Six auditors attack the artifacts, not the idea
Each auditor reads the PRD, specs, and the prototype's actual code, opens with an explicit kill-condition check, and files findings in the shared risk contract. Nobody returned "looks great." Nobody was asked to be nice.
What this stage wrote to disk
6 agents · 6 documents · 869 linesEach auditor's headline, verbatim
Business viability sharpest finding
"A CFO reading the PRD's three success metrics would notice they are all UX/adoption metrics — none of them is a business metric. That gap, not the segment size, is the sharpest finding here."
Accessibility release-gating
"Drag-to-reorder has zero keyboard/screen-reader/switch path in the demo — not a partial implementation, a complete absence" on the exact interaction the spec calls "a build requirement, not a nicety."
Ethics & trust/safety needs ratification
Two open design questions "will decide whether 'hide' functionally reduces what a user sees, or just relocates it while preserving the exposure Anthropic has a financial stake in."
Privacy & security kill check: not triggered
"Two things in this run do clear a 'lead with it' bar, and both live in the instrumentation, not the personalization."
Legal & compliance kill check: not triggered
"The nav-preference record itself is legally uninteresting; treat it like a theme setting." Four areas explicitly cleared with reasoning instead of being padded into the list.
Operational readiness unstaffed
"Operational readiness is close to unstaffed… no workstream produces a single piece of support tooling," so "my menu changed" becomes an unfixable ticket.
Note what the sweep is doing: the accessibility auditor recomputed the design spec's contrast ratios from the prototype's actual hex values rather than trusting them (they checked out), and the ethics auditor credited the PRD's guardrails before naming the one decision that could turn "hide" into a dark pattern. Praise and attack in the same documents, because the rubric rewards accuracy, not verdicts.