Skip to content

EEGT v0.2.0 — Experiment 002

Choose a tag to compare

View all tags
@CryptoJym CryptoJym released this 25 Sep 02:08
· 31 commits to main since this release
v0.2.0
77f0f11

EEGT v0.2.0 — Experiment 002

This release publishes a reproducible numerical pilot on public around-ear EEG:

  • 8 fixed recordings from two OpenNeuro CC0 cEEGrid collections
  • 40,500 two-second channel windows; 40,288 pass engineering QC
  • 3 representation families × 3 vocabulary sizes × 3 seeds = 27 fits
  • native slices, analysis windows, assignments, numeric codebooks, checksums and machine-readable results
  • a separate copied run with byte-identical model and assignment archives and identical numerical fields

The main result is disagreement between representation families. On the external source at K=16/seed17, spectrum/time-frequency ARI is 0.2894, spectrum/waveform is 0.0387, and waveform/time-frequency is 0.0189 on the same 7,938 QC-passing windows. High coverage does not establish a universal vocabulary, semantic meaning or clinical value.

Independent review of the prepublication packet found and documented a high-severity input-alignment defect: a same-length reordered evaluation index was accepted. The release includes the repair, explicit window IDs, digest-bound prepared inputs, regression tests, regenerated results, and repair-equivalence.json showing the numerical result is unchanged. The original finding is retained in protocol/amendment-002.md.

The integration acceptance receipt records 12 passing package tests, a 3,218-check regenerated-artifact audit with zero failures, a copied-run reproduction, and the exact review boundary. A fresh provider repair-review turn was attempted twice and was unavailable due to serverOverloaded; the release does not claim an independent PASS for that repaired turn.

These are LLM-authored numerical algorithms, not evidence that pretrained LLMs independently discovered a brain language. The recordings are cEEGrid data, not Neurable captures; the 12-channel/500 Hz adapter is an input-boundary test, not physical headset validation. Semantic and clinical studies remain deferred.

Code is MIT. Derived numeric artifacts from the declared CC0 sources are CC0-1.0 with upstream attribution retained. See the Research Notes site, DATA_CARD.md, and results/002/integration-acceptance.json for scope, provenance and limits.

Assets 20
Loading