Recruiting briefs with receipts — every claim cites a resume line — step 2 of 8

A claim with a receipt is a different object

Run it. That's a screening brief the way this studio builds them: the resume indexed by line, and every claim carrying either a pointer into it or an honest label.

Walk the structure:

• RESUME as numbered lines. The resume stops being a PDF you skim and becomes an addressable document. L2 is a specific, quotable line. When the hiring manager asks "where does it say multi-state?", the answer is a lookup, not a memory.
• receipt as a line id. "Has run multi-state payroll at scale" cites L2, and L2 actually supports it — 1,400 employees, six states. The claim and its evidence travel together. Anyone auditing the brief — you, the hiring manager, counsel two years from now — follows the pointer and reads the line.
• label as the honesty bit. The fourth claim has no receipt, because no resume line says anything about deadlines. The model inferred it from the vibe of the document, the way models do. Under this structure that's allowed — inference is a legitimate part of screening judgment — but it's labeled, so the human weighing the brief knows which statements stand on paper and which stand on pattern-matching. The label is what separates "here's my read" from "here's a fact," and briefs that blur that line are how a routine screen becomes an exhibit.

Notice what the structure makes impossible to do quietly. A claim with "label": "sourced" and "receipt": None is a contradiction you can detect in one if statement — a claim asserting it stands on the resume while pointing at nothing. A receipt like "L9" that doesn't exist in the resume is detectable the same way. And the audit checks the receipt before it reads the label: a claim labeled inference that nevertheless carries a valid receipt counts as SOURCED — the label records the drafter's confidence, the receipt records the evidence, and evidence outranks confidence in both directions. The audit loop you'll write in this lesson catches both, which means the failure mode of every AI screening summary — confident claims with no traceable source — becomes a mechanical check instead of a careful read you have to remember to do per candidate, per req, forever.

One thing the structure deliberately does not contain: a score. No ranking field, no fit percentage, no advance/reject flag. The brief describes; it doesn't decide. Why that absence is load-bearing — and regulated — is step six.