Governed AI-Assisted Authoring

The Metricis study assistant is feature-gated. It helps transform protocol material into reviewable draft configuration; it does not autonomously approve, release, or operate a study.

Workflow

1. An authorized user supplies protocol or structured source material.
2. The assistant extracts relevant entities and proposes configuration.
3. Metricis represents the proposal as a ChangeSet with source provenance.
4. Deterministic validation checks identifiers, references, rules, and supported contracts.
5. A person reviews the source, diff, warnings, and unresolved decisions.
6. An authorized person accepts, edits, rejects, and eventually releases the configuration through the normal lifecycle.

Appropriate uses

• drafting study metadata, schedules, events, forms, and mappings;
• identifying ambiguity or missing protocol decisions;
• proposing reusable modules and translations;
• comparing a protocol revision with released configuration; and
• preparing an inspectable starting point for human review.

Boundaries

The assistant must not be treated as the source of truth for eligibility, randomization, clinical judgment, consent language approval, safety decisions, or regulatory interpretation. Model output may omit, invent, or misread details. Source citations, deterministic checks, role-based approval, and audit provenance are required controls—not optional polish.

Deployment considerations

Enabling the assistant requires an approved model/provider configuration, data-handling review, access controls, retention policy, prompt/source provenance, evaluation against representative protocols, and a fallback authoring process. No protected or confidential material should be sent to a provider without an approved data-processing basis.

See Forms and governed data capture, Core concepts, and Trust and governance.