Set the last four months side by side and a shape emerges. Deployment counts rose steadily across every category we track. Published randomized evidence rose barely at all. Governance infrastructure — registries, safety officers, monitoring — grew faster than either, which is arguably the healthiest signal in the set.

The categories where deployment and evidence are most mismatched are the ones worth watching: behavioral health conversational tools, staffing and scheduling models, and patient-facing triage. Each is being adopted on the strength of a capacity argument rather than an outcomes argument, and each has a plausible mechanism for harm that current monitoring would not detect.

The categories where the two are converging are less discussed and more encouraging. Documentation, pharmacy, and imaging assistance all now have real deployment experience and a growing body of evaluation, including null results that were published rather than buried.

A field that publishes its null results is a field that is maturing. That has been the quiet development of this spring, and it is more consequential than any individual model release.