Posts on pov

Review time is the new bottleneck. The fix isn't trusting the model, it's putting an adversarial agent in front of every AI output so trust becomes a property of the process, not a feeling.

We rolled back from Opus 5 to Opus 4.8. Memory loss, path confusion, confidently wrong answers, issues we hadn't seen since Sonnet 4.x. Newer isn't always better. Here's the two-part system we use to grade every model release before it costs our customers time, energy, and money.