The open register in a new dated file of 535 rows: 487 dataset rows and 48 model rows. The four datasets checked end to end on the data-quality page become this site’s rows at step 4 of the dataset ladder, each with its MOABB record folded in: Kim beta-range SSVEP and MIND with their results published, and Damm finger motor imagery and NeBULA with their scores held. Three models scored on the pilot-studies page enter the model ladder at step 4: NeuroRVQ, EEG-Captioner and AXON. Five more models are identified from their papers at step 1, each with its hold: BandVQ, EEG-PRIME, FAME, STST-JEPA and USE-FM. open-register-20261020.json stays served byte for byte.
Notes Kim beta-range SSVEP: its row links its dataset page and its case on the data-quality page, and holds models whose output depends on amplitude, in place of the onset-timing hold its MOABB record carried. MIND: both licences side by side (CC BY 4.0, CC BY-NC 4.0), with original terms and pinned version as gaps at step 3. Damm finger motor imagery: rights review and consent or ethics statement as gaps at step 3. NeBULA: its Figshare record at its only version, with no gap at step 3.
Notes NeuroRVQ, EEG-Captioner and AXON: each links its panel on the pilot-studies page and has no method page; pretraining exposure is unknown for all three. Terms by part: NeuroRVQ: code CC BY-NC 4.0, weights CC BY-NC 4.0; EEG-Captioner: code MIT, weights Apache-2.0; AXON: module code Apache-2.0, project code BSD-3-Clause, weights Apache-2.0. BandVQ, EEG-PRIME, FAME, STST-JEPA and USE-FM: each links its paper; EEG-PRIME is held because its code repository could not be reached, and the other four because no official release of code or weights was found. Not published: the register use statements drafted for some rows, which stay internal, and third-party conversions of NeuroRVQ’s checkpoint.
Seven more follow-ups on SSVEP frequencies a decoder never trained on, as cards 8 to 14 of the same question page, each read from its own audited aggregate: the cross-spectral round’s results, which take the place of card 8’s status entry; eigenvalue pooling; harmonic pooling, on BETA and on Wearable-102; filter-bank CCA with shorter windows and fewer electrodes; the gap between harmonics; a fixed decay over harmonics; and rising harmonic weights. All seven use route 3’s BETA cohort (70 people), and four of them also run on Wearable-102’s dry and wet headbands (102 people). Seen and held-out frequencies are separate tasks, reported side by side, and each set-up sits beside its untrained control. Each card carries its own interval method: card 12’s primary family is Bonferroni-adjusted over 9 comparisons, card 13’s margin test uses a 90% interval, and no other interval is adjusted. Card 11 adds CPU timings: warm, single-thread measurements of each component, never summed, and no sensor-to-decision latency.
Notes On the held-out 8-way task with 2-second windows and 8 electrodes, filter-bank CCA with no training reached 90.21% against 80.90% for CCA on the same windows (+9.31 points, +6.84 to +11.71; 70 people). The learned eigenvalue head differs from the harmonic sum, which needs no training, by +0.07 points (−0.45 to +0.62). Route 3’s questions-in-language page runs another filter-bank recipe on the same task, with 5 Butterworth sub-bands.
Notes BETA keeps its released record: CC BY 4.0, cohort aggregates only, with participant consent described for noncommercial scientific research. Wearable-102 keeps its released record: CC BY 4.0, cohort aggregates only. Filter-bank CCA is credited to Chen et al. 2015; the code that ran is the project’s own SciPy realisation, and no code or weights are distributed. This file’s card 8 replaces the status entry that new-ssvep-frequencies-update.json carries for it; that file keeps its bytes.
Notes Not published: per-person, per-fold, per-rotation, per-seed, per-window and per-frequency values of any kind, predictions, scores, features, fitted heads and the sealed run files; the cross-spectral round’s per-seed, canary and random-pairing accuracies; the fixed-decay retrospective ratio table; all-trial decision counts beyond the harmonic-gap shares, and that card’s Delta means and distractor margins; run times of whole evaluations and audits, memory, CPU-time readings and every host, processor, system and runtime label; the comment thread’s predictions on these cards; the commenter’s name; and the handoffs, protocols, freezes, receipts, reviews and approvals themselves, pinned by hash in the release manifest because several carry private storage paths.
A separate panel on the pilot-studies page, outside the pilot count and every leaderboard: NeuroRVQ’s released backbone, kept frozen and read by one fixed ridge head, on three existing tasks: EEGMAT’s arithmetic and rest (36 people, 2,160 two-second windows), motor imagery against rest on ds003810 (10 people, 1,200 windows) and P300 target against non-target on ds006593 (21 people, 2,520 one-second epochs), 5 participant-held-out folds each, nothing tuned. Each task sets the backbone’s 800 pooled values beside simple features of the same windows, with participant-mean balanced accuracy and macro-F1 for each, and one paired difference with its participant-bootstrap interval. Pointers on the three protocol pages lead to it and carry no figure; the core matrix keeps its rows.
Notes NeuroRVQ minus its comparator, in balanced-accuracy points: on EEGMAT +1.53 (−3.52 to +6.67) against band power: the interval spans zero, so neither is shown to be better, and the two are not shown to be equivalent; on motor imagery −15.25 (−20.17 to −10.25) and on P300 −3.57 (−5.56 to −1.51) against temporal means, both below zero. Pretraining exposure is unknown: none of the three datasets is named in the paper’s list of 13 pretraining corpora, 1 of them proprietary. There is no random-initialisation arm, so the panel says nothing about pretraining. The authors’ inference wrapper does not restore strictly (8 normalisation tensors missing), so what ran is the released backbone with parameter-free branch normalisation and mean-token pooling.
Notes NeuroRVQ’s code and weights are CC BY-NC 4.0, the licence file of both repositories at the pinned revisions; only aggregate figures are published, no weights are distributed, and no endorsement by NeuroRVQ’s authors is implied. EEGMAT, ds003810 and ds006593 keep their released records. Not run: semantic-target (the original calibration evidence is still missing); beta-4ch (physical units and reference are unresolved); beta-8ch (physical units are unresolved, and PO5 and PO6 are missing from the model’s channel list); idle (its inputs are missing, and it needs its own calibration, threshold and split protocol); sleep-scalp (deferred to a separate study). Not published: per-person, per-fold and per-window values of any kind, fold assignments, identifiers, labels, extracted values, fitted heads, predictions and the sealed files; how many people each arm scored above 50%; per-arm intervals and macro-F1 differences, none of which was computed; measured compute; the checkpoint and the authors’ source; third-party conversions of the checkpoint and the licence tags they carry; and the handoffs, protocols, configurations, receipts and reviews themselves, pinned by hash in the release manifest because several carry private storage paths.
Another separate panel on the pilot-studies page, outside the pilot count and every leaderboard: two frozen encoders on ten-second EEGMAT windows of arithmetic and rest (36 people, 72 recordings, 432 windows, 5 subject-disjoint folds), each read by one fixed ridge head and set against Welch band power on the same windows, in a subsection of its own: EEG-Captioner’s encoder (256 dimensions; only the encoder ran, with no caption) and AXON (512 dimensions). Participant-mean balanced accuracy and a paired difference with its participant-bootstrap interval for each; beside every score, 0 of 432 windows meet EEG-Captioner’s source annotation-QC rules. No difference between the two encoders is computed, and the panel is kept apart from the two-second results on this site.
Notes EEG-Captioner’s encoder reached 65.74% against 67.59% for Welch band power fitted in the same run (−1.85 points, −9.03 to +4.86). AXON reached 70.14% against the same band-power predictions, reused without refitting and matched for neither normalisation nor parameter count (+2.55 points, −5.79 to +10.42). Both intervals span zero. AXON’s own numerical input check passed 432 of 432 windows, a different check from the source QC, which flags whole recordings (72 of 72) and does not say that each window has an artifact. AXON ran after the EEG-Captioner and Welch outcomes were known; rest is recorded before arithmetic; pretraining exposure is unknown for both encoders.
Notes EEGMAT keeps its released record: Open Data Commons Attribution License 1.0, aggregate results only. EEG-Captioner’s source code is MIT. EEG-Captioner’s encoder weights are Apache-2.0 as its Hugging Face weight metadata states. AXON’s checkpoint is Apache-2.0 as its Hugging Face licence tag states; its Braindecode module carries per-file Apache-2.0 terms within a BSD-3-Clause project. No weights are distributed, and no endorsement by either model’s authors is implied. Not published: per-person, per-fold and per-window values of any kind, fold assignments, identifiers, features, fitted heads, predictions and the sealed files; confusion matrices and pooled macro-F1; the source-QC breakdown beyond its totals; forward counts, head-fit counts, CPU and wall time and memory; a difference between the two encoders; AXON’s paper reference, until it is checked; the checkpoints and weights, whose SHA-256 identifies each file and grants nothing; and the handoffs, protocols, approvals, reviews, audits and receipts themselves, pinned by hash in the release manifest because several carry private storage paths.
Four more public EEG datasets checked end to end on the data-quality page, each read from its own files. Kim beta-range SSVEP is a worked case: training-free CCA on all 40 participants (9,600 trials, 40 targets from 14.0 to 21.8 Hz), with a dataset page of its own. MIND is a conditional case on record IDs: CSP + LDA on two stored-code pairs beside an ordinal lookup and a majority control, with record-ID resampling ranges and the facts of its label schedules. Damm finger motor imagery and NeBULA are one-person status cases: design, counts and file facts, with every score held. The one-person rule is stated once on the page, before eldBETA, Damm and NeBULA.
Notes Kim: mean participant accuracy 53.40% (45.89% to 61.01%), against chance of one in 40; the amplitudes are as stored, and models whose output depends on amplitude are held. MIND, primary panel of 28 record IDs and 3,360 trials: CSP + LDA 56.96% on stored codes 4/5 and 51.85% on 6/7, against 50.00% for the majority control; the ordinal lookup sits below the majority control because each pair follows one of only two complementary 20-trial label schedules. Record IDs are not verified people. Damm: 150 delivered cues over five fingers and 62 scalp channels. NeBULA: 89 delivered trials, 127 EEG channels and 11 EMG rows where the metadata states 16.
Notes Licences, from each record: Kim CC BY 4.0 (Figshare 28806815, version 3, which the research used; the paper cites version 2, also CC BY 4.0); MIND CC BY 4.0 and CC BY-NC 4.0 side by side, as the paper and the archive’s dataset description state the first and the MOABB adapter and the ScienceDB record’s metadata the second, with ScienceDB’s terms-of-use page not read; Damm CC0 as its dataset description states, while its README cites a LICENSE file the tree does not hold; NeBULA CC BY 4.0 (Figshare 27301629), the paper’s own licence covering the article only. Kim’s, MIND’s and NeBULA’s papers state ethics approval and consent, MIND’s without naming its committee; Damm’s dataset description names its ethics approval, and no consent statement was found. Not published: values per person, per record ID, per fold, per run or per trial, identifiers, features, fitted models, predictions and raw signals; every Damm and NeBULA score and any comparison of NeBULA’s order diagnostic with its EEG or EMG arms; MIND’s pooled metrics, confusion matrices, secondary schedule census and model-fit counts; Kim’s per-participant accuracies; compute, hosts and storage locations; and the research aggregates, audits, reviews, protocols and handoffs themselves, pinned by hash in the release manifest because several carry private storage paths.
The open register: every EEG dataset we know of and every EEG model and paper we track, each with the furthest step we have checked and what that step showed, in 527 rows (487 dataset rows and 40 model rows). Two ladders with their own step names: a dataset is listed, read, traced, run here and in the core matrix; a model is identified, has its terms recorded and its files checked, is scored here and in the core matrix. Each row keeps the evidence for every step as pass, gap or not applicable, with its source, so its step is the furthest one whose evidence is on file and its gaps stay visible. The dataset rows come in five groups: uploads on Hugging Face that hold brain-signal data, from a frozen snapshot of 28 search terms and 20 tag filters read on 19 October 2026 at 14:23 UTC (1,922 listings, 737 of them EEGDash mirrors, and 368 other uploads in scope); OpenNeuro’s mirrors on EEGDash, as one family row; this site’s own datasets; the catalog of MOABB release 1.7.2, one row per source record (206 loaders over 158 records, 18 of them merged into this site’s rows); and corpora that require an agreement or request, with their access terms as captured from the source. Anything can be listed here; the score tables keep their own rules.
Notes The rules, stated once on the register: gated data, or data that needs a signed or requested agreement, stays at step 1 unless we accept its terms; suspected re-uploads stay at step 1 whatever their gating, compared by metadata only, with none of their files downloaded; a finding about a re-upload is stated only when an independent check confirms it and a captured page of the original terms backs it, and then in factual words: that an upload matches another source’s metadata, or that its declared terms differ from the original terms; rows state facts; and every row has a correction route, a mail that names the row’s id. The CSV carries no findings; they are in the JSON. Repository ids and usernames appear in a row’s text, never in a link or an anchor; an uploader’s series of repositories is one family row; card text, download counts and modification dates are not shown.
Notes The register now lists sources that give a reader nothing to run, such as L-FAME (CC BY-NC 4.0), because it records what was checked. Each later register version is a new dated file and release; this file stays served and listed.
Route 3’s follow-ups on SSVEP frequencies a decoder never trained on, on a new question page: seven audited studies that answer questions from a reader’s comments on Hugging Face. Six reuse route 3’s BETA cohort and splits (70 people, 7 participant folds by 5 frequency rotations, 32 seen and 8 held-out frequencies; trained heads over 3 seeds), each testing one idea: asking with the reference spectrum; the head, the spectrum or the question; a random delay on a phase-carrying query; a random phase ramp with magnitude queries; where CCA’s lead comes from; and a bounded learned correction on top of CCA. The seventh runs the correction’s models unchanged on Wearable-102’s dry and wet headbands (102 people). Each card gives its question, what changed, the matched control, the result and its limit; every interval is its own study’s, and no multiplicity correction was applied. A status card for the cross-spectral round carries no figures; that round ships as its own release.
Notes On frequencies left out of training, temporal CCA with a small learned correction reached 86.15% against 80.90% for CCA alone, choosing among 8 frequencies (+5.24 points, 3.91 to 6.62; 70 people). Run unchanged on Wearable-102, the same models changed CCA’s score by +0.07 points on dry electrodes (−0.71 to 0.91) and −0.61 on wet (−1.48 to 0.23). Learned heads that read reference spectra or stimulus phase stayed far below CCA.
Notes BETA keeps its released record: CC BY 4.0, cohort aggregates only, with participant consent described for noncommercial scientific research. Wearable-102 keeps its released record: CC BY 4.0, cohort aggregates only. The frozen CBraMod features are the ones route 3 used; no weights are distributed. The stimulus phases of one query come from the BETA authors’ corrigendum, read at a pinned revision of the Hugging Face mirror Bingchuan/BETA. The Wearable-102 documentation cannot rule out that some of its participants also appear in BETA.
Notes Not published: per-person, per-fold, per-rotation, per-seed, per-window and per-frequency values, predictions, scores, features, fitted heads and the sealed run files; the adapter’s per-seed accuracies and the transitions of the cropped-match correction; the Wearable-102 contrast against the cropped-spectrum check; the commenter’s name; measured compute; and the handoffs, protocols, freezes, receipts and reviews themselves, pinned by hash in the release manifest because several carry private storage paths.
A new question page: what do the files show once a public EEG dataset is checked end to end? Three public datasets, each read from its own files. Shin2022 is the worked case: a participant-held-out diagnostic over 10 people with three arms (a participant-weighted training prior, a normalised spectral ridge on the first second of cursor feedback, and a retrospective control that reads only the feedback duration), scored on the signal-valid windows that its coverage chain counts from the runs in the release. eldBETA and ZuCo 2.0 are status cases with no score: one participant’s eldBETA recordings, as counts, and a timing hold on one converted ZuCo 2.0 run.
Notes Shin2022: in the release, 9 pairs of runs are byte-identical, so the files hold 101 distinct runs, and one incomplete run is held whole. On the signal-valid windows the normalised spectral ridge reached 52.59% (51.51–53.66) participant-mean balanced accuracy and the duration control 52.33% (50.87–54.02), against 50.00% for the prior. The scores cover the signal-valid windows only: feedback lasting at least one second, and a first second that is finite with positive 1–40 Hz power on all 16 channels. The label is the instructed target in closed-loop cursor recordings, which carry visual, controller and behaviour confounds. The duration control reads the whole feedback duration, known only after the trial ends, so it checks a confound; the spectral arm sits close to it, and no test between arms was run. Ten participants, one recording day each; the intervals are descriptive. The catalog’s six-second trial length is the feedback cap (MaxFeedbackDuration) that every run’s header sets.
Notes eldBETA: the raw-recording route runs end to end on one participant’s 7 blocks: all 63 trials pass the reader and the quality checks, and none is dropped. The continuous recordings hold 248,320 non-finite values outside the analysed windows, left as recorded. With one person every score is that person’s, so the accuracy is held, by the owner’s decision. The header calibration was applied, and the derived MATLAB epochs were not used. ZuCo 2.0: in one converted run on NEMAR, some fixation times fall outside the recording, so no EEG–text score is computed until the converter’s timing rules and a source-checksummed fixture are available.
Notes Licences, from each record: Shin2022 CC BY 4.0 (Figshare 20383716), eldBETA CC BY 4.0 (Figshare 18032669), ZuCo 2.0 CC-BY-4.0 as its NEMAR dataset description states, with the original terms on OSF and the terms of the quoted text excerpts unverified. Shin2022’s paper names its approving board and states written informed consent; eldBETA’s data descriptor gives its board’s approval number and states written consent; for ZuCo 2.0 no consent or ethics statement was read in this release. Not published: participant rows of any kind and raw signals; Shin2022’s per-participant figures, feedback-duration distribution and per-channel census counts; eldBETA’s one-participant accuracy and its comparator; ZuCo 2.0’s converted-run code, timing values and marker counts; compute; and the research exports, audits, reviews, protocols and handoffs themselves, pinned by hash in the release manifest because several carry private storage paths. MOABB’s catalog counts for Shin2022 are findings on its row in the open register.
A new section on the clinical-groups page, with its own question: do EEG case-control results survive age and recording-setting controls? IESSEEG, a public infant EEG dataset (100 people: 50 cases of infantile epileptic spasms syndrome and 50 controls, one awake minute each), on the publisher’s 5 participant-disjoint folds: six arms with and without EEG, including an age-only control, a recording-settings-only control and EEG after age is regressed out inside each training fold, and four arms that read frozen NeuroLM features with the same probes. Out-of-fold point estimates of balanced accuracy and AUROC, both shown for every arm, for everyone and for the 68 people aged 2 or under. The page’s Parkinson’s question, short answer and card keep their wording, and IESSEEG adds no dataset page or count.
Notes Point estimates out of fold; no interval or test was specified. One awake recording per person, first 60 s. Age and recording settings differ between groups; regressing out age is a sensitivity analysis. The probes and every score were reproduced independently; the NeuroLM features themselves were not re-extracted by a second implementation. NeuroLM’s preprocessing differs from the spectral pipeline: the two 2,000 Hz recordings are resampled to 200 Hz, then a 0.1–75 Hz band-pass, 50 and 60 Hz notches and division by 100. A difference between the two pipelines therefore mixes preprocessing with architecture. The checkpoint’s training stage and its prior exposure to these participants are unknown.
Notes IESSEEG: CC BY 4.0 for the dataset and annotations, from the release card and LICENSE at revision 1c32ee3a of Capur/IESSEEG, cited by the title the card requests: “IESSEEG: An Open EEG Dataset Towards Better Understanding of Infantile Epileptic Spasms Syndrome”. The release card states institutional review board approval of the retrospective study, with waivers of informed consent and HIPAA research authorization, and asks that no one attempt to re-identify participants. The publication boundary is the one the Parkinson’s comparison uses: aggregate scores only, with no participant rows, predictions, ages or per-group age and sex composition; the age ≤2 column gives the band’s size and its case and control counts. The card declares that the 400 case segments were previously analysed by Rajaraman et al. (2024); the participant-level crosswalk to that analysis is unresolved. NeuroLM-B is used frozen and identified by its checkpoint’s SHA-256; no weights are distributed. Not published: participant rows of any kind, per-fold metrics, per-group age and sex composition, intervals and tests, signals, code, the checkpoint and its features, and runtimes and storage locations.
A separate panel on the pilot-studies page, outside the pilot count and every leaderboard: NeuroLM-B kept frozen on EEGMAT’s arithmetic and rest (36 people, 2,160 two-second windows, whole participants held out in 5 folds), with one fixed ridge head fitted on each of three representations of the same windows: NeuroLM’s tokenizer features, its GPT hidden states and band power. Equal-person balanced accuracy and macro-F1 for each, and three paired differences with paired participant-bootstrap intervals. A pointer on the Arithmetic & rest protocol page leads to it; the core matrix keeps its rows.
Notes Tokenizer features minus band power is +0.60 points (paired 95% participant-bootstrap interval −5.32 to +6.44): the interval spans zero, so neither is shown to be better, and the two are not shown to be equivalent. Pretraining exposure is unknown: EEGMAT is NeuroLM’s Workload task, one of the 6 instruction tasks in its pinned code, and this checkpoint’s training stage is unresolved. The pilot has its own adapter and band-power features, so its scores sit in their own panel; rest is recorded before the task, so recording order moves with the label.
Notes EEGMAT keeps its released record: Open Data Commons Attribution License 1.0, aggregate results only. NeuroLM’s code is MIT at the pinned commit. The NeuroLM-B checkpoint is CC BY 4.0 as the licence tag of its Hugging Face repository states, and its 30-byte model card adds no terms. No weights are distributed, and no endorsement by NeuroLM’s authors is implied. Not published: per-person, per-fold and per-window values of any kind, demographics, measured compute, the checkpoint and the NeuroLM source, and the handoff, protocol, producer, receipts and reviews themselves, pinned by hash in the release manifest because several carry private storage paths.
EEGPT on the core matrix: its eight core-matrix cells, computed in the same runs as the published LaBraM rows and, except on idle, the CBraMod rows, with only the encoder swapped, published as computed, without a re-run, as a frozen row beside the core matrix, where the further foundation encoders sit. On seven protocols, balanced accuracy with its participant-bootstrap interval, macro F1 and the chance level; on the idle protocol, counts (idle false activations, commands detected within 3 s, people who always abstained) with the window balanced accuracy and AUROC, and no interval. Beside the published LaBraM and CBraMod rows, EEGPT’s pretraining-attribution rows on mi-rest and EEGMAT: the pretrained encoder against the mean of three random initialisations, under a fixed and a train-selected readout, with paired intervals. Grouped by family, never ranked: EEGPT is above or below a published row only where their 95% intervals do not overlap, and the gap between the intervals is given in percentage points. Each cell says whether its dataset is in EEGPT’s published pretraining list, checked on 12 October 2026 with the source (none of the seven core datasets is), and each footnote sits beside the cells it bears on.
Notes Several gaps are narrow. On sleep staging EEGPT’s interval lies above every core-matrix row’s, the nearest by 0.10 pp; on P300 it lies above LaBraM’s by 0.28 pp, on the semantic target above the spectral ridge’s by 0.78 pp, and on BETA with four channels above LaBraM’s by 0.90 pp. On BETA, EEGPT lies below standard CCA, which needs no training, at both channel counts. The idle row comes from the same run as the published LaBraM idle row; the published CBraMod idle row comes from another run, so a comparison with it is across runs. Pretraining attribution: on both datasets and under both readouts the pretrained encoder scored above the mean of its random initialisations, with paired intervals above zero; a frozen-encoder result conditional on these splits, readouts and initialisations, not a model ranking.
Notes Owner decisions of 12 and 13 October 2026: publish EEGPT’s eight core-matrix cells as computed, without a re-run, as a frozen row beside the core matrix, and its pretraining-attribution rows beside the published LaBraM and CBraMod rows. The rights reasoning is the pilot comparison’s of 12 October: independently computed aggregate results are a separate output from the weights; no publisher prohibition was found; the checkpoint’s licence is unresolved; no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. No author endorsement is implied. Four of the eight cells (mi-rest, EEGMAT and both BETA protocols) and EEGPT’s attribution controls have been in audit records in the public repository since 20 September 2026; the public attribution audit still labels the four EEGPT contrasts “hold-unresolved-weight-license”, as it did then, and keeps its bytes. With this release they are published.
Notes How EEGPT differs from EEG-DINO: EEG-DINO stays catalogue-only by the owner’s decision of 4 October 2026 — its upstream weights carry no licence, and it was neither downloaded nor run. EEGPT was evaluated in the same runs as the published rows and was held only because its checkpoint’s licence could not be read, a hold the reassessment of 12 October found to be this project’s own choice. EEG-DINO is not re-reviewed by this decision.
Notes Not published: per-trial, per-person, per-fold and per-class values, features, fitted heads, idle thresholds and calibration grids, and how many people had a negative attribution difference; the per-seed scores of the three random initialisations (their mean is published, as for LaBraM and CBraMod); BETA’s duration-group summaries; timings, devices and memory; EEGPT results outside the eight core protocols; the checkpoint and its tensor conversion, whose SHA-256 identifies them and grants nothing; and the run summaries, protocols, independent audits and verification records themselves, pinned by hash in the release manifest because they carry private storage paths. Nothing was re-run: experiments.json, deployment-topics.json and the further-encoder files keep their bytes.
Small pilots on four sources, each from its own approved aggregate export, all kept outside every leaderboard and model ranking. The Orthosis error-detection dataset (8 people, 83 runs, 2,407 recorded device events: one cohort, counted once) in three results: a 64-channel classical EEG baseline, the same baseline on the 58 channels EEGPT supports, and frozen EEGPT with the same fixed supervised probe against those matched classical features. BINGO’s Greek and English inner speech (17 source-defined subject fields): a fixed spectral baseline across people and across languages, a negative result near its 15-class prior. A mask-aware EEG analysis of an imagined-handwriting competition dataset (5 people, four released coded labels): with channels that sit flat at a digital rail masked, relative EEG is slightly above the training-majority control and above the missingness-mask and mask-plus-time controls, and not separated from elapsed time and trial order; a data-quality and methods case study, not handwriting recognition. REFED (two nominal participant folders, 30 stored clips): a computational diagnostic in which simple source-only controls have smaller errors than EEG band power. Nothing is pooled across datasets, configurations or conditions.
Notes EEGPT: the comparison is published from its approved export, after an independent check of that exact file. The earlier hold on EEGPT aggregate scores was this project’s own conservative choice, not a publisher’s rule; this pilot comparison is the first EEGPT result on the site’s pages. The core-matrix EEGPT scores are not on the site’s pages, but some of them (mi-rest, EEGMAT and both BETA protocols, with EEGPT’s pretraining-attribution controls on mi-rest and EEGMAT) have been in audit records in the public repository since 20 September 2026. The checkpoint’s licence remains unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. The paired interval spans zero, so neither arm is shown to be better, and the two are not shown to be equivalent. The probe is trained on the other seven people’s labels (supervised transfer, not zero-shot), and overlap with EEGPT’s pretraining data has not been excluded. Corrected the same day: this note first said the pilot comparison was the only EEGPT figure published. The core-matrix cells followed on 13 October 2026, in eegpt-core-update-20261013.
Notes BINGO: the source states no licence. Consent is as the project’s public deliverable D3.1 (version 3.0) reports it: signed consent forms, stated for the project’s NATO-alphabet recordings in the same CERTH laboratory; the cross-language dataset’s own documentation has no ethics or consent statement, and no ethics-approval reference is published. The aggregates are published on that basis by the owner’s decision. The project documents artifact cleaning (ASR and FORCe) on the continuous recordings before trials were cut; that statement is printed beside the result, attributed to D3.1 (version 2.0), and was not re-verified against the file. The project’s own result on its NATO-alphabet dataset is printed as reported by the project, not run here.
Notes Handwriting: published by the owner’s decision with its rights gap stated. No ethics-approval or informed-consent statement was found, on the dataset card or in a participating team’s challenge paper, and the organisers’ paper is listed as forthcoming; to be rechecked when it appears. The earlier analysis of the same five people is printed as context, not as a second cohort.
Notes Not published: per-person, per-field, per-folder, per-clip and per-record values; the Orthosis response-clean subset’s scores (its counts are published); individual permutation and label-shuffle replicates (their mean and observed range are published); confusion matrices; the handwriting mask-only control’s interval, because with five people its bounds would show individual people’s scores; and the handoffs, approval records and independent audits themselves, pinned by hash in the review manifest because they carry private storage paths. Status only, with no figure: Orthosis EMG, TinyMyo and EEG–EMG fusion; handwriting fNIRS and fusion; REFED physiological and fusion evaluation; and a meditation dataset, in which RR means heartbeat RR intervals, not respiratory rate.
Locally executed analyses beyond scalp EEG, each its own species, signal and metric: none is compared with scalp EEG, none joins the core matrix, and they share no ranking. Human fNIRS (Bak et al., 30 people): executed right-hand, left-hand and foot tapping classified by a fixed shrinkage LDA in a person left out of training, against a training-prior comparator. One macaque’s motor-cortex spiking (MC_RTT, prepared for the Neural Latents Benchmark; one session): x/y finger velocity decoded by a fixed ridge on held-out trials of the same session, against the training mean, on five metrics, including the one on which the ridge is worse. A numerical coverage check of human fNIRS recordings (ds007420, 12 people, 60 recordings): counts only, no decoding. Two sources are prepared and not scored: FALCON H2 human intracortical handwriting and Rabut et al. human functional ultrasound.
Notes Bak: the participant-macro balanced accuracy is well above the training prior, with a descriptive interval that resamples fixed participant scores (not a refit bootstrap). Executed movement, not motor imagery; offline over a 15-second window, and systemic or motion effects are not excluded. MC_RTT: not an official Neural Latents Benchmark score; with one animal every figure is that animal’s, and no interval is reported. ds007420: coverage counts say nothing about signal quality, and no classifier score exists, because its task codes have no verified lookup and its event durations disagree.
Notes Licences, from each repository’s record: Bak CC BY 4.0, MC_RTT CC BY 4.0, ds007420 CC0, FALCON H2 CC BY 4.0, Rabut et al. CC BY-NC 4.0 (non-commercial). Bak’s ethics and consent statement was read from an Internet Archive capture of the publisher’s article page, because the publisher refused automated requests. MC_RTT is animal data collected under procedures approved by institutional animal care and use committees, so informed consent does not apply. Not published: per-person, per-recording, per-channel, per-trial and per-row values; confusion matrices; signals, features, predictions and coefficients; measured compute; demographics; and the handoff, reviews and independent audits themselves, pinned by hash in the review manifest because they carry private storage paths.
Route 3 of the decision-research roadmap, questions in language: can an EEG model be asked its questions in words, including ones it was never trained on? One small question-conditioned head over frozen features is asked each question by a question number, a label template or a natural-language description. SSVEP on BETA (a plain spectrum and frozen CBraMod features) and sleep on BOAS (frozen CBraMod features) are primary, three seeds each; EESM19, OpenBMI motor imagery, REVE-L features, a second text encoder, Chinese wordings, a negation probe, a contiguous held-out band of frequencies and Wearable-102 are secondary. Seen questions in their training wording, rewordings and questions the EEG head was never trained on are reported apart, never pooled into one zero-shot number, and every entry carries two flags: whether its interval excludes zero, and how it stands against a margin fixed before any result (2 pp for accuracy; a ratio of remaining error of 1.25 for the unseen sleep questions; the primary comparisons with the numeric code, the neighbour average and the shuffled templates are read for a difference only). “Jev-style” here means the interface of a decision model like TypeSafe’s Jev, applied to EEG: encode the recording once, then answer several explicit, typed questions about it, each with a probability. The roadmap started from a vision paper (Yu & Yao, 2026) that applies that idea to images.
Notes Mostly a negative, boundary result. On sleep, asking a seen question with a template or a description was equivalent to a question number within 2 pp; on SSVEP it cost accuracy with both kinds of features. A new sentence frame cost nothing measurable on sleep, while synonyms and new descriptions lowered accuracy on both datasets (on SSVEP with frozen CBraMod features the synonyms stayed within the margin). Questions the EEG head was never trained on were not answered well: in every wording, each unseen sleep question did worse than adding up the head’s own stage answers, and every comparison of an unseen flicker frequency with CCA, which needs no training, fell far short. Negated wordings were answered as if they asked for what they negate.
Notes Before any result, the pre-run permutation canary failed twice (revisions 3 and 4; every failing cell was on BOAS, one of them below chance). The owner overrode revision 4’s “no further revision” and approved revision 5, which passed; revision 4’s centring diagnostic was not run, and both failed records are kept and disclosed. For the Chinese text encoder the SSVEP token rule was reported instead of stopping the run (decision S0b-6); an engineering amendment before the first fit (EA-1) changed no question, partition, wording, arm, metric, margin or gate; the declared numeracy prediction was not activated, because its premise did not hold. There are 35 pre-declared primary comparisons, not corrected for multiplicity, and most secondary results are single-seed and were declared likely inconclusive before the run.
Notes BOAS figures are published under the owner’s approval of 7 October 2026, extended to route 3 the same day (D11): with its three stated gaps and its attribution, participants pseudonymised in the public release, and every BOAS cell pooling at least 20 people; neither Bitbrain’s headband nor its automatic scoring is used or evaluated. Not published: per-person, per-night and per-fold values; the S10 per-frequency and interior-only breakdowns, which no independent audit covers; any value of a pre-run check, which is published as pass or fail and counts only; measured compute; and the handoff, the aggregates, the protocol and the four independent audits themselves, pinned by hash in the review manifest because they carry private storage paths. The wording lists, the held-out partition and the derangements of the shuffled control are published in their own file.
Four results under one question — does a decoder trained on an earlier session still work later? — each from its own approved release, and none uses labels from the later session. WBCIC-SHU (62 people in two separate cohorts, 51 with two-class and 11 with three-class motor imagery): a source-majority prior and a relative spectral ridge trained on each person’s recording session 1 and tested on session 3, and frozen CBraMod with a ridge readout fitted on session 1, on the same people and trials. A longitudinal RSVP dataset (15 people): one ERP baseline trained at the first visit and scored at the publisher’s nominal Day 7, 80 and 200 visits. Forenzo’s continuous cursor-tracking dataset (23 admitted records in two cohorts, two response arms): a fixed spectral ridge from each record’s earliest to its latest complete session, against a source-mean comparator, published as a negative result. Nothing is pooled across datasets, cohorts or response arms, and the results share no ranking.
Notes WBCIC-SHU: in both cohorts the spectral ridge is above the source prior and frozen CBraMod above the spectral ridge, each with a paired interval above zero. The CBraMod arm was frozen after the spectral ridge results existed and reuses their test trials; there is no matched random-weight control, and whether this checkpoint saw WBCIC-SHU in pretraining is not established, so it does not show that pretraining caused the gain. Session numbers are recording-session ordinals, not a guaranteed time gap.
Notes Longitudinal RSVP: AUROC is lower at the nominal Day 200 visit than at Day 7, with a paired interval below zero, and some people lost 0.05 AUROC or more; targets are rare, so average precision is printed beside AUROC. Day 7, 80 and 200 are the publisher’s nominal labels, and nothing here shows that elapsed time caused the change. Forenzo: the spectral ridge has a higher overall error than a constant source-mean comparator on every admitted record, in both cohorts and both response arms, with means far above medians; velocity imitates the publisher’s historical decoder output, not intended motion, and no online-control claim is made.
Notes All three datasets are new to this site. Licences, from each repository’s record: WBCIC-SHU CC BY 4.0, the RSVP dataset CC0, Forenzo’s KiltHub record CC BY 4.0; each paper states ethics approval and written or signed informed consent. Not published: per-person and per-record values (minimums, percentiles and medians, except the Forenzo ridge medians the handoff requires beside their means), confusion matrices, the Forenzo recorded-decoder strata, measured compute, and the handoffs, release decisions and independent audits themselves, pinned by hash in the review manifest because they carry private storage paths.
Route 2 of the decision-research roadmap, one representation and several questions: can one shared encoder with a fixed linear head per question answer several questions as well as a question-conditioned head, or as separate models, for less? Independent models, fixed heads, a shared hidden layer, and that layer conditioned on the question’s identity, compared at matched data and compute: EEGNet trained from scratch and frozen CBraMod features as the primary levels, three seeds each, and CBraMod adapted by LoRA, one seed, as a secondary level. Motor imagery on OpenBMI and sleep on BOAS are primary; EESM19 is a crude one-seed replication. Every contrast carries two flags — whether its interval excludes zero, and whether it lies within the 2 pp margin fixed before any result — and nothing is ranked where intervals overlap.
Notes The fixed-heads sentence is supported on neither domain: the question-conditioned head was higher on imagery against rest, and on five-stage sleep scoring, while the fixed heads were the cheaper set-up in both. The question’s identity itself added nothing measurable: the same hidden layer without conditioning was equivalent to the conditioned head within 2 pp on both motor-imagery questions and non-inferior on sleep. On sleep that layer carries the lift over the fixed heads; on motor imagery its own contrast with the fixed heads was inconclusive. Against a separate model per question, sharing one small EEGNet trunk cost accuracy on which hand; on imagery against rest and on the sleep stage that comparison was inconclusive. Questions at floor are reported and never counted, and with 21 primary entries and no multiplicity correction about one in twenty entries with no true difference may show one by chance.
Notes BOAS is new to this site. Its rights review was approved by the owner on 7 October 2026 as publishable with stated gaps. The consent statement does not say whether participants agreed to public sharing or secondary use. The ethics and consent statements come from the publisher’s dataset description and README. No peer-reviewed paper describes BOAS. The ethics reference was added to the release in version 1.1.1 (May 2025), and the release does not say when it was granted relative to the recordings. Only cohort aggregates of at least 20 people are published. Participants are pseudonymised in the public release, and neither Bitbrain’s headband nor its automatic scoring is used or evaluated.
Notes E2 sleep — CBraMod adapted by LoRA on BOAS — ran on 7 October 2026 by the owner’s decision, after every other result was known; its design and run condition were fixed at the freeze and nothing about it was chosen from results. It ran with temporary private resume checkpoints: none outlived its fit, and none remains. Not published: per-window and per-person values, including the per-person percentiles the release candidate carries for OpenBMI; measured compute totals (GPU-hours, wall time per fit and per block, training steps per outer fold), while the cost ledger publishes, labelled indicative, the two times measured interleaved on a shared GPU; and the aggregates and the three independent audits themselves, pinned by hash in the review manifest because they carry private storage paths.
The v9 foundation-model evaluation: eleven further EEG foundation models — sixteen encoder checkpoints — run as frozen probes on the eight core protocols with the published LaBraM and CBraMod recipe and only the encoder swapped (the same windows, people, folds, heads and scoring), and nine of them adapted on EEGMAT with the 1 October recipe, a head trained on the frozen encoder against rank-4 LoRA over three seeds. New rows beside the core matrix, in their own files: one JSON, and a CSV per protocol whose first columns are the core results CSVs’. Grouped by family, never ranked: a row is above or below another only where their 95% intervals do not overlap. Every cell says whether its dataset is in the model authors’ published pretraining list, checked on 4 October 2026 with the source, and every model carries its weights licence.
Notes Sleep staging is the only protocol where new frozen cells — ST-EEGFormer Large and Base, and REVE Large — lie entirely above every published row. On the other six scored protocols no new cell lies entirely above the published non-foundation row with the highest point estimate; on BETA, no new row lies above standard CCA, which needs no training. On EEGMAT, frozen readouts and LoRA order the models differently: one fixed recipe on one task, with LoRA budgets that differ by model, not a ranking.
Notes BETA is in ST-EEGFormer’s published pretraining list and TMNRED in SingLEM’s, so those cells are flagged; ZUNA 1.1 does not list its pretraining data, so every ZUNA 1.1 cell is marked unknown. Every other cell, LaBraM’s and CBraMod’s included, is not in the authors’ published pretraining list as checked on 4 October 2026: a sourced statement, not proof that the recordings were never seen. Weights licences travel with the rows: REVE under its Responsible Use License, LUNA under CC BY-ND 4.0, ERP-FM for non-commercial use only, and ZUNA 1.1’s model card limits it to research use, not diagnosis or clinical use.
Notes Not published: per-trial, per-person and per-fold values, features, and timings and memory — measured on a shared GPU, so the CSVs leave the scoring-time column empty — the declared sensitivity runs beyond the row footnotes that state them, and the aggregate, the group summaries and the six independent audits themselves, pinned by hash in the review manifest because they carry private storage paths. Two BrainOmni cells were not run, because its tokenizer needs two-second windows: they are null with that reason, never zero. The released matrix keeps its bytes; the new rows never enter experiments.json.
Route 1 of the decision-research roadmap, reliable decisions: when a decoder should decline to decide, and what its confidence is worth. The saved test scores of models this site already publishes — spectral ridge, EEGNet and two LaBraM arms on EEGMAT (36 people); standard CCA, CBraMod and EEGNet on BETA (70 people) — rescored under a fixed confidence threshold, a coverage target, a certified selective risk and a learned reject option, with probability quality and a person-specific recalibration whose label cost is counted. No classifier was retrained, so every full-coverage accuracy equals the published one. Every protocol is balanced or uniform by design: these are method comparisons, never deployment error rates, and nothing is ranked where intervals overlap.
Notes The certified selective risk certified no outer fold on EEGMAT, so it accepted nothing there: no error rate, not a zero one. On BETA its realised error exceeded a certified fold’s own target in some folds: a nominal guarantee across people, whose failure rate is measured here, not assumed. The learned reject option had a lower error than the model’s own calibrated confidence for no method.
Notes ds003810 (ten people) is shown as crude, in a collapsed robustness panel beside secondary seeds, two sensitivity arms (their risk-certification fold counts only) and an exploratory reject head trained jointly with its classifier. Left out: the idle protocol, whose route-1 window proportions are a different unit from the trial-level idle figures on the same page, and every figure of the BNCI2015-001 arm, whose editorial hold stands.
Notes Not published: per-trial, per-person and per-fold values, including the release candidate’s person-level percentiles and its median temperature over folds; the accuracies of the matched-holdout, map-sensitivity and jointly trained arms; and the independent audits themselves, pinned by hash in the review manifest because they carry private storage paths. The primary audit’s one failed check is a documentation error, not a computation error, and is stated as such.
Two new sources, two separate questions, each with two fixed classical CPU baselines and no shared ranking. Dreem: five-stage sleep staging against the publisher’s consensus in 25 healthy sleepers (DOD-H) and 55 people with obstructive sleep apnoea (DOD-O), kept as separate experiments — a training prior and a spectral ridge, accuracy beside balanced accuracy, macro F1, Cohen’s kappa and the ridge’s stage-by-stage results. And OpenBMI: a motor-imagery decoder trained on a person’s first session and tested on their second after 0, 10, 20 or 40 labelled trials from it, in 51 people, with how many people declined beside every mean change. No foundation-model or fine-tuning result.
Notes The spectral ridge never predicts N1 in either cohort, so its N1 precision is not defined: it is published as null, never as zero. The training prior’s balanced accuracy sits slightly above one fifth because a night’s mean covers only the stages that night contains; it is a floor, not a chance level. Neural-network and foundation models on the Dreem cohorts are held, because the source’s physical units disagree.
Notes OpenBMI is an expanded cohort under the same fixed method — the original 40 people with their results unchanged, plus 11 later-eligible people — not an independent replication; an earlier 40-person snapshot was prepared but never published here and is not a separate result. The relative-PSD change at 40 trials has an interval that includes zero and is not established.
Notes Not published: per-person and per-night values (including the worst observed change, medians and 10th percentiles), Brier scores, log-likelihoods and calibration errors, the combined DOD-H plus DOD-O summary, and the independent audits themselves, which are pinned by hash in the review manifest because they carry private storage paths.
Twenty further people of the asynchronous SSVEP release, scored with a rejection threshold fixed on the four-person pilot and with a personal one: detection, coverage, correct-and-accepted output and false acceptance per non-control state, side by side. And a new source, LTRSVP: a P300 decoder trained on one recording at 5 or 10 images a second and tested on a different, 10-Hz recording, with the full rate-by-rate matrix. Fixed classical baselines; no foundation-model or fine-tuning result.
Notes The personal threshold raised detection balanced accuracy for 10 of 20 people and lowered it for 8, and did not raise the share of commands both accepted and correct. The image-rate difference has an interval that crosses zero, and rate and recording change together, as does time in the session (the original study presented the rates from the lowest to the highest, not randomised), so no rate effect is claimed.
Notes Not published: per-person values, thresholds, predictions, and the independent audits themselves, which are pinned by hash in the review manifest because they carry private storage paths.
New people, same task, zero labels from the test person: LaBraM adapted on mental arithmetic three ways — head only, last block, rank-4 LoRA — with the same checkpoint, folds, starting heads, batch order and five-epoch recipe, over three seeds. The next-day experiment is listed as status only.
Notes Printed beside the core matrix's frozen LaBraM readout on the same people and folds, because this batch's head-only arm is a short gradient-trained head and not the strongest frozen readout.
Notes Not published: memory figures (allocator samples are lower bounds, not per-method peaks) and any figure from the next-day experiment.
P300 calibration carried from a PC screen to a VR headset and back for 21 people; walking-speed scores from dry-electrode EEG printed beside a movement-nuisance comparator; and a four-person pilot of SSVEP windows accepted when no command was intended. Fixed CPU baselines; no foundation-model or fine-tuning result.
In-ear against scalp sleep staging, four posterior electrodes against sixteen, a physical-phantom motion test, and the status of planned adaptation experiments.
Notes Amended the EESM19 consent note (see data use). The core snapshot and the topic export were regenerated with one added citation; no score changed. Their current hashes, below, are the amended files.
BCI Report