On consequential tasks — client-facing copy, financial entries, lead classification, bulk data operations —
the agent generates k independent completions for the same task, scores them against the skill rubric,
and either accepts the consensus answer or flags divergence for a reflection cycle.
When k completions disagree, that disagreement is surfaced to the operator rather than resolved arbitrarily.
The operator's resolution updates the rubric for the next run. One generation with no vote is a chatbot.
k generations, scored, with divergence detection is a quality gate.
▸ k-vote fires on generation Before the output reaches anyone, independent completions are scored.
▸ Reflection fires on completion After the work runs, the agent checks if it solved the actual problem.
▸ System-of-record verify fires on declare done Re-fetch, sample, confirm. Self-reports do not count.