A New Build Enters the Lifecycle
Every candidate version starts here, regardless of size.
- 3: Change types seen (bug fix, minor update, retrain)
- 1: Entry point (single intake gate)
- 0: Unvalidated deploys (none allowed to skip review)
- All: Versions tracked (full build history logged)
Why every change enters the same gate
No version, however small, bypasses intake review.
Does This Touch Clinical Decision Logic?
Classification splits changes into non-clinical vs clinical-impact.
- 2: Classification axes (significance, decision-logic impact)
- UI, perf, logging: Non-clinical examples (no decision-logic change)
- Thresholds, model weights: Clinical examples (alters output behavior)
- High cost: Misclassification risk (under-validation risk)
Sorting signal from noise
A UI tweak and a retrain look similar in a diff, not in risk.
Scope Scales to Classification
Minor changes get streamlined checks; major ones get full validation.
- 3: Streamlined checkpoints (regression-only track)
- 6: Moderate checkpoints (expanded regression + subset)
- 10: Full checkpoints (full clinical validation track)
- Review board: Scope decision owner (sets required track)
Proportional, not maximal, validation
Over-testing trivial fixes wastes cycles; under-testing major ones risks harm.
Checkpoints Run at the Determined Scope
Each track runs its checkpoints until all gates turn green.
- 2: Checkpoint states (pending, passed)
- 1 active: Track parallelism (only chosen track runs)
- Full track only: New evidence needed (clinical evidence sometimes required)
- Sequential: Gate order (checkpoints clear in order)
Turning checkpoints green
Every checkpoint on the active track must pass before advancing.
Deploy Only After Clearing the Bar
Skipping proportional validation risks shipping unvalidated clinical change.
- 2: Deploy states (pending, approved)
- All checkpoints green: Approval condition (on the active track)
- Available: Rollback path (if post-deploy signal regresses)
- Retained: Audit trail (classification + test evidence kept)
The proportional bar is the safeguard
Deployment readiness reflects evidence, not just build completion.
Scaling validation to change impact is what keeps deployment fast for trivial fixes and rigorous for clinically-meaningful ones.