Story 02 · ML Engineer
Reproduce and investigate a failed AI execution
Reconstruct a bad production outcome from retained source evidence, then compare the new result without overwriting history.
The situation
A customer-support agent gave an incorrect refund recommendation in production. The ML engineer needs to reproduce the execution, test it against the current quality criteria, and retain an auditable investigation record.
The question
What happened, can we reproduce it, and what evidence explains the difference?
The governance workflow
From question to governed outcome
- 01
Find the source execution
Search tenant-scoped historical executions and select a completed execution with the evidence required for replay.
- 02
Validate replayability
BHANUJ - AI Governance Platform resolves the source again, checks its retained inputs and state references, then freezes an immutable replay configuration.
- 03
Create and submit the replay
A durable worker job reconstructs a new execution while preserving source-to-produced-execution lineage.
- 04
Run and compare evaluations
Evaluate the replay, compare it with the compatible source baseline, and inspect any drift.
- 05
Review evidence and audit
Follow the replay lineage, evaluation results, and bounded audit timeline for the investigation.
In Studio
The screens behind the story
The replay workflow begins with a governed historical source, not a blank rerun.
The frozen configuration and job lifecycle make the reproduction inspectable.
Source execution, replay, produced execution, and evaluation evidence remain connected.
The outcome
The investigation produces a reproducible, historical record—not a retry that silently changes the original evidence.
OSS capabilities used