Releases: h3ro-dev/eegt
Release list
EEGT v0.11.0 — Conditional timing controls and replication capacity
This exploratory supplement tests how much original-versus-phase landmark matching remains after event timing is randomized within one-second bins while retaining local counts. Across six previously exposed people and twelve existing blocks, mean block raw F1 is 0.8477986, the mean block-null median is 0.6634331, and the equal-person mean difference is 0.1843655 at 25 ms tolerance.
The result is descriptive. This null also breaks fine spacing, synchrony and waveform constraints, so the gap does not establish neural origin, semantic meaning or an audio-interference mechanism. The analysis adds no people, recordings or source hours, and does not change earlier Experiment 012/013 endpoints.
The metadata-only EESM19 audit identifies nine remaining ear-only candidate people (108 sessions), or nineteen candidates using ear contacts in PSG sessions (76 sessions). These are unqualified metadata upper bounds. Nine is an attainable significance floor for the prior two-sided exact participant sign-flip test with eight-test Bonferroni correction, not an adequate-power conclusion.
The archive has 98 explicitly selected files, including public derived events, unchanged source receipts, the original scientific runner, and a relocation-only replay adapter. It excludes raw waves, model weights, private EEG/media and vendor correspondence. Offline replay exactly reproduced all twelve scientific rows, all 99 timing draws at each of three tolerances, the person-level summary and the metadata capacity audit. Independent scientific and release reviews accompany the results. Historical candidate/draft labels in frozen receipts are retained deliberately; later review and approval receipts establish acceptance.
Verify the archive and companions against eegt-v0.11.0-SHA256SUMS.txt, then follow the reproduction guide. The bounded replay used 197.085428 CPU seconds and 477,675,520 bytes peak RSS on the recorded runtime. No DOI has been issued for this supplement.
EEGT v0.10.0 — Consolidated evidence and reproducible catalog
EEGT consolidated research catalog
This release consolidates the accepted numerical evidence through Experiment 013: 89 candidate recordings, 88 qualified recordings and 476.053296 qualified source hours, across four source datasets. People remain source-local; experiment joins add no recordings or hours. One quarantined record remains visible with unknown duration.
Experiment 013 retains all eight prespecified endpoints. Four new-session comparisons use six complete people and remain inconclusive after correction. The four new-person comparisons are not estimable: two complete people fail the fixed minimum of three. No semantic, diagnostic, universal-vocabulary or physical Neurable equivalence claim follows.
The small archive contains the SQLite catalog, all 36 hash-bound curator inputs, exact original endpoint receipts, builder, tests, SQL examples and synthesis. Large waveform/model/event archives remain in prior releases. Offline rebuilding runs no model or detector. The independent review and acceptance files accompany the immutable candidate snapshot; its historical candidate labels are preserved deliberately.
Verify every checksum, then follow REPRODUCE-CATALOG.md. Compare logical content and all substantive summary fields when SQLite runtime versions produce different file bytes. DOI metadata is prepared but not deposited; no DOI has been issued.
Experiment 013 — fixed validation on new recordings
Fixed-method validation on 20 pinned recordings, with separate new-person and new-session cohorts. Independent scientific output review accepted the complete extracted recorded-ledger replay and morphology checks.
The numerical record contains 9,600 candidate blocks, 4,117 quality passes, 174 selected blocks and 4,002 event partitions. CodeBrain and CBraMod each ran 650 forwards on 130 eligible new-session blocks. Six complete new-session participants support four comparisons; adjusted p values are 1.00, 0.25, 0.50 and 1.00. The new-person cohort has only two complete participants, below the frozen minimum; all four of its primary comparisons remain not estimable. A separate descriptive check covers the first 30 seconds from each of two 12-contact EESM19 recordings.
Download every data shard, the manifest and SHA256SUMS. The complete captured packet includes selected numeric waves, model outputs, event partitions, the SQLite index, frozen source/protocol, resource receipts and preserved failures. Full raw recordings and two pinned upstream checkpoint files are not redistributed; the checkpoints remain declared hash-checked verifier dependencies even for recorded-ledger replay. Follow REPRODUCE-013.md in the versioned source.
Recorded-ledger replay recomputes comparisons and the scientific summary; it does not regenerate detectors or repeat neural inference. Calibration, model-pretraining overlap and physical Neurable equivalence remain UNKNOWN. Numeric patterns and encoder agreement do not establish a universal vocabulary, semantic meaning or diagnostic performance.
The unchanged original replay reached its fixed CPU ceiling without a complete receipt. That failed attempt is retained. The independently reviewed single-payload-pass verification adapter completed all 174 blocks / 4,002 partitions, reproduced the scientific summary, and stayed within the original limits: 5,306.66 CPU seconds and 1,447,821,312 bytes peak RSS. The separate morphology check reproduced 7,656 rows. No frozen scientific function changed.
The five data shards are the immutable captured candidate snapshot. Also download eegt-v0.9.0-acceptance.tar.gz, its companion manifest and SHA256SUMS for accepted review/replay/extraction proofs, preserved failure evidence and the reviewed additive replay source. Follow the current versioned REPRODUCE-013.md rather than the earlier instructions captured inside the data shards. Numerical acceptance does not establish semantic meaning, diagnosis or physical-device equivalence.
Experiment 012: direct waveform transitions
Experiment 012: direct waveform transitions
This release measures numerical waveform landmarks and their relation to two fixed EEG encoder representations. It does not establish universal tokens, diagnose brain states or decode thought.
The frozen census is unchanged: 5,760 candidate blocks from 48 previously exposed recording hours; 2,973 quality passes; 122 time-selected thirty-second blocks. Five paired people and 110 primary blocks contribute to each endpoint. The four adjusted event-to-encoder tests are inconclusive (CodeBrain geometry .50, change .25; CBraMod geometry 1, change 1). Phase-randomized controls retain high landmark matching, which limits interpretation. Undefined passband comparisons and every failure remain visible.
Download all three data shards, the data manifest and SHA256SUMS. Follow REPRODUCE-012.md in this tag for checked extraction and offline replay. The package includes selected numeric waves/masks, archived model vectors, all 2,806 event partitions, the index, methods and review evidence. No new model forwards or paid API calls were made per window. Runtime dollar cost is UNKNOWN.
The corrected candidate passed complete extraction and scientific-ledger replay. Independent final release review is accepted with all six findings resolved. Final archive extraction and public download/page verification are recorded separately in the publication acceptance receipt.
Experiment 011 · Two pretrained EEG encoders
Research Note 011 compares two independently pretrained continuous EEG encoders on identical public around-ear EEG inputs.
- 122 selected 30-second blocks; five paired people and 110 primary blocks.
- 610 new CBraMod forwards and 610 reused CodeBrain outputs, with exact input/output hashes.
- Mean paired original-minus-phase contrast: geometry 0.0713; change profile 0.0947. Both adjusted p values are 0.125. These findings are inconclusive and do not establish universal tokens or semantic brain meaning.
- No additional source people or recorded hours. Reserved people and later sessions remain untouched.
The data archive includes the captured source/input layout, selected numerical waves, both latent archives, full candidate and analysis ledgers, frozen protocol, retained corrections, and independent review. Full raw recordings and model weights remain upstream. Extract into a new directory and follow the included README; offline reproduction requires no model call.
Independent review accepted every shipped source/report/data hash. A fresh final extraction reproduced scientific values and SQLite table contents; file headers can vary with SQLite versions. Verify the assets against SHA256SUMS-v0.7.0.txt.
EEGT v0.6.0 — time-distributed pretrained geometry
Experiment 010 extends a frozen quality census across the already examined first four hours of twelve EESM23 recordings. Of 5,760 thirty-second candidates, 2,973 qualify, 122 are selected by a fixed time rule, and the pretrained CodeBrain backbone completes 610 original/control forwards. Five people and 110 selected blocks support the paired estimates. No new source recordings or people were acquired.
Geometry correlations with waveform shape, spectrum and sensor coordination are weak (0.0815, 0.0381, 0.0419), each adjusted p = 0.003 under the prescribed conditional cyclic-shift tests. Change-profile tests do not clear adjustment (0.387, 1.000, 0.639). These results do not establish matching inflection points, universal tokens, brain meaning or physical Neurable equivalence.
The archive contains all candidate outcomes, 122 prepared numeric inputs, 610 embeddings, aligned frozen descriptors, SQLite comparisons, source/code/checkpoint receipts, reproduction scripts and review evidence. Source EEG is open CC0; project code is MIT and vendored model source retains Apache-2.0. The full original recordings and pretrained weights remain upstream. Raw-to-prepared reproduction additionally needs the v0.4.0 feature bundle and its pinned raw sources.
Reproduction and verification commands are in README.md. Independent review accepted the 93-file source snapshot and recomputed 22,354 numerical assertions. Root reproduced the summary, page and plot byte for byte from the extracted candidate bundle; 66 tests pass with one optional encoder-suite skip. The reviewer did not rerun the full raw preparation or model forwards. The publication receipt separately records final extracted-bundle and public byte verification. Native local tests are distinct from GitHub Actions, which is organization-disabled. EESM23 participants 007–010 and unexamined later sessions remain reserved.
EEGT v0.5.0 · Pretrained encoder feasibility
Experiment 009 adds a pinned CodeBrain EEGSSM encoder to EEGT's public research ledger. It preserves the frozen selection and its insufficient-data result.
- 240 candidate 30-second segments from two hours of already exposed ear EEG; 9 segments pass the common quality rules (3.75%, or 4.5 minutes).
- 45 model forward passes: original waves plus four matched waveform controls.
- Zero complete eligible participant pairs. All six primary comparisons report INSUFFICIENT_PARTICIPANTS; no participant-level p values are reported.
- All candidates, exclusions, 27 segment/view comparisons, 36 control comparisons, wave derivatives, embeddings and reproducible code are retained.
The model is a continuous EEG encoder backbone, not a full discrete tokenizer or LLM decoder. Its scalp-to-ear spatial transfer and pretraining overlap remain unvalidated. This release does not establish universal tokens, clinical meaning or physical Neurable compatibility. The source reports 50 Hz mains; the frozen recipe's 60 Hz notch does not specifically remove that component.
Independent review reproduced all 54 segment correlations and 1,431 shifted-time values. A fresh forward agrees within 1e-5 but is not byte-identical. General tests, separate encoder tests, release source bindings and extracted-bundle reproduction are recorded with the release. SHA256SUMS identifies expected download bytes. GitHub Actions is disabled by the organization and is not claimed as passing CI.
Read the Research Note, frozen protocol and reproduction instructions.
The data archive includes nine derived public-source wave blocks, aligned descriptors and all encoder outputs. Full original recordings and pretrained weights remain upstream. SHA256SUMS and the manifest identify the exact release bytes. Earlier releases remain unchanged.
EEGT v0.4.0 · Repeated-night ear EEG
EEGT v0.4.0 adds a first repeated-night test to the raw ear-EEG research database.
- 12 new open EESM23 recordings, six participants and two nights each; 86.38 qualified recorded hours.
- Frozen analysis of the first four hours of each file: 48 hours, 345,564 multichannel windows, 306,392 QC-pass windows, and 47,476 transition events.
- Expanded queryable source index: 67 candidates, 66 qualified recordings, 309.85 qualified hours across three datasets; the prior quarantine remains visible.
- Same-person nights were closer on average in all three numerical views. Adjusted pairing p-values were 0.0917 for shape, 0.2833 for spectrum and 0.5542 for coordination. This six-person comparison does not establish universal tokens or semantic meaning.
- Four participants and later sessions remain reserved. No labels entered numeric discovery, no pretrained checkpoint inference ran, and no physical Neurable input was admitted.
- Official CodeBrain/REVE checkpoint compatibility audit, frozen protocols, source hashes, feature arrays, SQLite evidence and an independently reviewed Research Note.
See Experiment 008 and the source note. Original CC0 EEG remains at OpenNeuro ds005178 v1.0.0. Verify the release manifest and SHA256SUMS before extracting the derivative bundle. Earlier experiments and releases remain unchanged.
EEGT v0.3.0 — Continuous EEG and transition geometry
EEGT now includes the full pinned around-ear corpus: 55 source recordings, 54 qualified, and 223.47 decoded sample-hours. The run analyzed 1,608,671 multichannel windows; 1,077,399 passed the engineering quality gate. One source remains quarantined and visible.
Three numerical views locate partly overlapping transitions. On 16 previously unused external records, median timing F1 is 0.487 for shape/spectrum, 0.245 for shape/coordination and 0.207 for spectrum/coordination. These are timing agreements, not diagnostic accuracy or proof of universal geometry. Phase and nuisance controls, all selected scales, per-record results and five geometry descriptors are retained.
Read the results · Research notes and interpretation · Reproduce · Next priorities
Assets:
eegt-v0.3.0-database.tar.gz: source and transition SQLite databases, source/protocol/model/control/transfer receipts, per-record results and notes.eegt-v0.3.0-features.tar.gz: 54 numeric feature files and their binding receipt. Extract both archives into the source checkout to inspect the full derivative database.eegt-v0.3.0-database-manifest.jsonandSHA256SUMS-v0.3.0.txt: per-file and archive identities.
Original 21.54-GB raw files remain at the public archives with a pinned, checksum-verifying downloader. Code is MIT; the derivatives of these CC0 datasets retain upstream attribution. No private recordings are included.
Validation: 44 tests and 2,412 numerical/source-join audit checks pass locally and in a separate review environment. GitHub Actions did not run because the organization disables it. Physical Neurable collection, independent pretrained-LLM discovery, repeated-session reliability and clinical/semantic claims remain unverified. Earlier releases are preserved.
EEGT v0.2.0 — Experiment 002
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.