01 / CAPTURE
A run or evaluation fails
Record the failure, affected inputs and artifacts. Preserve the original run as evidence.
run_id · failure type · evidence refsFrom failure to the next candidate
A failed run can start
a better-informed next run.
Recover the process, protect the current release and turn reviewed failures into useful training or configuration changes.
02 / QUALIFY
Decide whether to react
A trigger you set up picks which failures count. Duplicate events are ignored and the usual checks apply.
event id · policy · retry limit03 / DIAGNOSE
Choose the right response
Retry a transient failure, correct invalid data or propose a model/configuration candidate.
rules you write · limited investigation04 / BUILD
Create a new candidate
A separate declared pipeline prepares corrected inputs or runs a targeted improvement.
parent run · dataset revision · candidate05 / EVALUATE
Test the proposed improvement
Compare against a baseline and held-out criteria. A regression blocks advancement.
quality · cost · safety · baseline06 / DECIDE
Approve, reject or escalate
Someone with approval rights can advance a passing candidate. Otherwise retain the current version.
approval · promotion · rollback evidenceClose the feedback loop: monitor the approved result. A later eligible failure can start another bounded cycle with a new run identity.
Example / demand forecasting
A forecast misses the quality threshold.
Evidence from the failed evaluation starts a declared review-and-improvement pipeline. The current approved forecast remains in use while a candidate is tested.
Ready for review, not automatically released.
The candidate meets the configured evaluation criteria. A reviewer with approval rights decides whether to publish it; the evidence links back to the failed run.
Deduplicate failure eventsBound attempts and spendKeep held-out evaluationPreserve approval requirementsEscalate repeated failure
Interactive illustration, not a live self-healing system. Trigger wiring and recovery behavior must be verified for the selected pipeline and backend. Retrying infrastructure is not the same as improving a model. The workflow does not rewrite itself or relax its own policy.
Run it on your own infrastructure.
Apache 2.0 licensed, self-hosted. The quickstart gets a control plane with demo data running; a call gets your first pipeline onto it.