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

Ying Zhou; Mengting Liu

Dataset	Subj.	Raw	Retained	Excl.	Ch.	Samples	Hz	Patterns	Inactive trials
BNCI2014_001	9	2,592	2,592	0	22	600	200	1	0
Cho2017	52	10,520	10,492	28	64	600	200	223	4,710
Lee2019_MI	54	21,600	21,600	0	62	600	200	6	34
PhysionetMI	109	4,918	4,918	0	64	600	200	1	0
Shin2017A	29	1,740	1,740	0	30	600	200	1	0
Total	253	41,370	41,342	28	—	—	—	232	4,744

Appendix

Table S1: Detailed immutable queue and operator inventory. “Raw” is the prepared H5 queue before the fixed trial gate. “Patterns” is the number of distinct retained active-channel masks.

Target ( $n$ )	Canon. dyn.	Src.-only dyn.	Change (pp)	Src.-only stat.	Dyn.–stat. (pp)
BNCI2014-001 (9)	64.336	62.423	$-1.913$	64.668	$-2.245$
Cho2017 (52)	58.172	58.141	$-0.031$	56.868	$+1.273$
Lee2019-MI (54)	60.318	60.392	$+0.074$	59.361	$+1.031$
PhysionetMI (109)	58.247	58.020	$-0.227$	56.724	$+1.296$
Shin2017A (29)	63.253	63.322	$+0.069$	60.943	$+2.379$
Equal-target mean (253)	60.865	60.460	$-0.406$	59.713	$+0.747$

Appendix

Table S2: Access-boundary sensitivity: canonical target-geometry construction versus a strict source-only geometry basis. Values are subject-mean balanced accuracy averaged over five seeds. “Change” is source-only dynamic minus canonical dynamic; it isolates the cost of withholding the target montage until after the source fits are sealed.

Method	BA	Macro-F1	AUROC	Metric availability
Graft-EEG dynamic	60.865	58.886	66.447	all three sealed
Graft-EEG static	59.806	58.015	64.588	all three sealed
BIOT + LR	52.684	46.954	56.131	all three sealed
CBraMod frozen + LR	50.740	—	—	BA only
EEGConformer	50.448	44.322	50.808	all three sealed
ODEBrain (scratch)	50.174	39.996	51.060	all three sealed

Appendix

Table S3: Complete strict five-target discrimination summary. Values are percentages. All 15 methods completed 25/25 target–seed cells and are strict-eligible on 5/5 targets. Missing secondary metrics are not imputed.

Method	BNCI	Cho	Lee	PhysioNet	Shin	Equal-target
Graft-EEG dynamic	64.336	58.172	60.318	58.247	63.253	60.865
Graft-EEG static	64.923	56.817	59.448	56.717	61.126	59.806
BIOT + LR	56.250	50.086	52.524	51.741	52.816	52.684
CBraMod frozen + LR	52.392	50.241	50.341	50.001	50.724	50.740
EEGConformer	50.455	50.183	49.958	50.321	51.322	50.448
ODEBrain (scratch)	50.085	49.871	50.621	50.210	50.080	50.174

Appendix

Table S4: Target-resolved subject-mean balanced accuracy for every strict-clean 5/5 method. Values are percentages after averaging the five registered seeds within target. The last column weights targets equally.

Method	BNCI	Cho	Lee	PhysioNet	Shin	Available mean	Strict	Exclusion
EEGPT + LR	65.664	49.315	60.985	—	60.264	59.057	4/5	overlap
LaBraM + LR	53.997	50.584	51.267	—	52.000	51.962	4/5	overlap
CodeBrain + LR	—	—	—	—	—	—	0/5	unresolved

Appendix

Table S5: Executed but strict-incomplete pretrained encoders. A numeric cell is shown only when the target is clean under the checkpoint-provenance review. The available-target mean is descriptive and is not inserted into Table S4 .

Method	Source BA	Positive-rate range	$H_{\mathrm{hard}}$	$H_{\mathrm{prob}}$	Single	Coverage
Graft-EEG dynamic	60.8	26.4–56.2	0.974	0.902	0	25/25
Graft-EEG static	58.6	39.8–55.4	0.993	0.948	0	25/25
BIOT	53.6	26.9–76.8	0.870	0.859	0	25/25
EEGPT + LR	61.2	13.3–54.4	0.870	0.830	0	20/20
LaBraM + LR	52.4	37.0–75.6	0.929	0.978	0	20/20
CodeBrain + LR	N/A	N/A	N/A	N/A	N/A	25/25 †

Appendix

Table S6: Post-seal implementation and prediction-health audit. Source BA is the mean source-validation balanced accuracy (%). “Positive-rate range” is the minimum–maximum target positive prediction rate across eligible target–seed cells (%).

H_{\mathrm{hard}}

and

H_{\mathrm{prob}}

are mean hard-label and mean probability entropy in bits. “Single” counts single-class target cells; “coverage” is completed/eligible health-audit cells. N/A values are not imputed.

Comparator	Balanced accuracy	Macro-F1	AUROC	Targets / subjects	Holm decision
Exact same-parent layout-aware coordinate control	$+9.96$ [ $+8.20,+11.75$ ]	$+12.31$ [ $+10.22,+14.39$ ]	$+14.88$ [ $+11.95,+17.90$ ]	5 / 253	reject
Graft-EEG static	$+1.06$ [ $+0.58,+1.57$ ]	$+0.87$ [ $+0.35,+1.41$ ]	$+1.86$ [ $+1.40,+2.30$ ]	5 / 253	reject
CSP–LDA	$+10.87$ [ $+9.22,+12.59$ ]	$+25.52$ [ $+23.58,+27.47$ ]	$+16.39$ [ $+14.01,+18.90$ ]	5 / 253	reject
FBCSP–LDA	$+10.86$ [ $+9.18,+12.61$ ]	$+25.49$ [ $+23.55,+27.45$ ]	$+16.14$ [ $+13.80,+18.67$ ]	5 / 253	reject
Riemann–LR	$+10.87$ [ $+9.25,+12.65$ ]	$+25.52$ [ $+23.59,+27.49$ ]	$+16.41$ [ $+13.95,+18.93$ ]	5 / 253	reject
EEGNet	$+10.96$ [ $+9.28,+12.67$ ]	$+15.88$ [ $+13.95,+17.86$ ]	$+16.66$ [ $+14.24,+19.21$ ]	5 / 253	reject

Appendix

Table S7: Secondary subject-paired physical-dynamic-minus-comparator effects in percentage points. Each entry is effect [95% paired subject-bootstrap interval], based on 10,000 replicates. The exact same-parent row uses its registered target-stratified subject bootstrap; the primary dynamic–BIOT and dynamic–static contrasts in the main text instead use dataset–subject hierarchical resampling.

Target	Seed	BA D	BA S	$\Delta$ BA	F1 D	F1 S	AUROC D	AUROC S
BNCI	2027	63.310	64.931	$-1.620$	61.234	63.947	69.688	70.054
BNCI	3407	62.693	64.429	$-1.736$	60.029	63.240	70.720	70.224
BNCI	4513	64.738	64.660	$+0.077$	63.406	63.684	69.076	69.882
BNCI	5623	62.384	64.429	$-2.045$	59.294	63.240	71.056	70.224
BNCI	6737	68.557	66.165	$+2.392$	67.474	65.274	74.157	71.401
Cho	2027	58.073	56.794	$+1.280$	55.579	54.249	62.714	60.348

Appendix

Table S8: Registered target–seed cell metrics for dynamic (D) and static (S) Graft-EEG . BA, macro-F1, and AUROC are percentages;

\Delta

BA is D minus S in percentage points. Every target contains the same five preregistered seeds.

Target	$n$	Dynamic BA	Static BA	BIOT BA	Dynamic–BIOT	$>0$	Dynamic–static	$>0$
BNCI	9	61.94 [56.32,73.89]	64.72 [54.79,78.12]	53.82 [51.39,57.99]	7.29 [4.65,9.31]	77.8%	$-1.04$ [ $-3.26$ ,1.53]	33.3%
Cho	52	56.88 [52.49,62.38]	55.08 [51.46,61.45]	49.90 [47.70,51.38]	7.00 [2.44,13.30]	82.7%	1.25 [ $-0.03$ ,2.74]	73.1%
Lee	54	58.08 [53.12,66.50]	56.27 [52.64,64.75]	51.62 [50.31,53.30]	6.20 [3.10,13.24]	90.7%	0.78 [ $-0.09$ ,1.88]	72.2%
PhysioNet	109	56.25 [51.61,64.35]	54.60 [50.00,62.83]	51.28 [48.51,55.24]	5.89 [0.30,12.87]	75.2%	1.13 [ $-1.07$ ,3.89]	61.5%
Shin	29	61.33 [54.67,71.33]	59.00 [53.33,69.67]	51.67 [48.33,56.67]	10.00 [1.00,18.00]	79.3%	2.33 [0.00,4.67]	72.4%

Appendix

Table S9: Subject-level heterogeneity. Entries are median [Q1,Q3] BA or paired BA differences in percentage points; “

>0

” gives the descriptive proportion of participants with a positive paired difference.

Contrast	Full [95% CI]	omit BNCI	omit Cho	omit Lee	omit PhysioNet	omit Shin
Dynamic–BIOT	8.182 [6.382,10.196]	8.205	8.206	8.279	8.600	7.618
Dynamic–static	1.059 [0.013,1.892]	1.470	0.985	1.106	0.941	0.792

Appendix

Table S10: Leave-one-target-out sensitivity, in percentage points. “Full” is the five-target equal mean with its primary hierarchical 95% interval; the other columns omit the indicated target.

Figure S1: Post-score-seal equal-information FI/SSI controls. a , Active-sensor geometry and native versus quotient-referenced C3, Cz, and C4 traces from the sealed Cho2017:0 replay; 60 of 64 sensors were active, and all traces are shown on the same physical scale. b , Fixed-17 FI and SSI mappings for the same label-blind trial after active-native common-average referencing, mask-specific mapping, fixed-17 common-average referencing, scaling in 100-

\mu

V units, and clipping at

\pm 10

. The lower panel shows Welch relative power spectral density averaged over C3, Cz, and C4 from 1 to 40 Hz. c , Distributions of balanced-accuracy differences for all 253 paired subjects under each matched control. Each subject contributes once per control; colours and symbols denote target datasets, and black bars show the median and interquartile range. d , Target-level paired estimates with 95% subject-bootstrap intervals and equal-target subject-cluster pooled effects from 20,000 bootstrap replicates. Positive values favour dynamic GRAFT-EEG. All controls covered 5/5 targets (253 unique subjects; 1,012 control-specific paired rows); tangent used one registered split seed and EEGNet used five seeds. These post-score-seal controls do not revise the primary inference.

Head	Target	Subjects	Matched control	Graft-EEG dynamic	Effect (pp)
FI + tangent	BNCI	9	59.645	64.336	$+4.691$
	Cho	52	59.032	58.172	$-0.860$
	Lee	54	58.519	60.318	$+1.799$
	PhysioNet	109	58.930	58.247	$-0.683$
	Shin	29	60.862	63.253	$+2.391$
SSI + tangent	BNCI	9	56.327	64.336	$+8.009$

Appendix

Table S11: Target-resolved equal-information controls. Values are subject-mean BA percentages. “Effect” is dynamic Graft-EEG minus the matched control. Tangent controls retain their registered single-seed denominator; EEGNet controls retain five seeds.

Figure S2: All-metric equal-information inference. Each panel reports dynamic Graft-EEG minus one matched equal-information control for balanced accuracy, Macro-F1, AUROC, and AUPRC. Points are equal-target effects and bars are 95% target-stratified subject-cluster bootstrap intervals (20,000 replicates; five targets; 253 subject clusters; audit seed 20260824). Balanced accuracy, shown in bold, was the prespecified primary metric for this supplementary analysis. Hollow circles denote tangent-space controls; filled squares denote EEGNet controls. The exact plotted estimates and intervals are printed at right and are taken directly from the sealed 16-row source table without refitting or resampling.

Figure S3: Calibration and registered stress. a , Trial-pooled 15-bin reliability curves reconstructed from all 25 sealed target–seed cells, together with equal-target NLL, Brier, and ECE. b , Target-resolved BA changes in percentage points under the registered stress families; hatched entries are structural N/A. c , The five target effects in percentage points within each family, with the equal-target mean and 95%

t

interval. d , The seed-cell execution ledger; structural N/A entries retain their zero numerator and are not converted into effects.

Target	NLL raw	NLL cal.	Brier raw	Brier cal.	ECE raw	ECE cal.
BNCI	0.6329	0.6378	0.2212	0.2233	0.0649	0.0760
Cho	0.6883	0.6746	0.2457	0.2404	0.0741	0.0342
Lee	0.6597	0.6568	0.2338	0.2325	0.0459	0.0317
PhysioNet	0.6953	0.6754	0.2479	0.2408	0.0844	0.0374
Shin	0.6394	0.6480	0.2243	0.2281	0.0390	0.0578
Equal-target	0.6631	0.6585	0.2346	0.2330	0.0617	0.0474

Appendix

Table S12: Target-resolved source-only calibration. Lower is better for NLL, Brier, and ECE. Each entry averages five frozen target cells; calibration parameters were fitted on source predictions only.

Condition family	BNCI	Cho	Lee	PhysioNet	Shin	Equal-target	Exec.
Random channel drop 25%	$-1.22$	$-1.03$	$-0.36$	$-0.46$	$-4.01$	$-1.42$	75/75
Random channel drop 50%	$-7.38$	$-2.11$	$-3.71$	$-0.33$	$-6.20$	$-3.95$	75/75
Spatial contiguous drop 25%	$-6.76$	$-6.50$	$+1.25$	$-3.35$	$-10.75$	$-5.22$	25/25
Noise +10 dB	$+0.43$	$-0.46$	$-0.66$	$-0.31$	$-2.79$	$-0.76$	25/25
Noise 0 dB	$-3.46$	$-2.42$	$-3.34$	$-1.87$	$-7.71$	$-3.76$	25/25
Noise $-10$ dB	$-11.84$	$-6.27$	$-7.77$	$-5.28$	$-13.48$	$-8.93$	25/25

Appendix

Table S13: Target-resolved robustness family matrix. Values are dynamic BA change in percentage points relative to the unperturbed target cell. “Exec.” is complete/registered seed cells. Structural N/A is not coded as zero.

Figure S4: Legacy stability and execution boundaries. This registered diagnostic is separate from the exact same-parent physical-leadfield control in Fig. 3 . a , Physical-minus-comparator cross-seed standard-deviation contrasts for coordinate and legacy no-physics diagnostics. All four estimates are positive; negative would mean that the physical model is more stable, so the observed direction is adverse rather than supportive evidence. b , Seed-level mean BA-change fingerprints for random dropout, spatial dropout, noise, and reference families. c , Completed/registered cells by target and condition family. Structural N/A remains hatched with its zero numerator and is never imputed as an effect.

Strategy	Labels	BNCI	Cho	Lee	PhysioNet	Shin	Equal-target
Frozen linear probe	1%	58.939	52.879	56.608	53.211	54.060	55.139
	5%	61.231	56.606	59.947	53.949	57.075	57.762
	10%	63.718	57.856	60.804	54.934	58.774	59.217
Full GRAFT head fine-tune	1%	66.102	58.309	60.736	57.776	63.692	61.323
	5%	65.917	58.424	61.163	57.915	63.964	61.477
	10%	66.887	58.821	61.398	57.888	63.614	61.722

Appendix

Table S14: Target-resolved few-shot BA (%). Label fractions denote target trials used for adaptation; all remaining trials form a disjoint evaluation set.

Strategy	Labels	BA	Macro-F1	AUROC	ECE	Cells
Frozen linear probe	1%	55.139	54.872	57.149	28.017	25/25
	5%	57.762	57.713	60.836	19.790	25/25
	10%	59.217	59.202	63.131	13.730	25/25
Full GRAFT head fine-tune	1%	61.323	61.189	66.096	4.772	25/25
	5%	61.477	61.379	66.221	3.898	25/25
	10%	61.722	61.701	66.594	3.477	25/25

Appendix

Table S15: Equal-target few-shot metrics (%). ECE is the 15-bin expected calibration error for the positive class. Each row summarizes 25 completed target–seed cells.

Figure S5: One-shot external evaluation on 60 held-out participants. a , C3/Cz/C4 waveforms from six label- and prediction-blind SHA-selected real Dreyer2023A trials. b , Welch spectra of the same six trials. c , The prepared 27-channel montage; the prepared export has no per-trial mask. d , Paired dynamic and static BA for all 60 participants. e , The complete participant-level dynamic–static distribution with the sealed mean and interval. f , Two preregistered contrasts and descriptive absolute BA estimates. The dynamic–static contrast is inconclusive; the physical-operator contrast is positive.

Result type	Quantity	Estimate	Lower	Upper	Decision
Absolute BA	Graft-EEG dynamic	62.359	59.562	65.239	descriptive
Absolute BA	Static physics	61.789	59.033	64.638	descriptive
Preregistered contrast	Dynamic $-$ static	$+0.571$	$-0.082$	$+1.230$	inconclusive
Preregistered contrast	Full physics $-$ no physics	$+4.029$	$+2.338$	$+5.773$	positive

Appendix

Table S16: Dreyer2023A one-shot external summary. Absolute BA rows are descriptive; the two contrast rows report the preregistered decisions. Values are percentages or percentage points with 95% intervals.

Figure S6: Jansen–Rit validation with known sources. a , Locked source trajectories and matched physical, row-shuffled, and spectrum-matched recoveries for one held-out BNCI seed-2027 trial at a 22-channel, 0-dB condition. b , A 25-ms lag embedding of the same left-M1 trajectories. c , Mean source correlation across registered retained-channel and SNR conditions; each cell reports the mean and completed/25 count, hatching retains structural N/A, and orange area encodes the fraction entering the preregistered scientific failure region. d , Five-target stress trajectories at a 22-channel budget across the five sealed seeds. e , Physical-operator advantage over row-shuffled and spectrum-matched controls by target; intervals are descriptive 20,000-draw bootstrap intervals. f , Failure-region placement for the 705 executed physical-operator conditions: 600 channel/noise conditions, 25 physical negative controls, and 80 reference conditions; 473/705 entered the preregistered failure region. The full executed ledger additionally contains 25 row-shuffled and 25 spectrum-matched-random negative controls, all in the failure region, giving 523/755 overall. All trajectories and estimates come from the sealed known-source archive. The experiment evaluates the explicit inverse on synthetic known sources and does not imply that the frozen real-EEG classifier recovered cortical sources or supports clinical localization.

Target	Block	Registered	Executed	Structural N/A	Recoverable	Failure
BNCI	Channel/noise	150	90	60	15	75
	Negative control	15	15	0	0	15
	Reference	20	15	5	15	0
Cho	Channel/noise	150	150	0	45	105
	Negative control	15	15	0	0	15
	Reference	20	15	5	15	0

Appendix

Table S17: Target-by-block accounting for the 925 Jansen–Rit conditions. “Recoverable” and “failure” partition the 755 executed conditions; structural N/A remains in the registered denominator.

Block	Registered	Executed	N/A	Recoverable	Failure
Channel/noise	750	600	150	152	448
Negative control	75	75	0	0	75
Reference	100	80	20	80	0
Total	925	755	170	232	523

Appendix

Table S18: Known-source execution totals. The complete channel-budget-by-SNR grid is shown in Figure S6 c and retained without rounding in the corresponding source-data table. “Recoverable” and “failure” partition executed conditions; structural N/A remains in the registered denominator.

Figure S7: Label-blind replay of a real retained trial. a , Native 200-Hz Cho2017 trial Cho2017:0 , its 64-channel geometry, trial-level hard-QC mask, and prespecified C3/Cz/C4 traces; 60 channels remain active. b , The same active signals before and after gauge projection, followed by the frozen active operator

A_{S}

,

B=P_{S}A_{S}

, and

G_{\rho}

. c , The resulting real

8\times 600

transported latent. d , The exact

11\times 24

windowed spectral sequence supplied to the frozen SSM. e , Mean decay, input gate, hidden-state norm, and a read-only prefix replay of the bounded residual and logit. Prefixes 1–10 are diagnostic only; the final point is the sole canonical scored output and matches the sealed logit within

1.49\times 10^{-8}

. No label or outcome was used to choose the dataset, trial, channels, time range, axes, or colour limits.

Figure S8: Source-fitted parameter anatomy for one prespecified held-out-target fold. a , Standardized logistic-regression coefficients across eight motor modes and three spectral bands. b , Condition-neutral state retention implied by the 16 frozen base-decay parameters. c , Frozen candidate, gate, and input-conditioned decay weights, with inputs ordered by band and motor mode. d , Frozen weights on terminal and sequence-mean states in the bounded residual readout. These panels describe stored source-fit parameters; they are not target-derived attribution maps or performance estimates.

Figure S9: Three-dimensional geometry-conditioned inverse kernels. The eight panels show rows L1–L4 and R1–R4 of the locked Cho2017 inverse operator

G_{\rho}

for one prespecified active mask (60 of 64 native channels). Sensor positions follow the recorded head-frame montage. Active sensors are colored by their corresponding

G_{\rho}

coefficients, and the scalp surface interpolates the same values using one max-absolute normalization over the complete

8\times 60

matrix. Ring-and-cross markers identify the four inactive channels. The ellipsoid, nose, ears, lighting, and shadows provide orientation only and are not participant anatomy. Colors represent inverse-kernel coefficients—not learned Spat-Conv weights, cortical activity, source localization, statistical significance, or participant-specific anatomy.

Figure S10: Real-EEG atlas, deterministic rank 1. One SHA-selected trial from each of BNCI2014-001, Cho2017, Lee2019-MI, Shin2017A, and PhysionetMI. Every dataset row shows the recorded montage and active set, three available central-motor sensors, the complete native-channel input, and the spectrum of the same trial. Shaded bands mark 8–13 and 13–30 Hz; no sleep-stage, event, ERD/ERS, or source-localization label is inferred.

Figure S11: Real-EEG atlas, deterministic rank 2. The second independently SHA-ranked subject/trial for each dataset, rendered under the same locked channel, time, spectrum, colour-limit, and annotation rules as Figure S10 . Trial identifiers and selection hashes are printed in the figure and retained in the source manifest.

Figure S12: Real-EEG atlas, deterministic rank 3. The third SHA-ranked subject/trial for each dataset. Together, Figures S10 – S12 show 15 real trials with complete trial IDs, native channel counts, selected sensor names, display limits, waveforms, and Welch source data.

Figure S13: Strict target–method matrix and baseline-health diagnostics. a , Target-resolved BA for the 15 methods with a clean 5/5 denominator, followed below the rule by EEGPT and LaBraM with four clean targets and CodeBrain with unresolved checkpoint provenance. CodeBrain hatching denotes provenance withholding rather than a structural gap; partial denominators are not rank-mixed. b , Source validation, target positive rate, hard- and probability-entropy, completed eligible cells, and single-class counts. Hatched N/A remains missing and is never replaced by zero.

Figure S14: Coordinate-perturbation sensitivity from sealed real-data evaluations. a , Target-resolved dynamic BA change for independent tangent and rigid head-frame perturbations; T and R identify the families and the number gives millimetres. b , Magnitude profiles; points denote the five targets and lines equal-target means after averaging seeds. c , Reduced-operator relative Frobenius change versus dynamic BA change across 50 target–condition units (Spearman

\rho=-0.449

; target-cluster 95% CI

-0.760

to

-0.282

; target-stratified permutation

P<0.001

). d , Execution boundary. No training was performed, tangent and rigid conditions were not pooled, and the displayed operator change is for reduced

A

, not the full forward matrix

G

.

Family	Magnitude	Mean $\Delta$ BA	Target range	Mean operator change	Artifact origin
Independent tangent	1 mm	$-0.055$	[ $-0.195,+0.045$ ]	0.0154	new extension
Independent tangent	2 mm	$-0.157$	[ $-0.610,+0.107$ ]	0.0309	sealed read-only reuse
Independent tangent	5 mm	$-0.393$	[ $-1.165,+0.142$ ]	0.0763	sealed read-only reuse
Independent tangent	10 mm	$-0.570$	[ $-1.736,+0.341$ ]	0.1511	sealed read-only reuse
Independent tangent	20 mm	$-1.064$	[ $-3.935,+0.371$ ]	0.2964	new extension
Rigid head-frame	1 mm	$-0.058$	[ $-0.115,-0.009$ ]	0.0155	new extension

Appendix

Table S19: Equal-target coordinate-sensitivity summary. BA changes are percentage points from the frozen unperturbed parent; ranges are across the five held-out targets after averaging seeds. The operator column is the target mean of

\lVert\Delta A\rVert_{F}/\lVert A\rVert_{F}

.

Component	Setting	Selection / role
Preprocessing	notch; 0.5–45 Hz; 200 Hz	fixed for all datasets
Epoch	$[0.5,3.5)$ s; 600 samples	fixed
Hard channel gate	nonfinite; $>5000~\mu\mathrm{V}$ ; $<0.1~\mu\mathrm{V}$ p–p	fixed before inference
Trial gate	$\leq 20\%$ inactive; at least 9 active	fixed before inference
Motor latent	4 left + 4 right modes	fixed dimension $d=8$
Transport ridge	$\rho\in\{10^{-6},10^{-4},10^{-2},10^{-1}\}$	source-selected

Appendix

Table S20: Fixed implementation and selection settings. Any quantity marked “source-selected” is optimized without target signal, label, mask distribution, or empirical statistic.

Method	BA (%)	Trainable	Fit time (s)	Timed cells	Scope / device
Graft-EEG dynamic	60.865	1,281	289.93 [271.56, 413.96]	25	joint static–dynamic / CUDA
Graft-EEG static	59.806	N/R	289.93 [271.56, 413.96]	25	same joint record / CUDA
EEGConformer	50.448	64,482	56.19 [35.06, 67.25]	25	selection + freeze / CUDA
EEGNet	49.901	1,714	30.94 [24.57, 37.40]	25	selection + freeze / CUDA
CSP–LDA	50.000	3	—	—	comparable timer absent
FBCSP–LDA	50.005	9	—	—	comparable timer absent

Appendix

Table S21: Available downstream parameter counts and source-fit times. Times are seconds per target–seed cell, reported as median [Q1,Q3]. Static and dynamic share one joint fit record and must not be summed. A dash marks unavailable comparable timing, not zero compute.

Step	Process	May read	Must emit	Failure rule
1	Source fit	source signals/labels; declared geometries	checkpoint + fit metadata	target store unavailable
2	Fit barrier	all expected checkpoints	global fit manifest	any missing/hash drift stops run
3	Blind inference	sealed fits; target signals; trial masks	probabilities/logits + queue IDs	labels unavailable
4	Raw barrier	all expected raw predictions	raw-prediction manifest	any missing/hash drift stops score
5	One-shot join	raw IDs; isolated target labels	paired score source	no fit artifact is writable
6	Scorer	paired source; fixed metrics	cell metrics + subject metrics	no tuning or threshold selection

Appendix

Table S22: Fail-closed execution and access separation. Each stage may read, but cannot rewrite, outputs accepted at an earlier gate.

Object	Scientific role	Rows / cells	SHA-256 prefix
Figure2 subject pairs	Dynamic–BIOT held-out subjects	253 rows	0096b584532b
Figure2 benchmark matrix	18 methods $\times$ 5 targets	90 rows	c53c30d75146
Figure2 target effects	two contrasts $\times$ 5 targets	10 rows	080901a117ed
Hierarchical summary	20,000-replicate primary inference	1 JSON	ad5334718806
Strict target table	strict 15-method target matrix	15 rows	a46860b6bc08
Secondary paired inference	10 comparators $\times$ 3 metrics	30 rows	7a2d26ab3c01

Appendix

Table S23: Fourteen principal objects among the 17 immutable inputs behind the manuscript tables. The accompanying JSON manifest enumerates all 17 inputs and stores full digests, byte counts, paths, and aggregation rules; 12-character SHA-256 prefixes are shown here.

Method family	T/S/L	Tgt. tune	Common ch.	Tgt. geometry	Pretraining	Executed source fit	Trainable	Tgt.
Graft-EEG static	none	no	no	pre, geometry only	none	25 cells	N/R	5/5
Graft-EEG dynamic	none	no	no	pre, geometry only	none	25 cells	1,281	5/5
BIOT + LR	none	no	no	fixed 18-map	clean	25 cells	257	5/5
CBraMod frozen + LR	none	no	no	checkpoint map	clean	25 cells	201	5/5
CBraMod source-only full FT	none	no	no	checkpoint map	clean	25 cells	full ckpt.	5/5
EEGPT + LR	none	no	no	fixed 19-map	overlap 1/5	25 cells	513	4/5

Appendix

Table S24: Protocol-fairness ledger for access, adaptation, geometry, pretraining provenance, compute budget, and target denominator. Parameter counts include trainable downstream weights only; “pre” and “post” are defined relative to the model-fit lock.

Paper role	Code identifier	Equal-target BA
Main dynamic method	mosaic_dynamic	60.865%
Static anchor	mosaic_static	59.806%
Primary pretrained comparator	biot	52.684%
Exact same-parent layout-aware coordinate control	mosaic_dynamic_exact_no_physics_missing_v2	50.907%
Matched-spectrum random operator	matched_spectrum_random	51.167%
Electrode-row-shuffled operator	electrode_row_shuffle	49.624%

Appendix

Table S25: Non-interchangeable result identities used in the manuscript.

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

Abstract

Figures & tables

Appendix

Cross-Session EEG Decoding

Brain foundation model-guided source-selective domain adaptation for cross-subject EEG decoding

Cortical-SSM: A Deep State Space Model for Motor Imagery Decoding from EEG Signals

Lightweight Adaptation of EEG Foundation Models for Stroke Motor Imagery Decoding: Domain Shift and Subject-Level Robustness

Abstract

Cross-Session EEG Decoding

Explore similar work

Brain foundation model-guided source-selective domain adaptation for cross-subject EEG decoding

Cortical-SSM: A Deep State Space Model for Motor Imagery Decoding from EEG Signals

Lightweight Adaptation of EEG Foundation Models for Stroke Motor Imagery Decoding: Domain Shift and Subject-Level Robustness