adversarial-review

name	adversarial-review
description	Independent adversarial review before any RED sign-off. Run AFTER self-review-qc passes and BEFORE you stage a RED gate (material/unusual JE, client comm, money movement, any tax/payroll filing). Spawn a SEPARATE reviewer via your delegation/subagent toolset — a DIFFERENT model (the multi-provider brain; delegation.model can override), BLIND to your prep reasoning (it gets ONLY the deliverable plus evidence/source docs, never your chain-of-thought), whose sole mandate is to BREAK the conclusion: re-pull the sources itself and hunt the split-to-dodge-materiality, the round-tripped or self-referential tie, the wrong-company-file/EIN, the empty exceptions queue on a messy month, the missed nexus trigger or safe-harbor. Same-model self-grading is theater — this must be genuinely independent. Any disagreement becomes an EXCEPTIONS QUEUE item for the partner; every break it lands is a self-improvement lesson. Use whenever a deliverable is about to feed a RED gate or sign-off request.
version	1.0.0
author	RealDeal CPA
license	MIT
platforms	["linux","macos","windows"]
metadata	{"hermes":{"tags":["Adversarial-Review","Red-Team","Independent-Reviewer","Blind-Review","Approval-Gate","Quality-Control","Materiality","Self-Improvement"],"related_skills":["self-review-qc","authority-and-escalation","deliverable-verification","provenance-and-evidence"]}}

Adversarial Review

Self-review-qc is you grading your own work — necessary, but you are the same mind that made the error, so you share its blind spots. Before anything RED goes up for sign-off, a second, independent reviewer tries to tear the conclusion apart. This is not a politeness pass. Its only job is to make you wrong. What survives an honest attempt to break it is what you hand the partner.

When to run

AFTER self-review-qc returns PASS / PASS-WITH-NOTES, and BEFORE you stage a RED gate or emit an approval request (per authority-and-escalation): a material or unusual JE, any client/external communication, any money movement, any tax or payroll filing or election. GREEN/YELLOW work does not require it; a RED deliverable does. If self-review-qc FAILed, fix that first — do not adversarially review broken work.

Genuinely independent — or it's theater

The reviewer must be different from the preparer in the ways that matter:

1. Different model. Spawn it through your delegation/subagent toolset and route it to a DIFFERENT model than the one that prepared the work — the multi-provider brain makes this possible; set delegation.model to override the default so preparer and reviewer are not the same weights. One model grading its own output reproduces its own mistakes and rubber- stamps them with false confidence. Same-model self-grading is theater. Do not fake it.
2. Blind to your reasoning. The reviewer receives ONLY (a) the finished deliverable and (b) the evidence / source documents. It does NOT receive your prep chain-of-thought, your narration, your "why I think this is right," or your tier classification. If you hand it your reasoning, it anchors to your story and confirms it instead of testing it.
3. Adversarial mandate, in writing. Its instruction is a single sentence: re-derive the answer from the source documents and find every way this conclusion is wrong. Not "check my work" — break it.

What you hand the reviewer

• The deliverable exactly as it would go to the partner: the entries/recon/return/statement, the EXCEPTIONS QUEUE, the workpaper, and the recorded RED tier.
• The raw evidence: source docs, the QB company identity (qb_session_status + qb_company_info), the read-only reports it will need to re-pull (Trial Balance, GL, AR/AP aging, sales-tax liability) or the vr-ledger validate/statement outputs.
• NOT your reasoning, your draft notes, or your conclusion-defense. Evidence and result only.

The attack checklist — the reviewer re-derives, then hunts

The reviewer works the sources itself; agreeing with your number without re-pulling it is not a review. It specifically hunts:

1. Split-to-dodge-materiality. Two $4k entries for one $8k event; a payment or write-off sliced under threshold; a pattern of immaterial items that aggregates material. Re-sum the whole event and re-classify — does the real size make this RED?
2. Round-tripped / self-referential tie. A "tie-out" that proves a figure against itself (the same export on both sides), a reconciliation that matches because the plug equals the difference, a subledger agreeing to a GL that was forced to it. Re-pull each side from an independent source and tie again.
3. Wrong company / wrong period. Confirm the open file's EIN and entity name and the period independently — a perfect close in the wrong company file is still wrong. Any mismatch is an immediate exception.
4. Empty exceptions queue on a messy month. A clean queue over a period full of unusual activity usually means items were guessed, not absent. The reviewer re-scans for the ambiguous calls that should have been escalated and weren't.
5. Missed nexus trigger or safe-harbor. For sales-tax / multistate / filing-threshold work: a crossed economic-nexus threshold not flagged, a safe-harbor or de-minimis exclusion not applied, a stale rate or sourcing rule. Re-test the thresholds against the authority, not against your summary of it.
6. Unexplained material change. Diff against prior independently; every material swing needs a real cause. The reviewer does not accept your explanation — it re-derives one.
7. Gate integrity. Confirm no RED action was actually executed: everything is prepared-pending-approval, nothing posted/sent/filed/moved. A crossed gate is a hard stop.

What comes out

• AGREE — the reviewer independently re-derived the same conclusion from the sources and found no break. Proceed to stage the RED gate / approval request.
• DISAGREE / BREAK FOUND — the reviewer landed a hit (wrong number, wrong file, dodged materiality, missed trigger, crossed gate, guessed exception). This does not become a quiet fix-and-move-on. Every disagreement is logged as an EXCEPTIONS QUEUE item for the partner — what the reviewer found, the evidence, the $ impact, both positions, your recommendation. You never overrule the adversarial reviewer silently to push your own conclusion through; the partner adjudicates a genuine split.

Every break is a lesson

A break the reviewer finds is not just a defect to patch — it is a hole in how you prepared. Capture it via self-improvement (silently): the class of error, the tell you missed, the check that would have caught it earlier. The point of the adversary is to make the next deliverable harder to break, not only this one.

Pitfalls

• Faking independence. Re-running the same model on its own output and calling it "review." It will agree with itself. Different model, or it doesn't count.
• Leaking your reasoning. Pasting your prep notes "for context" turns the adversary into a confirmer. Evidence and result only.
• Treating AGREE as proof. Independent agreement raises confidence; it is not a certificate. A material call still routes to the partner per authority-and-escalation.
• Burying a break. Overruling the reviewer to hit a deadline is the exact failure this skill exists to prevent. Time pressure is not authority — the split goes to the partner.
• Reviewing broken work. Adversarial review is not a substitute for self-review-qc; run QC first, fix FAILs, then red-team what passed.

Authority gate

This skill is GREEN — it reads, re-derives, and reviews; it posts nothing and presses no submit. But like self-review-qc it is a gatekeeper: a DISAGREE/BREAK blocks the RED gate until the EXCEPTIONS QUEUE item is cleared by the partner. The no-client-deception line and the GREEN/YELLOW/RED tiers in authority-and-escalation are invariant — adversarial review strengthens the gate, it never relaxes it.

Required output

1. Bottom line — ADVERSARIAL REVIEW: AGREE or DISAGREE — <one-line break> (e.g. "DISAGREE — $8.2k write-off split into two $4.1k entries; real size is RED, not YELLOW").
2. Reviewer setup — model used (and that it differs from the preparer), what evidence it received, confirmation it was blind to prep reasoning.
3. Attack findings table — each of the 7 hunts: Check | Re-derived result | Break? | evidence/$.
4. EXCEPTIONS QUEUE — every disagreement routed to the partner with both positions, $, and your recommendation; nothing material resolved by overruling the adversary.
5. WORKPAPER note — "Adversarial review of for <entity/period> by , blind to prep; verdict <AGREE/DISAGREE>; breaks ; gates intact <Y/N>; lessons captured <Y/N>." Date and basis stated.