PaperlayerView exact revision on arXiv

Source-linked AI summary

PhysioTRACE: Provenance-Aware Stress Tests for Physiological Foundation Models

Ayana Mussabayeva, Anuar Aimoldin, Olivier Oullier, Xue Liu, Kun Zhang

arXiv:2609.34466v1cs.LG

TL;DR

Recording provenance can be decoded from physiological foundation-model embeddings, but that does not show whether a task head uses it. PhysioTRACE tests reliance by changing only the provenance–target association and selectively certifies removal; its results show that accessibility and out-of-distribution utility are not consistent proxies for reliance.

  • Problem

    Shifted transfer and provenance probing do not establish whether a task head actually relies on recording provenance, although that distinction matters when acquisition conditions are associated with diagnosis.

    Source: §1 · pp. 1–2

  • Method

    PhysioTRACE reverses the provenance–target association on the same held-out records with fixed heads, then tests train-localized removal against matched random controls and a declared utility margin.

    Source: §1 · p. 2, §3 · p. 5, §5 · p. 10

  • Results

    The Recover–OOD utility relation changes sign across benchmarks, while ECG audits certify one remedy that removes 91% of ECGFounder’s excess vulnerability and return all three verdicts.

    Source: §1 · p. 3, §5 · p. 10

  • Takeaways & Limitations

    Selective verification matters because removing decodable provenance by default can also remove target-relevant signal, whereas certified remedies preserve utility within a declared margin.

    Source: §5 · p. 10

    The certified remedy supports the evaluated association and linear intervention, not safety under every deployment shift; regression and forecasting remain untested extensions.

    Source: §5 · p. 10

Abstract

from arXiv · show

Physiological foundation models encode how a signal was recorded alongside the physiology it reflects. When recording conditions are associated with diagnosis, this acquisition provenance can become a shortcut, yet the usual evidence, shifted transfer and provenance decodability, does not show whether a predictor uses it. We introduce PhysioTRACE, a four-axis behavioral audit for frozen encoders that separates what a probe can decode from what a fixed task head relies on. Recover scores how decodable provenance is; Stress reverses only the provenance-target association on the same held-out records; Intervene removes a train-localized provenance component; and Verify certifies that removal only if it beats matched random projections within a declared utility margin. Each audit thus ends in one of three verdicts: no reliance, or reliance with the remedy certified or refused. Across EEG and ECG, five training objectives, and five frozen foundation models, the relation between Recover's calibrated score and out-of-distribution utility changes sign between datasets, so neither can stand in for a reliance test. On paired EEG views where the shortcut is known by construction, the audit detects it (the exposed head loses about 0.2 AUROC when the association is reversed, while a control head is unaffected) and certifies removal of a rank-two component that restores control-level behavior without measurable utility loss, for both encoder objectives tested. On real ECG device metadata it returns all three verdicts: it certifies a remedy that removes 91% of one model's excess vulnerability, finds no reliance where device and diagnosis are barely associated, and refuses the remedy for a second model whose localized direction also carries task signal. Robustness to how inputs were recorded therefore needs a behavioral test, and PhysioTRACE provides one that can pass, fail, or refuse a remedy.

1 Introduction

Physiological foundation models retain acquisition provenance, but common transfer and probing measurements do not establish whether a task head relies on it. PhysioTRACE tests fixed-head reliance and evaluates whether a train-localized remedy can remove it without sacrificing utility.

1 Introduction

  • The paper positions reusable frozen encoders as useful only if task heads read physiology rather than recording circumstances.The proposed audit is released with leakage-aware splits, calibrated probes, stress weights, localization operators, and machine-readable results.Source: §1 · pp. 1, 3
  • PhysioTRACE reverses only the provenance–target association on the same held-out records with task heads fixed, then tests whether train-localized removal reverses the failure without costing ordinary utility.Removal is certified only when it beats matched random controls and stays within a declared utility margin.Source: §1 · pp. 2–3
  • Across benchmarks, the relation between provenance accessibility and out-of-distribution utility changes sign, so neither measurement establishes whether a task head relies on provenance.ERP/P300 shows a negative relation, PTB-XL reverses it, and HMC sleep staging decouples the two.Source: §1 · p. 3
  • On paired EEG views, reversing the reference–diagnosis association lowers exposed-head AUROC by 0.20–0.22, while control heads are unaffected for both encoder objectives.Removing a rank-two subspace restores control-level behavior without changing in-domain or cross-reference AUROC.Source: §1 · p. 3
  • On observational ECG, the audit certifies a remedy removing 91% of ECGFounder’s excess vulnerability, finds no reliance for a barely associated device pair, and refuses CLEF-Small’s remedy.The CLEF-Small localized direction also carries task signal; removing it by default would have cost 0.039 AUROC.Source: §1 · p. 3

2 Related Work and Positioning

PhysioTRACE distinguishes accessible provenance from behavioral reliance and positions its fixed-head audit against shifted-transfer, probing, and concept-erasure approaches. It also incorporates acquisition metadata into provenance groups for physiological benchmarks.

PhysioTRACE positioning

  • The four audit axes separate provenance decodability, fixed-head stress, train-localized intervention, and verification against matched controls and a utility margin.The algorithm returns no reliance, reliance with a certified remedy, or reliance with a remedy not certified.Source: Figure 2 · p. 4, §2 · p. 5

Probing and usable information

  • A probe score measures accessibility rather than use, so decodable provenance alone does not establish that a task head relies on it.LPI is a calibrated linear instance of predictive V-information for acquisition provenance, estimated per native recording unit with identity-aware uncertainty.Source: §2 · pp. 3–4

Spurious correlations and stress tests

  • Unlike training or head-refitting approaches, PhysioTRACE changes only the provenance–target association at fixed marginals on held-out records and compares fixed exposed and control heads.This comparison is designed to isolate association reliance from ordinary distribution shift.Source: §2 · p. 4

Concept erasure

  • Concept erasure can remove target-relevant population signal and does not by itself establish head reliance or preserved task utility.PhysioTRACE adds train-localized suppression and certifies it only when it beats matched random controls and stays within a declared utility margin; remedies can be refused.Source: §2 · p. 4

Physiological benchmarks and metadata

  • Physiological benchmarks often fold acquisition into domain labels, while PhysioTRACE turns reference, device, site, and preprocessing metadata into provenance groups with identity-disjoint splits.The paper builds on BIDS and dataset-documentation frameworks.Source: §2 · p. 4

3 The PhysioTRACE Protocol

PhysioTRACE audits frozen encoders by separating provenance accessibility from task-head reliance, then testing and selectively suppressing reliance. Verification requires both specificity against matched random projections and utility preservation within a declared margin.

3 The PhysioTRACE Protocol

  • PhysioTRACE takes an identity-disjoint train/test split, freezes exposed and control task heads, and compares them under declared provenance-factor roles.The exposed head is fitted under an association; the control head uses independence weights, and shared behavior is not attributed to that association.Source: §3 · p. 5

3.1 Recover: Linear Provenance Information

  • LPI measures calibrated linear provenance accessibility per native unit, with identity-aware bootstrap intervals rather than treating trial views as independent.On ERP, trial resampling would narrow intervals 14–31-fold; LPI depends on its probe family and calibration and does not show that a task head uses provenance.Source: §3.1 · pp. 5–6

3.2 Stress: fixed-marginal association shift

  • Stress reweights the same held-out records across fixed-marginal regimes and measures excess vulnerability as the exposed head’s sensitivity beyond the control’s.Because records, labels, embeddings, and heads remain identical across regimes, G measures reliance on the association rather than ordinary distribution shift.Source: §3.2 · p. 6

3.3 Intervene: train-localized suppression

  • Intervene removes a train-localized component while leaving directions outside it unchanged, using paired-view contrasts or target-conditioned observational localization.Paired localization uses no task labels; observational localization conditions on target because provenance groups may differ in target composition.Source: §3.3 · p. 6

3.4 Verify: specificity, utility, and verdict

  • Verify certifies suppression only when its reduction in excess vulnerability beats matched random subspaces and ordinary-utility loss stays within the declared margin.The audit returns no reliance, reliance with a certified remedy, or reliance with a remedy not certified; reduced decodability alone does not enter the decision.Source: §3.4 · p. 7

4 Applying PhysioTRACE

Across benchmark transfer, paired-reference EEG, and observational ECG device metadata, accessibility and transfer do not establish head reliance, while the audit distinguishes reliance, certified remedies, and refused remedies.

4.2 Q2: Does the audit recover a known shortcut?

  • Under induced reference association, exposed heads degrade from aligned to reversed (G = 0.20–0.22) while controls are unaffected; exposed decisions flip for 44.5% of records versus 13.8% for controls.The reliance follows the association under which the head is fitted, not the reconstruction objective.Source: §4.2 · pp. 7–8

4.1 Q1: Can accessibility or transfer stand in for a reliance test?

  • Across ERP/P300 and PTB-XL, provenance accessibility and OOD transfer decouple, with their relationship changing sign between datasets; neither measures whether a head relies on provenance.Figure 3 compares LPI with OOD AUROC; LPI is comparable only within a dataset.Source: §4.1 · p. 7, Figure 3 · p. 8

4.2 Q2: Does the audit recover a known shortcut?

  • Removing the rank-two reference subspace brings G within 0.003 of zero and restores control-level behavior, while matched random subspaces leave G unchanged and AUROC changes by less than 0.002.Both tested objectives are certified.Source: §4.2 · p. 8, Figure 4 · p. 9

4.3 Q3: Does the audit work on observational data, and can it refuse?

  • For ECGFounder, removing the rank-one direction reduces G by 91% beyond all random controls, flattens responses across association strengths, and stays within the utility margin.The remedy removes only the component the head relies on, leaving other device-associated signal in place.Source: §4.3 · p. 9
  • For a site-2 device pair with odds ratio 1.04, stress yields G = 0.00007, below ε and not bounded away from zero, so the audit stops before Intervene.The stress procedure does not manufacture excess vulnerability where the head has no association to exploit.Source: §4.3 · p. 9
  • For CLEF-Small, rank-one removal reduces G by 63% but costs 0.039 AUROC against a 0.01 margin, so Verify refuses the remedy despite its beating random controls.The localized direction carries task signal together with device signal.Source: §4.3 · p. 10

5 Discussion, Scope, and Conclusion

PhysioTRACE treats provenance robustness as a behavioral question: transfer and probe scores cannot establish whether a fixed head relies on acquisition conditions. Its conclusions are bounded by the evaluated associations, interventions, endpoints, and data settings.

5 Discussion, Scope, and Conclusion

  • Transfer and probe scores cannot establish reliance: PhysioTRACE tests fixed-head behavior, and a certified remedy supports the evaluated association and linear intervention—not safety under every deployment shift.The audit can also refuse a remedy when localization captures task signal.Source: §5 · p. 10

Ethics statement; Reproducibility statement

  • The experiments reanalyze published datasets and existing checkpoints; no participants were recruited, new human-subject data collected, or diagnostic models deployed.Code, benchmark schemas, experiment runners, and saved numerical artifacts are available from the project repository.Source: Ethics statement · p. 11, Reproducibility statement · p. 11

A.1 Shifted-utility reporting

  • Shifted utility for mixed-provenance heads reports mean performance across target groups and worst-group utility separately from source-specific estimates.The score is computed across target groups; worst-group utility is the minimum target-group score.Source: §A.1 · p. 18

A.2 Fixed-marginal stress and reliance estimands

  • Stress regimes keep target and provenance marginals fixed to the training design; observational ECG regimes vary device–target odds from aligned through reversed.The regimes include aligned, independent, and reversed associations, with intermediate regimes for ECG.Source: §A.2 · p. 18
  • The paired reliance statistic is a class-balanced, same-record logit contrast under provenance change while physiology and diagnosis remain fixed.It is reported before and after intervention.Source: §A.2 · p. 19

A.3 Train-localized suppression

  • Paired localization uses training records and provenance views, without diagnosis labels, validation or test embeddings, or task-head weights.For NMT, the predeclared rank is K − 1; observational ECG localizes device information in target-residualized training embeddings.Source: §A.3 · p. 19
  • The projection suppresses a localized component in the full standardized embedding, while the already-fitted task heads remain fixed and need no target or provenance labels at application.The operator builds on linear concept-erasure methods; PhysioTRACE adds acquisition-specific localization and behavioral verification.Source: §A.3 · p. 19

A.4 Verification criteria

  • Verification requires targeted suppression to beat rank- and energy-matched random controls and the lower 95% confidence bound for utility change to remain within the declared margin.The criteria support selective suppression of a behaviorally relevant linear component, not complete linear or nonlinear invariance.Source: §A.4 · p. 20

A.5 Relation to linear erasure and amnesic probing

Table 2 compares PhysioTRACE with INLP, Amnesic Probing, and LEACE. The paper distinguishes the reused projection operator from the audit elements it adds.

A.5 Relation to linear erasure and amnesic probing

  • The projection operator itself is not new.PhysioTRACE reuses a projection approach rather than claiming a novel operator.Source: §A.5 · p. 20
  • PhysioTRACE adds the specification of what to stress, which comparator to use, and what qualifies as a certified remedy.These elements define the behavioral audit beyond the projection operator.Source: §A.5 · p. 20
  • Table 2 compares PhysioTRACE with INLP, Amnesic Probing, and LEACE.The comparison concerns the methods’ relative features.Source: Table 2 · p. 21

B Worked example: one NMT record through the four axes

The worked NMT example shows how Recover, Stress, Intervene, and Verify distinguish accessible reference information from a shortcut used by the task head. A rank-two intervention removes the observed reliance while preserving measured utility.

Recover

  • 93.3% reference-probe accuracy on 555 held-out views (chance 33.3%) and LPI = 1.31 [1.22, 1.38] bits per view show reference accessibility, not decision reliance.The prevalence baseline is log2 3 = 1.585 bits per view.Source: B Worked example: one NMT record… · p. 21

Stress

  • 0.828 to 0.599 AUROC: the exposed head degrades as the association reverses, while the control head stays at 0.768 across aligned, independent, and reversed regimes.The resulting excess vulnerability is G = 0.228.Source: B Worked example: one NMT record… · p. 21

Intervene

  • A rank-two basis removes 2/96 = 2.1% of localization energy; after suppression, the exposed head assigns all three views of record 513 normal probabilities.The refitted reference probe falls to 37.8%.Source: B Worked example: one NMT record… · p. 21

Verify

  • G falls from 0.228 to −0.002, while the ten matched random subspaces leave it between 0.221 and 0.231; mean in-domain AUROC changes by +0.002 [−0.003, 0.008], within the −0.02 margin.The example therefore meets the stated verification criteria for this encoder.Source: B Worked example: one NMT record… · pp. 21–22

C Calibrated accessibility on ERP/P300 and PTB-XL

The calibrated LPI grid measures held-out provenance accessibility, not task-head reliance. Its relationship with transfer and its uncertainty depend on dataset and physical-identity-aware analysis.

Coverage and inference scope

  • The Recover grid covers 55 cached-embedding conditions across 15 model–benchmark summaries, but its analysis does not establish task-head reliance by itself.The locked analysis retains negative estimates and does not retrain encoders or task heads.Source: C Calibrated accessibility on ERP/P300… · pp. 22, 25

Measure and interpretation

  • LPI is a held-out log-score gain over the training-prevalence baseline, not exact mutual information or evidence that a task head uses provenance.Zero or negative LPI means this probe procedure did not improve on that baseline for the evaluated support, not that every decoder lacks provenance information.Source: C Calibrated accessibility on ERP/P300… · p. 22

Results and sensitivity

  • ERP and PTB show no universal accessibility–transfer ordering: masked reconstruction has the largest ERP LPI across five seeds, while PTB patterns motivate joint reporting.The authors describe the ERP pattern as descriptive, not a causal ablation or population-level superiority test; PTB site analysis is separate from the ECGFounder device audit.Source: C Calibrated accessibility on ERP/P300… · p. 23

Sensitivity

  • ERP leave-one-subject-out LPI ranges vary across four subject omissions, and BIOT’s positive point estimates do not override its zero-crossing cluster interval.These are sensitivity ranges, not confidence intervals or refitted folds.Source: C Calibrated accessibility on ERP/P300… · p. 23

Identity-aware uncertainty

  • ERP identity-aware intervals are 14–31 times wider than naive intervals, separating 9 of 28 model pairs rather than 27; PTB intervals are 1.1–1.5 times wider.BIOT’s naive interval excludes zero, while its reported identity-aware interval includes it.Source: C Calibrated accessibility on ERP/P300… · p. 24, Table 6 · p. 24

D Complete broad transfer screens

The broad transfer screens report transfer and classification-probe results separately from the calibrated LPI grid. The mixed-reference exposure baseline is not comparable with source-held-out gaps.

Interpretation boundary

  • These transfer and classification-probe tables enter no audit verdict, and objective-family comparisons are not causal ablations of reconstruction.Reported standard deviations across five fits are distinct from the cluster-bootstrap intervals used elsewhere.Source: D Complete broad transfer screens · p. 25

Mixed-reference exposure

  • The mixed-reference exposure baseline reaches 0.681 ± 0.004 mean-target BA and 0.674 ± 0.004 worst-target BA, but its gap is not comparable with source-held-out gaps.Every target reference is seen in training.Source: D Complete broad transfer screens · p. 25

EEG transfer screens

  • The broad screens report balanced accuracy across three EEG reference views for NMT and domain-transfer summaries across three ERP/P300 domains; frozen external EEG encoders are included.LaBraM, BIOT, and CBraMod are external frozen-encoder audits, not trained objective-family baselines.Source: Table 8 · p. 26, Table 9 · p. 26

PTB-XL cross-site screen

  • The PTB-XL cross-site screen reports task AUROC and multiclass site-probe accuracy for ResNet runs and frozen ECG encoders, separately from the device audit.ECGFounder and MERL are frozen encoders; their site-probe balanced accuracies are 0.563 and 0.599, respectively.Source: Table 10 · p. 26

E Five-seed paired NMT audit

The paired NMT audit tests fixed heads under a controlled reference–diagnosis association and evaluates targeted removal against random controls and a utility margin. Targeted rank-two removal eliminates measured excess vulnerability for both objectives while preserving ordinary AUROC.

Paired design and localization

  • The audit uses five-seed EEGNetSmall encoders with 128-dimensional embeddings and fixed heads; reference association is stressed on the same held-out records.Train-only, equally weighted within-record reference contrasts localize a rank-two basis without diagnosis labels or task-head weights entering localization.Source: E Five-seed paired NMT audit · p. 27

Recovery

  • Reconstruction mean G falls from 0.2249 [0.1928, 0.2557] to 0.0007 [−0.0019, 0.0036], while task-only G falls from 0.1973 [0.1727, 0.2214] to −0.0005 [−0.0029, 0.0017].The paired reductions are 0.2242 [0.1923, 0.2542] and 0.1978 [0.1730, 0.2221], respectively; targeted residual G beats matched random controls.Source: E Five-seed paired NMT audit · p. 27, Table 11 · p. 28

Utility preservation

  • Reconstruction in-domain AUROC changes from 0.7774 to 0.7788 (+0.0014 [−0.0015, 0.0050]), satisfying the prespecified −0.02 utility margin.Task-only changes by +0.0001 [−0.0035, 0.0040], with all five per-seed checks passing; reconstruction cross-reference AUROC changes by +0.0005 [−0.0018, 0.0029].Source: E Five-seed paired NMT audit · pp. 27–28

Scope

  • The induced association is a controlled positive stress test, not an estimate of clinical deployment prevalence.The between-encoder comparison also includes separately selected head regularization, so it is not a reconstruction-only causal contrast.Source: E Five-seed paired NMT audit · p. 28

Recover analysis

  • Both objectives retain substantial linearly accessible reference provenance, but the reconstruction-minus-task LPI difference is seed-dependent and its crossed interval includes zero.LPI measures the fitted linear probe’s log-score gain, not exact mutual information or task-head use.Source: E Five-seed paired NMT audit · pp. 28–29

F Regime-level NMT results for both objectives

Across the three reference-association regimes, the exposed heads gain over controls when aligned and lose under independence; targeted removal rescues reversed-regime AUROC for both objectives.

Measurement

  • Regime summaries average five-seed weighted held-out AUROC over three shortcut-reference rotations on the same 185 records; flip rate measures cross-view prediction changes.Random rows average ten rank-two, energy-matched Haar subspaces per encoder.Source: F Regime-level NMT results for both… · p. 29

Regime-level results

  • Targeted removal rescues reversed AUROC by 0.139 [0.121, 0.158] for task-only and 0.159 [0.137, 0.183] for task-plus-reconstruction encoders.Under alignment, exposed heads gain 0.056 [0.039, 0.074] and 0.062 [0.043, 0.083] over controls, respectively; under independence, they lose 0.037 [0.026, 0.049] and 0.042 [0.030, 0.055].Source: Table 13 · p. 29, Table 14 · p. 29

Post-removal behavior

  • After targeted removal, exposed-head behavior is numerically indistinguishable across regimes and within 0.003–0.004 AUROC of the control; random removal stays within 0.005 of full-representation values.The control head is invariant to the regime by construction because its fitting weights make reference independent of diagnosis.Source: F Regime-level NMT results for both… · pp. 29–30

G Second observational ECG model: CLEF-Small

The CLEF-Small audit detects behavioral reliance on device-associated provenance, but its rank-one remedy fails the utility check. The localized direction removes task-relevant variance as well as device-associated variance.

G Second observational ECG model: CLEF-Small

  • The audit uses a frozen single-lead encoder on 8,888 ECGs from 7,237 patients, with cohort, folds, heads, and procedures matched to ECGFounder.Sensitivity analyses retain positive G on common calendar support and human-validated ECGs, but use smaller cells than the primary analysis.Source: G Second observational ECG model:… · pp. 30–31
  • G is 0.0599 [0.0549, 0.0650], about 2.6 times ECGFounder’s value on the same cohort, confirming excess vulnerability under association reversal.The exposed head loses 0.113 [0.102, 0.125] AUROC, while the control loses 0.053 [0.043, 0.063].Source: G Second observational ECG model:… · p. 30
  • Held-out conditional device AUROC, averaged over target strata and folds, was 0.897 for CLEF-Small versus 0.959 for ECGFounder; under association stress, excess vulnerability was G = 0.0599 [0.0549, 0.0650].Source: G Second observational ECG model:… · p. 30
  • Rank-one projection reduces G by 0.0375 [0.0327, 0.0426] (63%) and rescues reversed AUROC by 0.057 [0.046, 0.068], beating all 100 matched random directions.A refitted conditional device probe remains at 0.830.Source: G Second observational ECG model:… · p. 30
  • The rank-one remedy is rejected: the control head’s independent-regime AUROC falls by −0.0386 [−0.0443, −0.0334], beyond the −0.01 utility margin.The task-residualized direction removes task-relevant variance alongside device-associated variance.Source: G Second observational ECG model:… · pp. 30–31

H.1 Held-out HMC controlled reference-shift evaluation

The HMC reference-shift screen finds that accessibility and transfer do not have a consistent ordering across frozen models. The accompanying ECG, architecture, uncertainty, and artifact analyses define the scope and reproducibility of these evaluations.

H.1 Held-out HMC controlled reference-shift evaluation

  • BIOT has the strongest shifted and worst-reference utility, while CBraMod has the greatest reference accessibility and weakest transfer; LaBraM combines the lowest probe with intermediate transfer.The HMC audit holds recording, stage label, and active C4 electrode fixed while constructing four reference views.Source: §H.1 · p. 31, Table 17 · p. 32

H.2 ECGFounder: association specificity and confounding sensitivities

  • ECGFounder’s targeted projection reduces G from 0.0233 to 0.0021 [0.0007, 0.0034] and improves reversed AUROC by 0.0342 [0.0287, 0.0397].The control-head utility change, −0.0036 [−0.0059, −0.0012], remains within the −0.01 margin; the audited provenance includes device, era, and workflow, not isolated hardware causality.Source: §H.2 · p. 33

H.3 ERP architecture robustness

  • Across EEGNetSmall and ShallowConvNet, masked reconstruction has the weakest shifted utility and highest domain-probe accuracy among the compared ERP objectives.This is an architecture check, not a backbone-specific search or causal identification of the reconstruction term.Source: §H.3 · p. 34, Table 19 · p. 34

I Evidence units, seeds, and uncertainty

  • Uncertainty follows each experiment’s independent unit: encoder-seed and record resampling for NMT, patient clusters for ECG, and leave-one-subject-out sensitivity for ERP.Intervals condition on the split and selected probes rather than a newly sampled training corpus.Source: I Evidence units, seeds, and uncertainty · p. 35, Table 20 · p. 35

J Additional tables, model coverage, compute, and assets

  • The supplementary tables specify evaluation settings, factor roles, provenance-shift families, and a minimal provenance schema; they distinguish observed groups from fields that may be unavailable.The broad screens mix acquisition with population, interface, era, and workflow, so they are deployment-shift screens rather than causal estimates of a single hardware factor.Source: Table 21 · p. 36, Table 22 · p. 37, Table 23 · p. 37

J.2 Frozen-encoder preprocessing details

  • Frozen-checkpoint preprocessing and probe fitting differ from calibrated LPI: broad-screen probes use train plus validation, while LPI uses train-only fitting and validation selection and calibration.LaBraM EEG views are assigned together by record to prevent cross-view leakage.Source: §J.2 · p. 38

J.3 Artifact contents

  • The artifact separates reusable evaluation utilities from benchmark-local reporting and saved outputs; raw datasets and external checkpoints are not redistributed.It includes runners, locked manifests, saved predictions or probe states, bootstrap draws, and checked uncertainty estimates.Source: §J.3 · p. 38, Table 24 · p. 39

J.4 Physiological foundation-model pretraining landscape

  • The pretraining landscape varies in corpus coverage, channel handling, preprocessing, and objective, with reconstruction or masked prediction common among the listed models.The inventory is background rather than a ranking of the evaluated models.Source: §J.4 · p. 39, Table 25 · p. 40, Table 26 · p. 41

J.5 Evaluation settings

These appendices document the broad-screen settings, paired contrasts, transfer visualizations and tables, compute, and asset terms. They distinguish descriptive comparisons and realistic deployment-shift screens from causal estimates or interchangeable training replicates.

J.5 Evaluation settings

  • Broad-screen settings identify the provenance groups, utility metrics, and seeds, while dataset access covers NMT, ERP/P300, and PTB-XL.The listed settings distinguish independent encoder trainings from repeated evaluation fits for frozen checkpoints.Source: §J.5 · p. 41, Table 27 · p. 42

J.6 Paired seed-wise contrasts

  • Paired seed-wise deltas are descriptive checks, not causal isolation of an objective component or evidence of universal ordering.Probe deltas compare classification-probe accuracies, not calibrated LPI or downstream reliance.Source: §J.6 · p. 42, Table 28 · p. 42

J.7 Source-to-target transfer heatmaps

  • Figures 6–8 encode source acquisition groups as rows and target groups as columns, with diagonal in-domain and off-diagonal shifted evaluations.They visualize transfer matrices for NMT, ERP/P300, and PTB-XL.Source: §J.7 · pp. 42–43, Figure 6 · p. 43, Figure 7 · p. 43

J.8 Detailed target and transfer tables

  • Tables 29–32 separate NMT direct and pooled-target scores from genuine source-specific transfer matrices, whose off-diagonal entries evaluate held-out source groups.The mixed-reference NMT row is an exposure baseline trained on all three references.Source: §J.8 · p. 44, Table 29 · p. 44, Table 30 · p. 44

J.9 Compute resources

  • The experiments use local GPU and Apple-silicon workstations; evaluation, bootstrap intervals, and figure regeneration run on CPU from saved embeddings or numerical artifacts.The paired NMT encoder training takes under two minutes per d = 128 encoder.Source: §J.9 · p. 45

J.10 Existing assets, licenses, and terms of use

  • The artifact does not redistribute raw physiological datasets or external pretrained checkpoints, which users must obtain through official access paths and applicable terms.Code licenses are not assumed to license pretrained weights.Source: §J.10 · p. 45, Table 33 · p. 47

Related papers

Loading 2609.34466v1…