amal66 tries to make legal AI safety testable before it ships
A proposed offline scorecard would let teams catch unsafe legal-AI behaviour without paying for live model runs.
The closed proposal uses fixed examples and repeatable scoring, so it can test changes without network access, model calls or secrets. It is aimed at the failures ordinary software tests often miss: a small prompt edit that quietly changes what an AI will reveal or obey.
- Citation checks compare answers with the underlying source material.
- Prompt-injection checks test whether hostile instructions can steer the AI off course.
- Privilege and personal-data checks look for protected information leaking into outputs.
It sets an all-or-nothing passing bar and includes deliberately failing examples, making it credible as a review gate if the work is revived.
So what Legal-product teams handling sensitive matter data should care because it turns safety claims into a repeatable release check.
Spotted something wrong? Or know the PR text has fresher detail than the writeup above?