txmedai/gaugeflow-coin-evidence

GaugeFlow — Gauge Frontier Evidence

Per-seed evaluation evidence behind the preprint "GaugeFlow: Equivariance, Not Invariance, Is the Right Prior for Continuous Physical Gauges in Self-Supervised Medical Imaging" (Colin Son, MD — Seldinger, Inc.).

This is the evidence bundle: every number in the paper is one analyzer verdict over five seeds, and the raw per-seed, per-study metrics plus the aggregated verdicts are here so the results are independently checkable without rerunning. Code to reproduce is on GitHub.

What's the finding

Gauge-equivariant self-supervision helps iff the nuisance is a continuous physical gauge, the downstream utility is covariant under it, and the gauge acts as a recoverable transformation of the representation. On angiographic projection-angle retrieval — a continuous physical gauge whose utility is invariance-shaped — adversarial invariance is a no-free-lunch wall, equivariance dominates it as an auxiliary, and equivariance fails as a primary objective. The boundary is the contribution.

Projection-angle gauge (CardioSYNTAX), same-study/same-artery retrieval@1, 5 seeds:

Objective	Retrieval Δ vs. baseline	Gauge used? (true vs. shuffled)	Leakage R²
Adversarial invariance (GRL, single)	−0.044, p≈0	—	0.31
Adversarial invariance (GRL, dual-path)	−0.014	—	0.09
Equivariance, auxiliary	−0.018, p=0.11 (n.s.)	0.079 vs. 0.028 (used)	0.30
Equivariance, primary (canonical)	−0.080, p≈0	0.0089 vs. 0.0105 (none)	0.70
Equivariance, primary (pairwise-transport)	−0.050, p≈0	0.0093 vs. 0.0089 (none)	0.73

Contents

File	Contents
verdict_*.json	Aggregated analyzer verdicts: per-arm means, cluster-bootstrap CIs, paired delta, permutation p, shuffled-gauge control, gate checks. equiv = auxiliary, equiv_ctr = primary/canonical, equiv_ptr = primary/pairwise-transport, dp = dual-path adversary.
results_*.jsonl	Long-format per-seed, per-study-cluster metric + leakage rows feeding each verdict.
per_study_metrics/	Raw per-seed retrieval@1 and angle-R² per study, per arm.

Each row: {"arm", "seed", "cluster", "metric", "leakage"}. Treatment arm is named gaugeflow; negctrl is the shuffled-gauge control.

Methodology

• Analyzer: cluster-bootstrap 95% CI, permutation null on the paired per-cluster delta, shuffled-gauge negative control. A method passes only if the task is non-inferior, leakage is under cap (0.035), and the control reads null.
• Gauge continuity (verified pre-training): the CardioSYNTAX primary positioner angle takes 629 distinct values over [−46.4°, 47.5°]; within-study Δangle median 21.7°, only 5.4% exactly zero. The 11 coarse buckets used as a categorical label conceal a genuine continuous signal.
• Caveat: per-run baselines drift ~0.014 from MPS nondeterminism; the within-run paired delta is the valid statistic.

Not included

The CardioSYNTAX / DIAS / MSD Task01 imaging data are licensed by their providers and are not redistributed. This repo holds only derived evaluation metrics.

Citation

@misc{son2026gaugeflow,
  title  = {GaugeFlow: Equivariance, Not Invariance, Is the Right Prior for
            Continuous Physical Gauges in Self-Supervised Medical Imaging},
  author = {Son, Colin},
  year   = {2026},
  note   = {Seldinger, Inc.}
}

Code: https://github.com/SeldingerMed/gaugeflow