Your best customer can look like fraud
The error count doesn't tell you what the mistake costs. And the customer you turn away doesn't tell you what they would have been worth.
More issues
AI credit fraud looks suspiciously familiar
The payment can look good long enough for the compute to be gone. I think we have seen this problem before.
Compared to What?
Every impressive number is quietly missing its bottom half. A three-question test for “97% token reduction” claims.
Teaching an algorithm Sanzo Wada's eye
The colours were easy. Teaching the metric how Wada combines them took five tries
Answers Aren't Work
Models produce answers. Work is deciding which answers are allowed to touch production.
Prompting Is Not a Safety Boundary
A prompt is not a control. If the only thing between the agent and the blast radius is language, you don’t have a boundary.
"Green" Isn't Done
A passing suite can still be evidence of nothing. Green means the tests ran, not that the risk is covered.