A coronary agent now shows its evidence before its answer: the gate passes 76.7%, and rejects half of reference lesions

An agent for panvascular intervention is required to carry a checkable evidence state instead of only a plausible answer. The prior condition is a language model that states conclusions directly; retrieval-augmented generation guarantees neither consistent sources, endpoints, denominators and time windows nor abstention when evidence is insufficient. On the public ARCADE coronary angiography benchmark, the evidence gate accepts 76.7% of stenosis-frame candidates while a score-only comparator accepts 100%, compared in the same locked test environment.

The gate separates support from the predictive score and lets the system abstain explicitly. The gate does not retrain a model; it applies geometric checks to the candidate lesion mask and predicted vessel-support mask produced by a fixed model. A candidate is retained only when its area is at least 20 pixels, no more than 25% of the image, and meets an overlap condition; retain acceptance is 76.7%, while horizontal reflection, vertical shift and dilation give 0%, 0% and 0.3%.

Under controlled perturbations segmentation degrades while gate coverage stays non-zero, so coverage cannot serve as a performance guarantee. Previously only pooled metrics were available, leaving the effect of spatial or intensity change on lesion segmentation invisible. Under horizontal reflection, vertical shift and intensity attenuation, stenosis Dice is 0.195, 0.228 and 0.170, AUROC is 0.940, 0.952 and 0.944, and gate coverage is 0.633, 0.790 and 0.670.

A constructed fault bank makes the gate's response to mask transformations inspectable and exposes its rejection cost. Previously reference candidate masks and deliberately constructed wrong candidates were not compared inside one audit procedure. Reference candidate masks are accepted 44.3% and rejected 55.7% of the time; reflection, shift, random-frame candidate-mask substitution and dilation are accepted at 2.0%, 6.0%, 3.0% and 11.7%.

The benchmark addresses method-oriented readers and engineers in panvascular intervention research, and its evaluation scope is image-level tasks on the public ARCADE coronary angiography release, covering syntax and stenosis environments with 1,000 training, 200 validation and 300 test images each. The gate and action audit run on a fixed model, and the stenosis score threshold was selected on validation data from 0.10 to 0.90 in increments of 0.05 by mean image-level Dice, then locked. The authors propose a next step of a narrow, bed-specific study of device-vessel concordance, using coronary data for development and femoropopliteal data as a deliberately difficult external transfer, with geometry, mechanics, haemodynamics and biology measured separately before any composite is considered.

The authors state that the experiment does not establish recursive improvement, clinical correctness or a reduction in adjudicated scientific errors. The gate rejects more than half of reference candidate masks, which may reflect imperfect predicted vessel support, the area limits or the overlap heuristic, and the present audit does not disentangle those causes. Each representation setting was trained for three epochs with one random seed, so training variability and convergence are not characterized, and the projection penalty did not establish a consistent predictive advantage. Constructed fault labels do not establish true error status, and the controlled transformations are not external-cohort validation or a preregistered clinical trial. ARCADE contains no patient identifiers, procedural actions, device specifications, blood flow, longitudinal outcomes or clinical adjudication, so its masks cannot establish Suitcordance, a digital twin or treatment benefit.