The least calibrated confidence at which the suggestion is shown (see
DECISION_DISCOVERY_CALIBRATION). Both the Choice's confidence and the Likelihood that any
agent applies must reach it.
Set from the agent-discovery Decision Eval (plan Task 3.1, 2026-09-29): 258 labelled requests, 198
for one of 33 agents and 60 for none, three repeats per model, calibrated out of fold, and scored
as production would, so an answer after DECISION_DISCOVERY_TIMEOUT_MS injects nothing. At a
calibrated 0.85, Jev's suggestion covered 21.7% of the requests meant for an agent, and named the
right agent for 95.3% of those it covered. It was shown for 5.0% of the requests meant for no agent,
all of them multi-agent workflows. Every Jev discovery but the run's first, which loaded the
embedding model, finished within the timeout. LLM Decision finished within it for at least 67.7% of
requests, and covered at least 16.0% at 92.6% precision, shown for 5.0%. A wrong suggestion costs
more than a missed one, which is only the two-turn flow the agent used before, so the threshold
favours precision. On raw probabilities the old 0.7 almost never fired: Jev's raw any-applies
Likelihood never passed 0.8, so it covered 6.9%.
The least calibrated confidence at which the suggestion is shown (see DECISION_DISCOVERY_CALIBRATION). Both the Choice's confidence and the Likelihood that any agent applies must reach it.
Set from the agent-discovery Decision Eval (plan Task 3.1, 2026-09-29): 258 labelled requests, 198 for one of 33 agents and 60 for none, three repeats per model, calibrated out of fold, and scored as production would, so an answer after DECISION_DISCOVERY_TIMEOUT_MS injects nothing. At a calibrated 0.85, Jev's suggestion covered 21.7% of the requests meant for an agent, and named the right agent for 95.3% of those it covered. It was shown for 5.0% of the requests meant for no agent, all of them multi-agent workflows. Every Jev discovery but the run's first, which loaded the embedding model, finished within the timeout. LLM Decision finished within it for at least 67.7% of requests, and covered at least 16.0% at 92.6% precision, shown for 5.0%. A wrong suggestion costs more than a missed one, which is only the two-turn flow the agent used before, so the threshold favours precision. On raw probabilities the old 0.7 almost never fired: Jev's raw any-applies Likelihood never passed 0.8, so it covered 6.9%.