PaperlayerView exact revision on arXiv

Source-linked AI summary

Intervention anchors and scientific verification in synthetic vascular predictive representations

Lingsen You, Yujun Guo, Xinyu Zhong, Zisu Peng, Wentong Wang, Li Shen, Junbo Ge

arXiv:2610.11704v1cs.AI

TL;DR

Vascular digital twins motivate interpretable responses, but complete predictive coordinates do not alone establish intervention meanings. This mathematical and synthetic audit tests predictions and anchor-supported interpretations separately. Prediction agreement can coexist with substantial uncalibrated edit error, while coordinate transport restores the requested edit.

  • Problem

    Vascular digital twins motivate connecting observed vessel states to interpretable responses, while suitcordance labels do not validate learned predictive coordinates.

    Source: §1 · pp. 1–2

  • Method

    The paper audits a controlled linear setting by comparing complete orthogonal prediction transforms with anchor-based interpretation and coordinate transport.

    Source: §2.1 · p. 2, §2.5 · p. 4, §4 · p. 9

  • Results

    6.66 × 10−15 maximum paired output discrepancy coexisted with median 1.513 RMS error across three uncalibrated edits over 64 random orthogonal bases.

    Source: §3.1 · p. 7

  • Takeaways & Limitations

    Vascular anchor design should estimate observable perturbation signatures and their rank rather than assign ecology labels to factor indices.

    Source: §4 · p. 9

    The study is a synthetic numerical audit using conceptual coordinates, not measured vascular mechanisms, clinical validation, or physiological domain-shift testing.

    Source: §2.1 · p. 2, §2.5 · p. 4

Abstract

from arXiv · show

Complete orthogonal predictive coordinates do not by themselves bind a latent direction to a named intervention. We present a mathematical and synthetic audit motivated by vascular device-vessel suitcordance. Capacity-matched least-squares predictors were exactly equivalent under complete fixed output transforms, whereas an anchor-only observer recovered interpretations only within the span of known perturbation signatures. Six three-dimensional configurations across 64 seeds gave a maximum paired prediction discrepancy of 6.7e-15 but a median untransported edit error of 1.513. Coordinate transport removed that error. Noisy and weak anchors constrained calibration stability, and changing the representation basis required recalibration or verified transport. Across 256 additional fits in dimensions 3-24, prediction equivalence persisted within 4.0e-15. We then evaluated nine deliberate runnable fault classes across 64 seeds. All 576 faulty executions completed, but each violated at least one reconstruction, prediction, delivered-edit or scope contract; all 320 valid control records passed. Repeating a faulty implementation gave exact self-agreement despite error against the separately computed simulator expectation. For one omitted-direction defect, probe coverage followed its analytic law, and rank-aware abstention protected unsupported interpretations. Scalar-noise experiments exposed both missed weak faults and excessive rejection under narrow relative tolerances. These controls provide an executable separation of prediction, semantic support and scientific acceptance. They are synthetic numerical audits, not clinical validation, neural JEPA-Anything replication, agent learning or patient treatment-effect estimation.

1 Introduction

The paper uses vascular digital twins and device–vessel suitcordance to motivate interpretable predictions, while distinguishing predictive coordinates from validated intervention meanings. It audits what known perturbation signatures can identify in a controlled linear setting.

1 Introduction

  • Vascular digital twins motivate linking observed vessel states to interpretable responses for device selection or follow-up, and suitcordance names complementary device–vessel interaction aspects.The frameworks motivate scientific questions but do not validate a predictive model’s learned coordinates.Source: §1 · pp. 1–2
  • JEPA-Anything motivates examining complete predictive coordinates, but its factor geometry is distinct from causal meaning; this study is an independent linear control audit, not a neural-results reproduction.The cited work describes learned projectors, factor-specific prediction, and state synthesis across domains [3].Source: §1 · p. 2
  • The IABP cohort associated use with higher in-hospital mortality, but investigators emphasized residual confounding and an observational interpretation.The cohort is a motivation for evidential boundaries; no patient data were accessed and no simulator was calibrated to its estimates.Source: §1 · p. 2
  • The audit asks what paired perturbation signatures identify and how noise, weak signatures, and executable checks affect interpretation of named edits.It combines algebraic controls, ambiguity witnesses, dimensional stress tests, and a fault bank.Source: §1 · p. 2

2 Methods

The methods use a synthetic, dimensionless three-coordinate setting and restrict the observer to specified information. They compare equal-capacity least-squares predictors under complete orthogonal output transforms.

2.1 Study scope and information boundary

  • The three state coordinates are dimensionless conceptual analogues of mechanical, cellular, and physicochemical–immune ecology, not measured vascular mechanisms.The configurations are not validated vascular beds, and no patients, images, tissue assays, clinical outcomes, or device measurements were analyzed.Source: §2.1 · p. 2
  • The anchor-only observer receives A and B but not the synthesis matrix Q, additional named-state–latent pairs, transition eigenstructure, or biological supervision.If Q or sufficiently informative paired states are available, coordinate transport is already identified; the study does not claim universal non-identifiability from longitudinal data.Source: §2.1 · p. 2

2.2 Complete orthogonal coordinates and matched prediction

  • For complete orthogonal coordinates, u = Q^T x and x = Qu, with Q orthogonal and allowed either determinant sign.The analysis and synthesis maps are inverse transforms.Source: §2.2 · p. 3
  • Monolithic least squares fits W = D†Y, while three output heads with the same seven predictors fit V = D†(YQ); each predictor has 21 parameters.The equality checks a fixed complete output transformation, not learned nonlinear projectors, optimization, activity regularization, or independent nonlinear branches.Source: §2.2 · p. 3
  • A latent-coordinate edit and a named-state edit are distinct, and coordinate labels and state edits are not treatment effects.A latent unit edit maps through Q, whereas a specified named-state edit requires Q^T for its latent representation.Source: §2.2 · p. 3

2.3 Synthetic prediction configurations

Six synthetic configurations and a rotation analysis test prediction and edit behavior; a separate anchor analysis characterizes supported interpretations and ambiguity. The experiments use specified linear controls rather than physiological calibration.

2.3 Synthetic prediction configurations

  • The six synthetic configurations vary diagonal transition coefficients, with small fixed off-diagonal terms and arbitrary time units rather than physiological calibration.Each of 64 seeds uses 1,024 training observations and 512 independent test observations, with the two models sharing data and least-squares convention.Source: §2.3 · p. 3
  • The rotation controls distinguish invariant prediction residuals from coordinate activity: isotropic activity is unchanged by orthogonal transforms, while anisotropic activity can change.The 0.8 activity threshold is illustrative; the study does not treat coordinate activity as generally rotation-invariant.Source: §2.3 · p. 3

2.4 Anchors and supported interpretation

  • The anchor-only map recovers supported named-state changes within span(A), while the residual component outside that span is unsupported and calls for abstention.A zero residual permits interpretation only within the specified linear model and does not establish a biological mechanism.Source: §2.4 · pp. 3–4
  • For anchor rank below three, transformations can preserve all observed anchors while changing unanchored directions; a known SO(3) orientation constraint can remove the rank-two sign ambiguity.The construction gives explicit ambiguity witnesses in O(3).Source: §2.4 · p. 4

2.5 Noisy anchors and coordinate transfer

The methods compare unrestricted and orthogonal calibration from noisy anchors, test sensitivity to anchor conditioning, and examine representation-coordinate mismatch. These controls do not model physiological domain shift.

2.5 Noisy anchors and coordinate transfer

  • With full-rank identity anchors, noisy signatures are calibrated by either an unrestricted map or orthogonal Procrustes using a known orthogonal-map prior.Noise standard deviations are 0.05, 0.2, or 0.5, averaged over 1, 4, 16, or 64 measurements; reflection is allowed.Source: §2.5 · p. 4
  • Identity-anchor calibration targets overlap the calibration directions, making the noise-free result a reconstruction check rather than independent biological validation.Conditioning is tested by reducing singular-value scale while holding measurement-noise standard deviation at 0.02 and using 16 repeats.Source: §2.5 · p. 4
  • A map from a source representation basis cannot be applied directly to a different target basis; recalibration or known cross-basis transport restores consistency.This tests coordinate mismatch, not physiological domain shift or cross-hospital validation.Source: §2.5 · p. 4

2.6 Dimensional extension of prediction and anchor controls

The dimensional extension tested complete-coordinate prediction equivalence and anchor-span coverage, then specified executable contracts for distinguishing valid controls from deliberate faults.

2.6 Dimensional extension of prediction and anchor controls

  • For d = 3, 6, 12, and 24, the coverage control used ⌊d/2⌋ distinct unit-coordinate anchors, with the identity AB†QT = P checked in each dimension.Source: §2.6 · p. 4

2.7 Executable scientific contracts and controlled faults

  • The fault bank specified nine defective pipeline variants and five valid controls across 64 seeds, alongside a separately recorded execution-error variant.The design produced 896 runnable records, including 576 defective records and 320 valid controls.Source: §2.7 · p. 5
  • The fault-bank reference used 128 noiseless state-action samples per seed and checked reconstruction, prediction, and edit errors against a 10−10 tolerance.Each faulty case was paired with a passing full-span reference replacement as a validity witness, not a measured learner.Source: §2.7 · p. 5
  • Five contracts checked execution, coordinate reconstruction, state prediction, delivered supported edits, and truthful reporting of supported scope.The edit check measured named physical coordinates after synthesis, while the scope check compared declared dimension with numerical anchor rank.Source: §2.7 · p. 5
  • A consistent latent-unit change passed reconstruction because its analysis and synthesis maps satisfied DC = I, without requiring an orthogonal analysis map.The unit-conversion control used C = 2QT and D = Q/2.Source: §2.7 · p. 5
  • A fixed correction trace applied known coordinate transport, but did not test an agent’s ability to diagnose or learn from feedback.The trace recorded the candidate operation, contract residual, prescribed correction, and repeated checks.Source: §2.7 · pp. 5–6

2.8 Probe coverage for an unanchored defect

The probe experiment tested whether random unit-axis probes detect a defect that leaves rank-two anchors unchanged. A separate complement probe and scope gate addressed the unsupported direction.

2.8 Probe coverage for an unanchored defect

  • The candidate reversed only the third direction, leaving the two anchor observations unchanged and making anchor residuals unable to distinguish the extensions.The experiment drew unit-axis probes uniformly with replacement from {e1,e2,e3}.Source: §2.8 · p. 6
  • For probe budgets n = 1, 2, 4, 8, 16 and 32, the exact detection probability was 1 − (2/3)^n under the specified probe mechanism.Each probe misses the affected axis with probability 2/3; the probability is not a sensitivity claim for arbitrary faults or clinically important changes.Source: §2.8 · p. 6
  • One targeted probe in the unanchored complement could detect the sign defect, while a rank-two scope gate withheld interpretation of that unsupported direction.The complement probe provides evaluation information unavailable to an observer restricted to the original anchors.Source: §2.8 · p. 6

2.9 Tolerance behavior under controlled measurement noise

The tolerance analysis compared rejection rules under controlled Gaussian measurement noise and a sign-reversal fault. The expansion analyses were documented as post-inspection additions, with reproducibility materials and AI-assistance disclosures also reported.

2.9 Tolerance behavior under controlled measurement noise

  • The tolerance experiment compared absolute, relative, and hybrid rejection rules across simulator responses δ from 10−6 to 1 under Gaussian noise with σ = 10−4.Each of 25 signal settings used 16,384 measurements per condition, and the rules reused the same measurements.Source: §2.9 · p. 6
  • Healthy rejection was treated as a false positive and defective rejection as a true positive for the particular sign-reversal fault.The signal δ is a simulator response scale, not a clinical effect size or threshold of medical importance.Source: §2.9 · p. 6

2.10 Expansion provenance and reporting

  • The additional dimension, fault-bank, probe-budget, and tolerance analyses were specified after inspection of the original study and its first computational results.They were recorded in an expansion addendum before their own execution, rather than as an external preregistration.Source: §2.10 · p. 6

2.11 Reproducibility

  • The addendum introduced controls including signed-permutation bases, coefficient and action-response transport, and ten-step rollout composition, without observed longitudinal trajectories.Descriptive bands across seeds were not confidence intervals for patients or biological effects, and no significance testing was used.Source: §2.11 · p. 7
  • The ancillary archive supplies seeds, specifications, scripts, seed-level results, execution traces, and figures to regenerate the reported simulations.The listed environment used Python 3.12.14, NumPy 2.3.5, and Matplotlib 3.11.2.Source: §2.11 · p. 7

2.12 DECLARATION OF GENERATIVE AI AND AI-ASSISTED TECH-NOLOGIES

  • Codex assisted with manuscript assembly, editing, and consistency checks, and the authors reviewed and revised the assisted text.The authors take responsibility for the submitted manuscript.Source: §2.12 · p. 7

3 Results

The results show that complete fixed-coordinate transformations preserved predictions while named edits depended on transport and anchor support. In the synthetic audits, executable contracts rejected the specified faults, and probe and tolerance outcomes depended on their declared settings.

3.1 Prediction was preserved while uncalibrated edits changed

  • The maximum paired prediction discrepancy was 6.66 × 10−15, while the median RMS error across three uncalibrated edits was 1.513 and transported edit error was at most 6.91 × 10−16.Across six configurations and 64 seeds, the median test MSE of the monolithic predictor was 0.002519.Source: §3.1 · p. 7

3.2 Anchor coverage and noise constrained interpretation

  • Rank-one and rank-two anchors supported one and two of three unit edits, respectively, while unsupported edits were marked for abstention and alternative complement-sign bases preserved anchor signatures.At measurement-noise SD 0.2, median inverse-map error fell from 0.324 to 0.039 and Procrustes error from 0.172 to 0.019 when measurements increased from one to 64; at SD 0.5 with one measurement, inverse-map error’s upper empirical percentile was 9.824.Source: §3.2 · pp. 7–8

3.3 Full rank and source calibration were insufficient controls

  • With true rank three, increasing condition number from one to 100 raised median inverse-map error from 0.0079 to 0.5860 and Procrustes error from 0.0043 to 0.2033.Applying a source calibration to an independently rotated target basis gave median RMS error 1.487; target recalibration reduced error below 9.74 × 10−16, and known cross-basis transport below 1.04 × 10−15.Source: §3.3 · p. 8

3.4 Complete-coordinate equivalence extended across dimensions

  • Across d = 3, 6, 12 and 24, the largest transformed-output prediction discrepancy was 4.00 × 10−15, while coordinate anchors supported ⌊d/2⌋ of d directions.The remaining directions lay outside the anchor span; the result does not establish equivalent behavior for learned nonlinear factorization.Source: §3.4 · p. 8

3.5 Execution and regression agreement did not certify the candidate pipelines

  • All 576 specified defective runnable records were rejected by at least one scientific contract, while all 320 valid records passed all-contract evaluations.Repeating a faulty candidate produced identical predictions despite maximum absolute oracle prediction error 3.519 in the seed-zero duplicate-gain control.Source: §3.5 · p. 8

3.6 An unanchored defect required informative testing or explicit abstention

  • Random-probe detection rates followed 1 − (2/3)^n with maximum absolute deviation 0.0142, and one targeted complement probe detected the constructed sign defect.With only the original anchors, the scope gate withheld the third-direction interpretation rather than assigning it a zero effect.Source: §3.6 · p. 8

3.7 Tolerance choices exposed different failure modes

  • At δ = 10−4, healthy false-positive rates for absolute, relative, and hybrid rules were 0.0000, 0.9590, and 0.0000, while defective detection rates were 0.0015, 0.9941, and 0.0015.These results concern an imposed sign defect under known noise, not a universal tolerance or clinical acceptance threshold.Source: §3.7 · p. 9

4 Discussion

The discussion separates predictive accuracy from justified intervention names and scientifically accepted edits. It recommends explicit anchor and verification checks while keeping conclusions within the synthetic audit’s scope.

4 Discussion

  • In the controlled linear setting, complete-state prediction can remain equivalent across bases while naming coordinates and justifying edits require explicit coordinate information.The audit does not contradict learned neural gains or establish a general impossibility theorem.Source: §4 · pp. 9, 11
  • Vascular factor reports should specify observable perturbation signatures and anchor support, with changes outside the calibrated span left unknown and basis changes transported or recalibrated.The proposed report includes the anchor matrix, estimated span, smallest retained singular value, measurement variability, held-out response checks, and model version.Source: §4 · pp. 9, 11
  • A simulator edit, experimental response, and patient treatment contrast are different estimands; the observational IABP association is not an intervention oracle.Patient treatment planning would require prospective or otherwise justified causal evidence.Source: §4 · pp. 9–10
  • A prediction-preserving representation can still yield an incorrect named edit, so verification must test the end-to-end physical quantity claimed.A verifier can also reject a correct pipeline if it mistakes a coordinate convention for the physical quantity.Source: §4 · p. 10
  • Repeated validation cannot expose a defect confined to directions outside the sampled anchor span; where informative probes are unavailable, abstention preserves the evidence limit.The constructed setting gives random probes a calculable coverage probability, distinct from targeted complement probes.Source: §4 · p. 10
  • The fixed correction trace records one coordinate-error repair but does not test harness adaptation, reinforcement learning, or researcher collaboration.Its successful correction follows from a supplied mathematical rule; learned repair would require separate learners, resource-matched comparisons, and held-out tasks.Source: §4 · p. 10
  • The results support algebraic checks under the declared simulation assumptions, not future discovery capability or transfer to clinical research.The oracle avoids a deliberately shared code defect but retains the same mathematical world; independent implementation, biological measurement, and clinical validation are distinct evidence sources.Source: §4 · p. 10
  • The audit’s added dimensions and fault checks do not expand simulator biological validity; its transparent linear assumptions exclude nonlinear response, heterogeneous patients, and hidden confounding.The contribution is a reproducible translation of coordinate and calibration mathematics, not a new foundation model or validated medical device.Source: §4 · pp. 10–11

5 Mathematical appendix

The appendix shows that anchors identify mappings on their span, while unanchored directions remain ambiguous. It also bounds inverse-calibration stability under a finite-noise condition.

5.1 Exact span recovery

  • Every orthogonal synthesis map agreeing with the anchors agrees on their span, but transformations acting on the complement remain indistinguishable from anchor observations.In three dimensions, the complement has O(3−rank(A)) freedom; at rank two this includes a sign reflection when handedness is unspecified.Source: §5.1 · p. 11

5.2 A finite-noise bound for the inverse map

  • For square invertible A, if anchor noise satisfies ∥E∥2 < α, the perturbed anchor matrix remains invertible and the inverse calibration error is governed by E and eB−1.Here α is A’s smallest singular value; the bound applies to the unrestricted square inverse, not every pseudoinverse or Procrustes estimator.Source: §5.2 · p. 11
  • Calibration stability deteriorates as perturbation noise approaches the weakest signature scale, and a numerical rank declaration alone does not ensure a small denominator.The stated deterministic bound has the square-inverse norm condition and should not be generalized to every estimator.Source: §5.2 · p. 11

6 Declarations

The declarations describe the study’s archive and synthetic-data status, ethics and consent, funding, and competing interests.

6.1 Data and code availability

  • The ancillary archive provides seeds, specifications, executable analyses and audit scripts, seed-level results, and original figures; generation scripts recreate the synthetic observations.Serialized observation arrays are omitted from the compact public archive.Source: §6.1 · p. 11

6.2 Ethics and consent

  • The study used no human participants, animals, or identifiable personal records, and obtained no ethics approval or participant consent..Source: §6.2 · p. 11

6.3 Funding

  • The work reports support from the National Natural Science Foundation of China, Fudan University’s Medical Engineering Joint Fund, and its AI for Science Foundation.The declaration lists grant identifiers T2288101, 82170342, yg2023-01, and FudanX24AI003.Source: §6.3 · p. 12

6.4 Conflicts of Interest

Related papers

Loading 2610.11704v1…