QA and the readout: the audit Deloitte wishes it ran

Every fact in the final deliverable gets exactly one of three stamps. VERIFIED: a human saw the source with their own eyes - the fact stays, and the log row records who checked it. VENDOR-CLAIMED: the fact comes from a party with an interest in it (a vendor's own benchmark, a press-reported company statement) - it stays, but the deliverable labels it as a claim in the text itself. CUT: no source found. No source, no sentence - however plausible, however well it fits the story. The invented statistic from b3 is plausible and well-formatted by construction; that is exactly why "it sounds right" is not an outcome.

And who runs it: not the author. Fresh eyes catch what familiarity forgives - the author reads what they meant, the auditor reads what is on the page. Timebox it, run it from a checklist, and run it every time, not just when the deliverable feels risky. The Deloitte lesson, retold in two sentences: an academic found fabricated references and a misattributed court quote in a government report, errors Deloitte attributed to generative AI use it later disclosed. The partial refund was modest; the headline was the real cost.

Steal this and adapt it: (1) list every number, named quote, and external reference in the final document - this is the fact inventory; (2) for each, find its source-log row from b3; (3) open the source and confirm the fact says what the deliverable says it says; (4) stamp VERIFIED, VENDOR-CLAIMED, or CUT; (5) for every VENDOR-CLAIMED, confirm the label appears in the deliverable text; (6) for every CUT, either find a real source or delete the sentence; (7) sign the audit sheet with your name and the date. Step 7 is the point - Rail 2 says every fact has a human owner, and a signature makes ownership visible.

Why this works: the model is excellent at generating challenge patterns - it has read every objection ever published. Ask for "the ten hardest questions a skeptical CFO would ask about this recommendation" and it will reliably surface the cannibalization worry, the pricing-anchor challenge, the capacity question, the why-now push. This is squarely inside the frontier from session 1: pattern generation, not fact retrieval.

The discipline is what turns the parlor trick into QA: every question gets a one-line answer WITH its source row. "Enterprise-style seats retain at 94% vs 71% for individuals - source log row 12" is an answer. "Retention is strong" is not; the scorer from b5 already taught you why. And the two questions you cannot answer this way are the most valuable output of the exercise: each becomes either week-6 work (go find the answer) or an honest caveat in the appendix (say what you do not know). Both beat improvising in the room.

One more rehearsal, from b7's difficult news: role-play the $15-tier partisan - the stakeholder who wanted the cheaper tier and did not get it. Have AI play them, argue their case at full strength, and rehearse the disagreement before the room does it live. You know this person will be in the readout. Meet them in private first.

A second rehearsal prompt, for the night before: "Assume this recommendation failed in 12 months - write the post-mortem." The model will produce a plausible failure story: enterprise sales cycle longer than modeled, the 210-account pipeline converting worse than hoped, a rival dropping price below $30. Then check the deck: does it answer those failure modes, with a caveat, a mitigation, or a monitoring plan? Where it does not, you have found week-6 work while it still costs nothing. The pre-mortem finds structural risks the question list misses, because it forces a narrative of failure instead of a list of doubts.

Open with the answer - in the first two minutes, restated plainly, because the pre-read already delivered it and pretending otherwise wastes the room's patience. Exhibits on demand, not on parade: the appendix holds every chart you might need; the meeting shows only the ones a question summons. The Q&A is the main event - not the risk at the end of the presentation but the reason the meeting exists, and the part you rehearsed ten questions deep.

What AI does during the readout: nothing. What it did before: everything rehearsable - the question list, the pre-mortem, the pre-read draft, the timing run. That division is session 1's frontier map applied to a meeting: preparation is inside the frontier, the room is not.

And close with the ask, stated as a decision: approve the $30 per seat tier pilot for the 210 accounts already in procurement or security review, name the owner, set the review date. A readout that ends without an ask was a seminar. Northbridge does not bill for seminars.