Lab Notes / Technical field journal
Governed AI Repair
Designing useful AI-assisted repair without silent changes to operational knowledge.
01
Goal
Explore how AI-assisted repair could help maintain operational knowledge without allowing a model to silently change it.
02
Setup
Gnojo's governed repair pattern begins with an integrity finding and the source's current state. A proposed repair can then be explained, previewed, validated, reviewed, and recorded with its provenance before approval.
03
What I tried
I separated assistance from authority. Curator can inspect a problem, explain it, propose a repair, and preview the effect. It cannot publish independently, approve its own repair, override the reviewer, or hide provenance.
04
The design challenge
Generating a correction was only part of the problem. A trustworthy workflow also needed to show what changed, why it changed, who reviewed it, whether it passed validation, and where recovery boundaries remained.
05
Result
That approach led to governed repair workflows and the Fix Wizard: findings become guided repair paths, proposed changes can be inspected before acceptance, and consequential decisions stay with a human reviewer.
Takeaway
What I carried forward
AI can be useful in repair workflows without becoming the authority that decides whether a consequential change becomes real.