PaperlayerView exact revision on arXiv

Source-linked AI summary

GRAFT-EEG: Gauge-Aware Physical Field Transport for Cross-Dataset Motor-Imagery EEG

Ying Zhou, Mengting Liu

arXiv:2610.23620v1cs.LGeess.SP

TL;DR

Cross-dataset EEG decoding faces changes in montage, reference, and active sensors that alter the observation operator. GRAFT-EEG removes the active-set common-reference component and transports native masks through a template motor basis, with a static spectral classifier and bounded residual. Across five held-out datasets, it outperformed strict-clean BIOT, while the evidence remains bounded to the tested domains and operators.

  • Problem

    Cross-dataset EEG methods must address montage, reference, and active-sensor changes that alter the observation operator, while evidence on decoding unseen montages without target signals or empirical statistics during fitting is limited.

    Source: §1 · pp. 1–2

  • Method

    GRAFT-EEG removes the active-set common-reference component, transports native masks through a template motor basis, and pairs a source-fitted spectral classifier with a bounded state-space residual.

    Source: §1 · p. 2

  • Results

    60.865% equal-target BA for dynamic GRAFT-EEG versus 52.684% for strict BIOT across five leave-one-dataset-out targets.

    Source: §8 · p. 9

  • Takeaways & Limitations

    The tested exact same-parent operator replacements reduced BA by 3.985–11.241 pp, supporting a material contribution from the task-matched physical observation operator within this parent model.

    Source: §8 · p. 9

    Evidence is bounded to five binary motor-imagery domains and tested operators; the held-out dataset, not its participants, is the independent generalization unit.

    Source: §7 · p. 9

Abstract

from arXiv · show

Cross-dataset EEG decoding often treats montage, reference, and active-sensor changes as representation shift, although they also alter the observation operator. We study target-statistics-blind transfer: fitting receives no target signal, label, or empirical statistic, and electrode coordinates only instantiate a fixed prediction operator. method removes the active-set common-reference gauge and transports each native mask through a template motor leadfield. A source-fitted spectral classifier anchors the decision; a bounded state-space-model (SSM) residual is secondary. We establish restricted identifiability and an operator-perturbation bound. Across five leave-one-dataset-out targets (253 participants), dynamic method achieved 60.865% equal-target balanced accuracy versus 52.684% for strict-clean 5/5 BIOT (hierarchical difference +8.182 points; 95% CI +6.382 to +10.196; positive in every target). On four EEGPT-clean targets, the descriptive means were 61.520% and 59.057%. In four post-hoc exact same-parent tests, FULL exceeded coordinate-basis, matched-spectrum, row-shuffled, and wrong-anatomy controls by 9.959, 9.698, 11.241, and 3.985 points, respectively (all 95% CIs excluded zero; Holm-adjusted P<0.001). A separate 60-participant same-paradigm evaluation retained a positive physical-operator contrast (+4.029 points; +2.338 to +5.773), whereas dynamic--static was inconclusive; its preregistered status is partial confirmation rather than unrestricted external validation. The SSM added +1.059 points (+0.013 to +1.892). Claims remain bounded to five binary motor-imagery domains and the tested operators.

1 INTRODUCTION

Cross-dataset EEG decoding must contend with changes in montage, reference, and active sensors that alter the observation operator, not just the representation. GRAFT-EEG studies transfer without target signals, labels, or empirical statistics during fitting, using reference quotienting and physical-field transport with a static classifier and bounded residual.

1 INTRODUCTION

  • Transfer couples physiological and procedural shift with changes in observation coordinates; channel intersection discards native sensors, while learned adapters can obscure which target information fitted them.The paper tests whether an unseen montage can be decoded before any target signal or empirical statistic is opened.Source: §1 · p. 1
  • The access contract permits declared target coordinates to instantiate fixed operators but restricts fitting to source signals and labels.Target-derived fitting, normalization, calibration, and model selection are prohibited.Source: §1 · pp. 1–2
  • GRAFT-EEG removes the active-set common-reference component and transports the remaining field into a fixed bilateral motor basis.A static spectral anchor is paired with a bounded state-space residual.Source: §1 · p. 2
  • The evaluation covers five domains with strict and partial denominators, four exact same-parent operator controls, and a same-paradigm partial-confirmation external evaluation.It also reports domain-aware uncertainty and baseline-health diagnostics.Source: §1 · p. 2

2 RELATED WORK

Prior EEG systems accommodate heterogeneous layouts through pretraining, channel tokenization, coordinates, or learned channel-unification modules. This paper examines a distinct access-constrained question: decoding a full unseen montage before target signals or empirical statistics are opened.

2 RELATED WORK

  • Large EEG models use masking, tokenization, and diverse corpora, while newer systems also use coordinates or learned channel-unification modules.These approaches do not isolate whether a full unseen montage can be decoded before opening target signals or empirical statistics.Source: §2 · p. 2
  • GRAFT-EEG uses montage coordinates and a shared template motor basis, not participant-specific source localization, and its bounded temporal branch makes no streaming claim.The paper distinguishes this access-constrained question from protocols that are not uniformly target-statistics-blind.Source: §2 · p. 2

3 PROBLEM SETTING AND ACCESS CONTRACT

The target-statistics-blind contract withholds the held-out dataset’s signals, labels, trial queue, and empirical statistics during fitting, while allowing its declared geometry to instantiate a fixed operator. Target signals are opened only after source-fitted objects are committed, and labels only after predictions are frozen.

3 PROBLEM SETTING AND ACCESS CONTRACT

  • Each fold holds out one complete dataset; its geometry may instantiate the declared linear operator, but its signals, labels, trial queue, and empirical statistics remain closed during fitting.Source features, hyperparameters, classifiers, state-space weights, and calibration parameters are committed before target signals are opened.Source: §3 · p. 2
  • Target-statistics-blind means blind to target signals, labels, and empirical statistics during fitting, not blind to electrode coordinates.After lock, per-trial masks instantiate operators but are not used for refitting, normalization, calibration, early stopping, or threshold selection.Source: §3 · p. 2
  • Figure 1 depicts geometry available before lock, signals at inference, and labels only at scoring.Its access contract prohibits target-derived fitting, normalization, calibration, and model selection.Source: Figure 1 · p. 3

4 GRAFT-EEG

GRAFT-EEG removes the active-set common-reference component, transports each native mask through a template motor basis, and uses a static spectral classifier with a bounded dynamic residual. Its theoretical claims concern restricted identifiability and operator perturbations, not unique recovery of arbitrary nonlinear sources.

4.1 ACTIVE-SET REFERENCE QUOTIENT

  • EEG potentials have an additive common-offset ambiguity, so applying a fixed inverse to differently referenced observations can couple the representation to an arbitrary gauge.The same projection is applied to signal and matching lead-field rows, while failed channels are excluded rather than filled from target data.Source: §4.1 · p. 3

4.2 MASK-SPECIFIC TEMPLATE MOTOR-FIELD TRANSPORT

  • The observation matrix uses a fixed eight-dimensional template motor basis constructed from declared montage geometry and an fsaverage motor region, without EEG samples, labels, or target statistics.A mask-specific ridge inverse transports the field; its regularization scale is selected by source-subject validation.Source: §4.2 · p. 3

4.3 WHAT HETEROGENEOUS OPERATORS CAN IDENTIFY

  • When dynamics and operators are fixed and the joint observability matrix has rank d, noiseless stacked observations uniquely determine the initial latent state.If latent coordinates are learned jointly with dynamics and operators, identifiability is only up to an invertible similarity transform; arbitrary nonlinear sources are not thereby identifiable.Source: §4.3 · p. 4
  • Operator-perturbation analysis separates measurement noise from montage mismatch and motivates spectral regularization.Across 50 target–condition units, reduced-operator change tracked dynamic BA change (Spearman ρ = −0.449; 95% CI −0.760 to −0.282; permutation P < 0.001), characterizing tested operators rather than a universal placement tolerance.Source: §4.3 · p. 4

4.4 STATIC MOTOR-SPECTRAL ANCHOR

  • The static anchor represents each trial with full-window log power in three frequency bands for each latent component, yielding a 24-dimensional vector.Its weighted scaler and logistic classifier are selected on source subjects and refitted on eligible source trials.Source: §4.4 · p. 4

4.5 STABLE GATED STATE-SPACE RESIDUAL

  • The dynamic branch processes 11 short-window vectors through a diagonal recurrence and adds a bounded residual to the frozen static logit before the sigmoid.Its residual starts at zero, initially recovering the static classifier; optimization updates the dynamic branch while leaving the static logit fixed.Source: §4.5 · pp. 4–5, Figure 2 · p. 5

5 EXPERIMENTAL DESIGN

The evaluation uses five public motor-imagery datasets with native channel masks and a leave-one-dataset-out protocol. Comparator and operator-control designs hold target fitting blind to target signals, labels, and empirical statistics.

5.1 DATASETS AND PROTOCOL

  • Five public collections provide 41,342 left/right-hand trials from 253 participants, retaining native 22–64-channel montages rather than reducing them to a common intersection.Signals are filtered, resampled to 200 Hz, and epoched to [0.5, 3.5) s.Source: §5.1 · p. 5, Table 1 · p. 6
  • Each fold holds out one dataset, uses four source libraries, and evaluates static and dynamic models on shared target queues.The primary endpoint is subject-mean balanced accuracy; uncertainty uses a dataset–subject hierarchical bootstrap across five target datasets.Source: §5.1 · pp. 5–6

5.2 COMPARATORS AND PROTOCOL FAIRNESS

  • The strict comparison includes classical, neural, and pretrained baselines, with clean eligibility tracked separately from execution.BIOT is clean on all five targets, while EEGPT and LaBraM are clean on four because PhysionetMI overlaps with pretraining; partial denominators are not rank-mixed.Source: §5.2 · p. 6, Table 2 · p. 6
  • In the equal-information analysis, FI and SSI receive GRAFT-EEG’s active set and map masks to 17 fixed channels using source-fitted features or EEGNet.All 60 cells completed fitting, prediction, and scoring; Dreyer2023A is a no-tuning same-paradigm evaluation with partial-confirmation status.Source: §5.2 · p. 6
  • The exact same-parent ladder replaces the physical leadfield with coordinate-only, matched-spectrum random, row-shuffled, or wrong-anatomy operators while keeping the evaluation protocol fixed.The controls preserve the split, features, head, ridge rule, budget, and target queues; 75 fits were frozen before target evaluation, and the ladder is post-hoc rather than independent validation.Source: §5.2 · pp. 6–7, Figure 3 · p. 7

6 RESULTS

Across five held-out datasets, dynamic GRAFT-EEG outperformed strict-clean BIOT, and the physical leadfield exceeded each tested replacement operator. Secondary analyses qualify the result with baseline-health checks, a partial external confirmation, and coordinate sensitivity.

6.1 STRICT FIVE-TARGET COMPARISON

  • 60.865% equal-target BA for dynamic GRAFT-EEG versus 52.684% for BIOT, a +8.182 pp difference with a hierarchical 95% interval of +6.382 to +10.196.All five target effects were positive (+6.506 to +10.437 pp); Cho was the worst target for both methods.Source: §6.1 · p. 7
  • FULL exceeded the coordinate basis by +9.959 pp, the matched-spectrum random operator by +9.698 pp, the row-shuffled operator by +11.241 pp, and the wrong-anatomy leadfield by +3.985 pp.All four tests had Holm-adjusted P < 0.001; the controls do not establish superiority over every possible physical or learned operator.Source: §6.1 · pp. 7–8
  • Dynamic minus static was +1.059 pp, with a hierarchical 95% interval of +0.013 to +1.892; four target effects were positive, but BNCI was −0.586 pp.The SSM contribution is secondary to cross-montage harmonization.Source: §6.1 · p. 8
  • Under the stricter source-only-geometry sensitivity, dynamic and static BA were 60.460% and 59.713%, respectively, and the held-out montage was still required to instantiate the prediction operator.The method is therefore not geometry-blind.Source: §6.1 · p. 8

6.2 COVERAGE, FAIRNESS, AND EXTERNAL BOUNDARY

  • On the four targets clean for EEGPT, dynamic GRAFT-EEG averaged 61.520% BA versus 59.057% for EEGPT, a descriptive +2.463 pp difference not included in the complete 5/5 ranking.EEGPT and LaBraM are strict-clean on 4/5 targets because PhysionetMI overlaps with pretraining; CodeBrain is provenance-ineligible on all five.Source: §6.2 · p. 8
  • Source-health diagnostics found GRAFT-EEG healthy under the prespecified rule and without single-class target cells, while several baselines were source-weak or had systematic single-class predictions.No direct implementation defect was confirmed, and weak baseline results were not uniformly attributed to domain shift.Source: §6.2 · p. 8
  • In the separate 60-participant Dreyer2023A evaluation, physics versus no-physics was +4.03 pp [+2.34, +5.77], while dynamic versus static was inconclusive at +0.57 pp [−0.08, +1.23].The preregistered decision was partial confirmation, not unrestricted external validation; few-shot label-revealing results were not a fully supervised oracle or performance ceiling.Source: §6.2 · p. 8
  • Across 50 target–condition units, reduced-operator change tracked dynamic BA change (ρ = −0.449; 95% CI −0.760 to −0.282; permutation P < 0.001), with heterogeneous responses.At 20 mm, equal-target BA changed by −1.064 pp for tangent errors and −2.220 pp for rigid errors; no model was refitted.Source: §6.2 · p. 8

7 DISCUSSION AND LIMITATIONS

The central finding concerns the observation operator rather than the temporal head: GRAFT-EEG exceeded the strongest clean 5/5 pretrained comparator in every target, while operator controls reduced balanced accuracy. The evidence remains bounded by heterogeneous results and the tested domains and operators.

7 DISCUSSION AND LIMITATIONS

  • 8.18 pp: the native-montage model exceeded the strongest clean 5/5 pretrained comparator in every target.The central finding concerns the observation operator rather than the temporal head.Source: §7 · pp. 8–9
  • 3.98–11.24 pp: replacing the motor leadfield with four controls reduced balanced accuracy under the same parent and budget.The ladder is post-hoc and operator-specific, not proof that physics is universally optimal.Source: §7 · p. 9
  • The SSM contribution was modest and heterogeneous: +1.06 pp overall, four positive targets, −0.59 pp on BNCI, and an inconclusive external contrast.Synthetic known-source controls do not form a real-EEG same-parent classification ladder, and participant-specific source recovery lies beyond scope.Source: §7 · p. 9
  • Calibration lowered probability error without changing discrimination, stress losses reached −8.93 pp, and legacy stability diagnostics showed greater seed-to-seed dispersion than coordinate and no-physics comparators.Failed conditions also outnumbered recoverable ones in known-source tests; claims are confined to the evaluated operating range.Source: §7 · p. 9
  • Generalization is assessed across five held-out datasets, with dataset confounded by device and protocol; the primary task is binary hand motor imagery.The fsaverage basis is not participant-specific localization, and arbitrary montages, streaming, and clinical deployment remain outside the evidence boundary.Source: §7 · p. 9

8 CONCLUSION

Across five leave-one-dataset-out targets, GRAFT-EEG outperformed strict-clean BIOT while using native montages and active masks without target signals, labels, or fitting statistics. The conclusion is bounded to five binary motor-imagery domains and the tested operators.

8 CONCLUSION

A PROOFS OF THE RESTRICTED THEORY RESULTS

  • The propositions establish identifiability and perturbation bounds only for a fixed finite-dimensional linearization, not nonlinear EEG inversion or participant-specific cortical generators.The proof also limits identifiability to a latent-coordinate class under invertible changes of coordinates.Source: A PROOFS OF THE RESTRICTED THEORY… · p. 12

B DATASET AND OBSERVATION-OPERATOR INVENTORY

  • Retained active masks vary within datasets, making the operator inventory mask-specific rather than reducible to a single inverse per nominal montage.Cho2017 alone has 223 distinct retained masks, and 4,710 retained trials have at least one inactive native channel.Source: B DATASET AND OBSERVATION-OPERATOR… · p. 12

C STRICT SOURCE-ONLY GEOMETRY-BASIS SENSITIVITY

  • When the target montage was withheld until all 25 target–seed fits were sealed, source-only geometry lowered canonical dynamic mean by 0.406 pp; BNCI still showed a negative dynamic–static contrast.The target montage was then used once with the frozen fold-specific basis to construct the prediction operator.Source: C STRICT SOURCE-ONLY GEOMETRY-BASIS… · p. 13, Table S2 · p. 13

D COMPLETE STRICT FIVE-TARGET DISCRIMINATION SUMMARY

E POST-SEAL BASELINE HEALTH DIAGNOSTICS

  • The health audit found no confirmed code defect, but zero cells met the prespecified domain-shift attribution rule, supporting disclosure rather than post-hoc baseline removal or reranking.Several classical baselines had systematic single-class target predictions and unavailable source validation; near-chance source validation was not classified as healthy.Source: E POST-SEAL BASELINE HEALTH DIAGNOSTICS · p. 15, Table S6 · p. 15

F REGISTERED CELL REPRODUCIBILITY AND SUBJECT HETEROGENEITY

Across target omissions, dynamic GRAFT-EEG retained positive mean advantages over BIOT and the static model, while participant- and seed-level results show greater heterogeneity for the dynamic–static contrast. Supplementary controls, calibration, stress tests, and label-revealing adaptation probe distinct boundaries of the evaluation.

F REGISTERED CELL REPRODUCIBILITY AND SUBJECT HETEROGENEITY

  • Dynamic BA exceeded static BA in 22 of 25 target–seed cells, with all three reversals in BNCI; participant-level dynamic–static differences ranged from 33.3% to 73.1% positive across targets.Dynamic–BIOT differences were positive for 75.2–90.7% of participants in each target; seed cells are nested observations, not independent target domains.Source: F REGISTERED CELL REPRODUCIBILITY AND… · p. 16, Table S8 · p. 16, Table S9 · p. 17
  • Omitting one target at a time left the dynamic–BIOT estimate at 7.618–8.600 pp and the dynamic–static estimate at 0.792–1.470 pp.These are arithmetic means of the remaining four target effects; the primary interval remains based on all five targets.Source: F REGISTERED CELL REPRODUCIBILITY AND… · p. 17, Table S10 · p. 17

G POST-SCORE-SEAL EQUAL-INFORMATION FI/SSI CONTROLS

H CALIBRATION, STRESS, AND EXECUTION COVERAGE

  • Source-only calibration lowered pooled ECE, Brier score, and negative log likelihood without changing BA, AUROC, or AUPRC; stress losses reached −8.93 pp under −10 dB noise.Calibration did not improve every target: BNCI and Shin worsened on all three probability-error summaries, while seed-to-seed stability contrasts were adverse.Source: H CALIBRATION, STRESS, AND EXECUTION… · p. 20, Table S12 · p. 20, Figure S3 · p. 20

I FEW-SHOT TARGET-LABEL ADAPTATION BOUNDARY

  • The few-shot analysis reveals 1%, 5%, or 10% of target labels only after freezing the zero-target model, so its results are not target-statistics-blind performance.At 1% labels, full-head and PEFT BA were 61.323% and 61.130%, respectively; the frozen linear probe was 55.139%.Source: I FEW-SHOT TARGET-LABEL ADAPTATION… · pp. 21–22, Table S14 · p. 22, Table S15 · p. 22
  • In the separate 60-participant evaluation, the physical-operator contrast was positive, but dynamic–static was inconclusive and the study is classified as partial confirmation.Figure S5 reports paired dynamic and static BA for all 60 participants; Table S16 distinguishes descriptive absolute scores from preregistered contrasts.Source: Figure S5 · p. 23, Table S16 · p. 23

J SEPARATELY SEALED EXTERNAL EVALUATION

A separate 60-participant evaluation found a positive physical-operator contrast but an inconclusive dynamic–static contrast, and is reported as partial confirmation rather than unrestricted external validation. Supplementary analyses examine known-source recovery, frozen computations, real-trial displays, and coordinate sensitivity within the tested settings.

J SEPARATELY SEALED EXTERNAL EVALUATION

  • In Dreyer2023A, physics–no-physics was +4.03 pp (+2.34 to +5.77), while dynamic–static was +0.57 pp (−0.08 to +1.23; inconclusive).The evaluation used 60 held-out participants; the authors classify it as partial confirmation, not unrestricted independent external validation.Source: J SEPARATELY SEALED EXTERNAL EVALUATION · pp. 23–24

K KNOWN-SOURCE VALIDATION AND CONDITION LEDGER

  • Across 25 target–seed pairs, the physical operator increased mean source correlation by 0.190 versus shuffled-row controls and 0.192 versus spectrum-matched random controls.Of 755 executed known-source conditions, 232 met the recovery criterion and 523 entered the scientific failure region; these are synthetic-source results, not evidence of real-EEG cortical localization.Source: K KNOWN-SOURCE VALIDATION AND… · p. 24, Table S17 · p. 24, Table S18 · p. 24

L FROZEN TRIAL REPLAY AND SOURCE-FIT PARAMETER ANATOMY

  • The frozen replay shows one real retained trial passing through gauge projection, the active operator, transported latent, spectral sequence, and bounded residual; only the final point is the canonical scored output.Its final output matches the sealed logit within 1.49 × 10^-8; prefixes 1–10 are diagnostic only.Source: L FROZEN TRIAL REPLAY AND SOURCE-FIT… · pp. 25–26, Figure S7 · p. 26

M THREE-DIMENSIONAL GEOMETRY-CONDITIONED INVERSE KERNELS

  • The inverse-kernel visualization shows coefficients for the eight rows of one locked operator on a 60-of-64-channel mask; its rendered scalp surface is orientation and display, not participant anatomy or source localization.Surface interpolation does not enter training or evaluation.Source: M THREE-DIMENSIONAL… · p. 27, Figure S9 · p. 28

N DETERMINISTICALLY SELECTED REAL-EEG ATLAS

O STRICT BENCHMARK MATRIX AND BASELINE HEALTH

  • The strict target–method matrix distinguishes the 15-method 5/5 denominator from partial-denominator encoders and leaves unresolved-provenance CodeBrain separate; missing entries are not replaced by zero.Health panels retain missing N/A values rather than treating them as zero.Source: Figure S13 · p. 32, O STRICT BENCHMARK MATRIX AND BASELINE… · p. 32

P SEALED COORDINATE-PERTURBATION SENSITIVITY

  • Across 50 target–condition units, reduced-operator change tracked dynamic BA change (Spearman ρ = −0.449; 95% CI −0.760 to −0.282), but the result implies no universal millimetre threshold.The coordinate sensitivity is limited to five declared montages, shared fsaverage anatomy, and the tested synthetic perturbation models.Source: P SEALED COORDINATE-PERTURBATION… · p. 32, Figure S14 · p. 33, Table S19 · p. 33

Q IMPLEMENTATION AND REGISTERED INFERENCE

The section specifies the registered implementation, controls, uncertainty procedures, and audit gates. It also bounds efficiency claims and documents the access and provenance rules used for the evaluation.

Q IMPLEMENTATION AND REGISTERED INFERENCE

  • Trials are removed when more than 20% of native channels are inactive or fewer than nine remain, after channels are flagged by signal-quality thresholds.A channel is inactive if it is non-finite, exceeds 5,000 µV in absolute amplitude, or has peak-to-peak amplitude below 0.1 µV.Source: Q IMPLEMENTATION AND REGISTERED… · p. 33
  • The motor basis retains four leading right-singular vectors from gauge-projected lead fields across five declared montages, with source-library validation selecting ridge and classifier parameters.The dynamic branch uses source validation to select its epoch count before refitting on all source trials; Table S20 identifies source-selected quantities as optimized without target-derived information.Source: Q IMPLEMENTATION AND REGISTERED… · p. 34, Table S20 · p. 35
  • Four exact same-parent controls replace the motor leadfield with coordinate-only, spectrum-matched random, row-shuffled, or wrong-anatomy operators under shared refitting rules.The independent composite audit covered 253 participants and returned PASS; source-only geometry was excluded from the four-control Holm family.Source: Q IMPLEMENTATION AND REGISTERED… · p. 34
  • Uncertainty estimates resample the five targets and then subjects within each selected target, while calibration is fitted on source predictions only.Main effects average five registered seeds within subject and use 20,000 hierarchical bootstrap replicates.Source: Q IMPLEMENTATION AND REGISTERED… · p. 34
  • Synthetic validation varies channel budgets from 64 to 8 and SNR from clean to −10 dB, reserving test truth for registered performance metrics and failure gates.Shuffled-row and spectrum-matched operators serve as negative controls.Source: Q IMPLEMENTATION AND REGISTERED… · p. 34

R EFFICIENCY REPORTING BOUNDARY

  • Comparable GPU wall-clock summaries cover only executed neural fits; classical timing was not recorded over an equivalent scope, so no universal speed ranking is inferred.Table S21 leaves unavailable timings unfilled and does not sum the shared static and dynamic fit record.Source: R EFFICIENCY REPORTING BOUNDARY · p. 35, Table S21 · p. 35

S ACCESS CONTRACT AND AUDIT GATES

  • Execution proceeds through hash-matched stages, and completion requires normal exits, expected cells, passing group checks and replay, and P0 = P1 = 0.Table S22 specifies that each stage may read, but cannot rewrite, outputs accepted at an earlier gate.Source: S ACCESS CONTRACT AND AUDIT GATES · p. 35, Table S22 · p. 36

T REPRODUCIBILITY AND SOURCE-DATA MANIFEST

  • The source-data export contains 20 CSV files derived from 17 sealed source objects, while exported tables do not recreate the scientific anchors.The manifest records hashes and provenance; the tables do not rescore predictions, recalibrate probabilities, or change decisions.Source: T REPRODUCIBILITY AND SOURCE-DATA… · p. 36, Table S23 · p. 36

V NAME, CODE, AND RESULT IDENTITY

The section distinguishes the manuscript's named result identity from a later, separate parent namespace. It also records the fixed FI/SSI target montage and the provenance of quantitative figures.

V NAME, CODE, AND RESULT IDENTITY

  • GRAFT-EEG dynamic maps to the code identifier mosaic dynamic, and its static anchor maps to mosaic static; later graft parent FULL/A2–A8 is a different system.The paper's main claims do not use that later namespace or infer an A2–A8 effect without a corresponding result table; Table S25 records non-interchangeable result identities.Source: V NAME, CODE, AND RESULT IDENTITY · p. 37, Table S25 · p. 37

W FIXED 17-CHANNEL FI/SSI TARGET MONTAGE

  • FI/SSI maps each exact native active mask to a fixed 17-channel montage, then applies fixed scaling and clipping for EEGNet.The montage includes Fp1, Fp2, F7, F3, Fz, F4, F8, C3, Cz, C4, P3, Pz, P4, T7, T8, P7, and P8; tangent reference, covariance shrinkage, scaler, and classifier are source-fitted.Source: W FIXED 17-CHANNEL FI/SSI TARGET MONTAGE · pp. 37–38

X RESULT ARTIFACT MAP

  • The artifact map binds quantitative figures to locked objects, denominators, source files, and hashes, including Figure 3's 253 participants, five targets, and four exact controls.Figure S9's three-dimensional scaffold is marked visualization-only.Source: X RESULT ARTIFACT MAP · p. 38
Loading 2610.23620v1…