My biggest concern now is — if the issue they have is as vague as “reports of degraded quality”, how do they even approach fixing it? And what measurable criter

My biggest concern now is — if the issue they have is as vague as “reports of degraded quality”, how do they even approach fixing it? And what measurable criteria will they use, to declare that it is fixed? Would they take a vibes-check opinion poll? Curious why they can’t run some benchmarks with the model (if they suspect the issue is with the model itself) or some agentic coding benchmarks on Claude-code (if the issue might be with the scaffolding, prompts etc).