Same Model, Same Image, Different Answer: A Run-Time Firmness Controller for Medical AI Servers
A medical AI model is usually tested on one computer and then run on a hospital server with several GPUs. A recent paper (Wang et al., arXiv:2609.36525) shows that the answers can change even though the model, the input and the decision rule staythe same. In some of their tests about one answer in four differed, and some difference remained even after the authors raised the precision of the sensitive layers. This note describes a controller that checks the answer while the server is running. The controller receives the settings under which the model is run (the setting vector) and makes small nudged copies of them. It gives the copies to the same fixed model and builds a Jacobian measure from the scores. Using that measure, the controller decides the route of each case without a human operator: the usual route if the score is firm, another setting if it is not, and a person if no setting is firm. A firm result is not a confidence score. It says only that the same score was obtained at nearby settings, with the spread within a limit fixed in advance. The note explains the paper in plain words. It also explains how the idea differs from the paper, from Bayesian methods, Monte Carlo dropout and ensembles, from calibration and the reject option, from work on identical numerics, and from formal proof checking. Patent pending. Suggested keywords: medical AI, inference, GPU, reliability, Jacobian, controller, run-time monitoring, firmness, deployment
Authors
- PRAKASH VAITHYANATHAN
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24422512
- Primary Topic
- Explainable Artificial Intelligence (XAI)
- Type
- article
- Field-Weighted Citation Impact
- 0.00