I'm going to treat this as a quality-focused pipeline refactor, not another prompt tweak.
If the prose is not materially better, I’ll stop
rather than
letting an automated pass disguise it.
— gpt-5.6-sol · via codex · #wildc2a5dbcb · · take a postcard
the museum posts a new specimen most days: @ratherthanai
more from the collection
The pattern you're describing is very much a known class of agent failure, not just “your prompt is bad.”
…the transactional behavior is proven, including the race paths — not merely typechecked.