Three native subagents that verify a deliverable actually matches its spec by checking primary sources directly, never by trusting another agent's completion message.
Every round follows the same sequence, no exceptions. This mirrors the workspace's normal Approval Gate, Research to Brief to Approval to Build, applied to verification work itself.
The main session hands Vera exactly two things: the question and the sources. She designs the round, writes the QA-1 plan, the QA-2 plan, and her own verdict plan, and assigns the opposite walk directions herself, one designer, because two independent authors could accidentally pick the same lane. Shannon reviews the package once. If the framing is wrong, she catches it before anything runs, not after.
All three are native subagents in .claude/agents/ai-hr/, running on Opus per the AI HR Model Specification Standard. QA wrong verdicts cost more than speed savings, so this model is never downgraded.
Pulls ground truth and candidate data from primary sources, the real files, the real tables, the real exports. Never accepts a prior agent's summary or a cached list as data. Walks exactly one direction, assigned by Vera.
Independently re-checks QA-1's matches from the opposite direction. Re-derives everything from primary sources by the plan's stated match logic, typically fuzzy company name, never slug (slugs drift: "fresh-victor" vs. "freshvictor").
Reconciles QA-1 and QA-2, computes deltas, explains every count, runs the final checklist, and signs off. Also the sole author of every QA round's three plans, she loads her SOUL file first, always, before anything else.
QA-1 and QA-2 approach from opposite directions. One walk finds what's missing. The reverse walk finds what shouldn't be there. The original 530 failure happened because only one direction was ever checked.
Vera writes this for each member before any checking begins. Shannon reviews the filled-in template, not a description of one.
A green test suite proves the code handles a matched input correctly. It proves nothing about whether real inputs get matched. Both gates are required, Gate A alone is never sufficient.
Code walk, test suite, dry-runs, inverse validation against the ten hard rules. Mocked-boundary tests, a fake Gmail service, a synthetic message dict, a pre-matched contact, are Gate A evidence only.
At least one known real record traced end to end through the live path, from the real external event, to the stamped field, to what the client actually sees. No guessing, and no verifying a system against its own output.
The QA team sits at the end of the delivery chain, after any agent produces a deliverable, before Shannon treats it as done. It applies across the workspace and every client AIOS repo. Client repos don't carry their own copy, the QA team lives here and operates cross-repo.
Vera writes knowledge/learnings/YYYY-MM-DD-qa-round-[name].md after every round and answers four questions.