot check shifted RSA by up to 2.8e-3 at the
+final checkpoint (~1/3 of the across-seed noise sd), so precision is not mixed.