ReadonlycompetitorsOptional ReadonlyerroredRuns EXCLUDED because the turn errored / was rate-limited (detected by the
driver's runError), present only when > 0. These are NOT counted in n or
as misses — so rate reflects only valid runs. A large errored relative to
n means the measurement is thin (e.g. a Codex usage limit was hit); re-run.
Optional ReadonlyfalseFraction of irrelevant runs that wrongly fired (lower is better). Present only when TriggerRateSpec.irrelevantPrompts was given.
ReadonlynTotal relevant runs (prompts × trials).
Optional ReadonlyperPer-prompt stats for the irrelevant set. Present with irrelevant prompts.
ReadonlyperOptional ReadonlyprecisionrelevantFired / (relevantFired + irrelevantFired) — of all firings, the
share on the right prompts. Present only when irrelevant prompts were given
AND something fired (undefined when nothing fired at all). Pairs with rate
(recall) to catch a skill that fires on everything or nothing.
ReadonlyrateOverall fraction of relevant runs in which the behaviour fired (recall, 0..1).
Competitor skills co-installed via TriggerRateSpec.installSet.
0(the default) means the skill was measured ISOLATED — sorateis an UPPER bound on real recall andfalsePositiveRatea LOWER bound, because selection is competitive (a populated harness can evict or out-compete the description). A non-zero count is the whole-harness measurement.