The stack decides: local LLM gripes, a rival agent claim, and safety containment plans still not public

A quiet day for new research, but two threads stand out. A viral Level1Techs thread on local LLM underperformance and Inherent's claim that its Faraday agent beat Anthropic and OpenAI at replicating published research both point to the same idea recent papers have made: the stack around a model increasingly decides outcomes more than the model itself. Separately, a new study finds frontier labs still have little public documentation on how they would contain a rogue model, a gap that stands out given OpenAI is simultaneously pushing California to strengthen SB 53. For SMB builders the takeaway carries over from prior digests: inference configuration and evaluation setup often matter more than headline model choice. Safety governance at the policy level remains mostly aspirational, worth monitoring but not yet actionable.