Bilevel Autoresearch wraps an inner research loop inside an outer one that generates new search strategies as Python code at runtime — both loops powered by the same model, with no stronger model required. The recursion no longer needs a smarter supervisor.
The line of the issue is an aside: the economy is discovering that most of its institutions were workarounds for insufficient intelligence.