Readiness before authorization
A bounded builder checklist for intended use, release contraindications, evidence traceability, risk controls, monitoring, and qualified human authorization before a health AI prototype advances.
Mandatory hold conditions
Intended users, setting, and excluded uses are not explicit.
The system can influence diagnosis, treatment, dosing, or emergency decisions without qualified review.
Known failure modes lack detection, escalation, containment, or recovery controls.
Claims about performance or safety cannot be traced to current, applicable evidence.
Protected health information boundaries, retention, access, and deletion are unresolved.
No qualified person is authorized to accept residual clinical risk.
Describe who uses the system, for which task, in what setting, on which population, and for what decision. List excluded uses with equal precision. Marketing language and technical capability must not silently expand the intended use.
Identify credible failure modes, affected people, severity, time to harm, reversibility, and existing safeguards. Include omission, delay, hallucination, automation bias, subgroup performance, workflow disruption, and escalation failure.
Connect every material claim to a source, dataset, evaluation, reviewer, date, population, and limitation. Distinguish internal prototype testing from evidence suitable for a clinical or regulatory decision.
Test whether users understand limitations, notice uncertainty, verify consequential outputs, resist automation bias, and can escalate safely. A technically accurate output can still be unsafe in a poorly designed workflow.
Define access, logging, privacy, monitoring, incident response, rollback, model and prompt change control, and evidence retention. Assign owners and response deadlines before release.
Prepare a handoff that states intended use, excluded use, evidence status, unresolved risks, mitigations, monitoring, and the decision required. Authorization must be named, bounded, dated, and revocable.
Reviewer handoff
The handoff should include intended and excluded use, workflow, affected population, performance claims, evidence limitations, known failures, risk controls, privacy and security boundaries, monitoring, incident response, and every unresolved decision.