FAQ

Does MedAgentRx load my model? No. It reads predictions only, as Parquet or CSV. Model files such as pickles can run arbitrary code when loaded, so they are never opened.

Which models are supported? Version 0.1 covers binary classification risk models, the most common kind of clinical prediction model. Multiclass, survival, regression and segmentation are on the roadmap.

Why does a failed evaluation count against the budget? The locked test set loses its value if it is used to pick among many candidates. Counting every look, including failures, stops a person or an agent from trying until one passes by chance.

What is bio-creep? A series of updates, each only slightly worse than the one before, can drift a long way from the authorised model. LIFE-001 compares every update with the anchor as well as with its predecessor.

Can an agent use MedAgentRx directly? Yes. The guard lets an agent try only the modifications the plan marks automatic and spends a look per try. In thresholdout mode it answers with one noisy bit for the whole plan, and with the agent's own validation data it runs Thresholdout.

Is the output a regulatory submission? No. MedAgentRx supports a quality management system. Its reports are evidence for that system, not a submission.

Why is the dashboard read-only and local? It has no authentication in version 0.1. It binds to 127.0.0.1 by default and warns loudly if bound elsewhere.