Account-intelligence analyst (question → graded answer)

The insight: the other research plays fail by leaving a field blank; this one fails by answering. A brief that cannot find a company's headcount ships an empty field, and the gap is visible. But a question — "are they building an AI team?" — always has a plausible answer available from adjacent evidence: the homepage says "AI-powered", so the answer looks like yes. That is marketing copy, not a hiring signal, and nothing in the output would reveal the substitution. A confident answer from thin evidence is indistinguishable from a confident answer from good evidence, unless the skill reports what it actually observed.

So the question is never answered directly. It is decomposed into proxies — specific things that could be observed, each with a named arm that observes it and a declared cost — and the answer is graded by which proxies actually fired, not by how the prose reads. An account where nothing observable fired returns insufficient evidence, which is a real answer and the most commonly correct one.

Two consequences that shape every step below: