How the science will be tested

Verify authentication, tenant boundaries, real interfaces, resilience and operational rollback.

Artifacts: reproducible CI, runtime E2E, negative security cases and traceable logs.

Begin with synthetic scenarios. For real data, establish consent, minimization, access boundaries and appropriate research approval.

Evaluate reidentification risk, bias and harmful overconfidence; aggregation alone is not anonymity.

Compare simple or static rule-based approaches with hybrid policies and explicit uncertainty.

Predefine calibration, Brier score, missing-data stability, selective risk, alerts, detection delay and decision utility where applicable.

Study completion of the daily journey, comprehension of explanations, user autonomy and acceptability.

No improvement in mental-health outcomes, retention or ROI is asserted before observing and validating real data.

Maintain protocols, reproducibility, model configurations, negative results and independent review.

Use established guidance such as TRIPOD+AI where prediction reporting applies; reporting guidelines do not substitute for clinical/ethics approval.

Progress when engineering safety, data quality and pre-agreed decision value criteria are met. Stop or revise when access control fails, uncertainty is unsafe or no useful evidence emerges.

An inconclusive result should be reported, not rebranded as success.