Governance
·
3 min read
Model cards your risk committee will actually read
Most model cards are written for other engineers. The people who approve the release need four answers on one page.
By
Jordan Achebe
,
Head of Responsible AI & Governance
3 min read
·
A model card that runs to twelve pages of metrics is a document nobody on the risk committee will finish. We rewrote ours around the four questions committees actually ask.
What does it decide, and what doesn’t it?
One paragraph in plain language, with the decisions that stay with a person named explicitly.
How do we know it works?
The evaluation set, who wrote it, the headline result and the three failure modes we found, each with what happens when it occurs.
Who is watching it?
The named owner, the monitoring in place, the thresholds that trigger a review and how fast a rollback takes.
The technical appendix still exists for the people who want it.
Jordan Achebe · Governance
What changed since last time?
A short diff against the previous card, so a quarterly review is a ten-minute conversation instead of a re-read.
The technical appendix still exists for the people who want it. It just isn’t the first thing anyone has to read.
Written by
Jordan Achebe
Head of Responsible AI & Governance
View profile →
02 · Keep reading
More from the team.
03 · Start
Have a use case
like this one?
Most of what’s in this article came out of a real engagement. Tell us about yours.