Membership in the correct evaluation project and an assigned role
For evaluation programme teams
Run governed MedBoard evaluations
Build evidence in an isolated workspace: synthetic first, retrospective second, with blinded review and hard exit gates.
Who this is forProtocol authors, curators, reviewers, adjudicators, analysts and safety reviewers
- Study design
- Datasets
- Human reference
- Batches
- Exit
Before you begin
Three things to have ready
A predeclared protocol and separation of author/approver duties
No clinical workflow expectation—the workspace is evidence-only
Know the workspace
The screen at a glance
The navigation and information shown here mirror the role you selected. Your permissions determine which actions are available.
Evaluation workspace
KAStudy overview · exit readiness
ProtocolLocked
DatasetLocked
Exit gates2
Synthetic evidence workflowProtocol → reference → candidate → analysisReady
Retrospective evidence workflowAdmission → replay → blinded delta → exitGated
Open safety incidentsAll serious incidents must be resolved0
- 01
Select the evaluation project
Choose the intended project and confirm its scope, permitted data class and role assignments. Project membership never grants access to clinical hospital cases.
- 02
Lock the study design
Author the protocol, intended use, endpoints, thresholds, exclusions, subgroups, missing-data rules and stopping rules. A separate approver reviews and locks it.
CheckpointThe protocol cannot change after lock; create a new version instead. - 03
Build and lock the synthetic dataset
Draft cases, validate identity leakage, consistency, units, sources and eligibility, then complete independent curation. Include clinician-authored and template holdouts.
- 04
Create the human reference
Assign cases to two blinded clinical reviewers. Material disagreements go to an independent adjudicator before the reference version is locked.
- 05
Register the candidate and run the batch
Bind the immutable release, protocol, dataset and reference. The restricted worker executes without clinical side effects and records reproducible evidence.
- 06
Analyse, resolve incidents and approve synthetic exit
Run the predeclared analysis, investigate disagreement, resolve safety incidents and obtain an independent exit approval before admitting retrospective hospital data.
CheckpointRetrospective admission remains blocked until synthetic exit exists. - 07
Run the governed retrospective stage
Approve the hospital admission, freeze time-correct de-identified records, verify historical decisions, run blind MDT replay, collect two blinded delta reviews and adjudicate.
- 08
Approve retrospective exit and export
Every case must be adjudicated, the analysis independently approved and all incidents resolved. Only then can a controlled evidence export—or future shadow planning—proceed.
CheckpointA difference is classified clinically; it is not automatically an AI error.
Good practice
Do this
- Keep protocol, dataset, reference and candidate versions independent.
- Preserve blinding and separation of duties at every approval.
- Create a new version instead of editing locked evidence.
Avoid
Pause before this
- Do not import real records before synthetic exit approval.
- Do not expose reference answers to the candidate worker or reviewers.
- Do not describe an engineering pass as a clinical study result.