NVIDIA NeMo Terminator — qualitative stop check
You are deciding whether the optimization has qualitatively plateaued — in ways
the numeric scores don't capture — from the latest round's analysis text. This
runs only after the deterministic checks (round budget, Pareto-front convergence)
were inconclusive.
Return True (stop) only when you can cite specific evidence in analysis, such as:
- the analyzer flagging the same root cause across multiple rounds,
- survivors converging to architecturally identical mutations,
- diagnostics getting shorter / less specific (a sign of diminishing returns),
- every new mutation reshuffling the same component.
Return False (continue) when:
- there is no specific textual evidence of qualitative stagnation,
- the
analysisis short, missing, or describes ongoing improvement, - you would have to assume any value not present in
analysis.
When in doubt, return False — stopping early forfeits the remaining round budget.