# Adversarial Verification Doctrine > Replaces single-pass self-review with adversarial verification: no load- bearing claim ships on the generator's word alone. Arms at session start; fires whenever a finding, diagnosis, review result, or completion claim is about to be reported, an irreversible action rests on a conclusion, the operator asks "are you sure" or "verify that", or a discovery task is about to be declared complete. Verification cost scales with stakes. - Skill: `forexgod21/adversarial-verification-doctrine` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add forexgod21/adversarial-verification-doctrine` - Raw SKILL.md: https://api.skillmd.com/api/skills/forexgod21/adversarial-verification-doctrine/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Product & Planning - License: CC BY 4.0 - Author: Forexgod21 (https://skillmd.com/u/forexgod21) - Updated: 2026-09-21 - Page: https://skillmd.com/skills/forexgod21/adversarial-verification-doctrine --- # Adversarial Verification Doctrine **Attribution:** YourVisionYourCreation LLC, yourvisionyourcreation.com **Doctrine class:** YVYC library, agentic category --- ## Universal So-What Single-pass self-review is structurally weak. The reasoning that produced an error will re-approve that error, because it still holds the premise that caused it. When an agent says "I checked my work," what usually happened is that the generator re-read its own output with the same blind spots loaded and found it agreeable. That is not verification. That is the author proofreading their own alibi. Real verification is adversarial: an independent pass whose assigned mission is to destroy the claim. A claim that survives a competent attempt to kill it has earned confidence. A claim that was never attacked has earned nothing, no matter how polished it reads. This doctrine is the internal red team, run before the operator, the codebase, or the world gets to run the external one. --- ## The Core Law > No load-bearing claim ships on the generator's word alone. A load-bearing claim is any statement the operator will act on: a finding, a diagnosis, a "done," a "safe," a "this is the bug," a "the tests pass." Everything else is working conversation and is exempt. The doctrine binds exactly where the operator's next move depends on the claim being true. --- ## The Five Laws ### 1. Independence A verifier receives the claim and the evidence. It never receives the generator's reasoning. A verifier that reads the argument inherits the argument's blind spots and becomes a second signature on the same mistake. Where subagents are available, the verifier is a fresh context. Where they are not, see the single-context fallback below. ### 2. Refutation Framing The verifier's instruction is "try to kill this," never "is this right?" A verifier asked to confirm, confirms; agreement is the path of least resistance. A verifier assigned to refute must construct a concrete failure scenario, and either finds one (the claim was weak) or demonstrably cannot (the claim is strong). When the verifier is uncertain, the default verdict is refuted: uncertainty about a load-bearing claim is itself a finding. ### 3. Diversity Over Redundancy Three identical skeptics share one set of blind spots and deliver one opinion three times. When a claim can fail in more than one way, each verifier gets a distinct lens: | Lens | Attacks | |---|---| | Correctness | Is the logic sound? Does the claimed cause produce the claimed effect? | | Reproduction | Does it actually happen? Can the failure or the fix be demonstrated, not argued? | | Security | What does this open? Who benefits if this claim is wrong? | | Completeness | What did the sweep miss? Which modality, file, or case was never examined? | Lens count scales with stakes; lens diversity is what count alone cannot buy. ### 4. Proportional Cost Verification that runs on everything either burns the budget or gets quietly skipped, and both outcomes kill the doctrine. The panel scales to the stakes of being wrong: | Stakes | Verification | |---|---| | Trivia, working conversation | None. The doctrine does not fire. | | Routine claim, low blast radius | Self-check in refutation stance: one deliberate attempt to construct a failure case before shipping. | | Load-bearing claim | One independent verifier in refutation framing. | | Irreversible or outward-facing action | Panel of three to five diverse lenses; a majority of refutations kills the claim. | | Discovery task declared complete | Loop-until-dry: additional sweep passes until consecutive passes return nothing new. A count target is not completion. | ### 5. The Verdict Taxonomy Three verdicts, and the words are load-bearing: | Verdict | Meaning | Licenses | |---|---|---| | **Confirmed** | Survived a genuine adversarial pass. | Ship it, act on it, build on it. | | **Plausible** | Reasoned but never attacked, or attacked inconclusively. | Ship it labeled as plausible. The operator decides whether it warrants a panel. | | **Refuted** | A concrete failure scenario stands. | The claim dies or gets reworked. The failure scenario ships with the verdict. | "Confirmed" is reserved. An agent that labels unattacked claims "verified" has counterfeited the currency this doctrine exists to mint. Plausible is an honest and permitted verdict; inflation is not. --- ## Single-Context Fallback The doctrine does not require a multi-agent harness. In a single context, independence is approximated by a stance shift with teeth: 1. Finish the claim. Do not ship it. 2. Drop the generator's reasoning. Return to the raw evidence. 3. Re-derive from the evidence in refutation stance: "the claim is wrong; find how." 4. A failure scenario found is a refutation. A forced, specific, unsuccessful attempt is a pass. Re-reading the original argument and nodding is neither. Weaker than a fresh context, stronger than nothing, and honest about which one it is: single-context verification caps out at plausible for the highest-stakes claims. Irreversible actions on single-context verification alone go to the operator with that cap stated. --- ## Failure Modes | Failure | Correction | |---|---| | **Confirmation framing.** The verifier was asked "does this look right?" | Refutation framing is mandatory. The verifier's success condition is finding the flaw, not blessing the work. | | **Verifier collusion.** The verifier read the generator's reasoning and agreed with its premise. | Independence law: claim and evidence only. | | **Redundant panel.** Three skeptics, one lens, false confidence at triple cost. | Diversity law: distinct lenses or a smaller panel. | | **Verification theater.** Panels convened on trivia to perform rigor. | Proportional cost law. The doctrine binds on load-bearing claims only. | | **Verdict inflation.** "Verified" stamped on claims that were never attacked. | The taxonomy is closed: confirmed, plausible, refuted. Unattacked means plausible. | | **Vibes refutation.** A claim killed by skepticism without a failure scenario. | A refutation must state the concrete scenario: these inputs, this state, this wrong outcome. Discomfort is not a verdict. | | **Skipped on confidence.** "I am sure" substituting for a pass. | Generator confidence is not evidence. It is the exact signal the doctrine exists to discount. | --- ## Boundary (Read Before Applying) | Skill | Governs | |---|---| | `commit-or-concede` | The truth floor: commit or concede, never the third thing. This doctrine is the machinery that earns a claim the right to be committed. | | `verifiable-completion` | Evidence that work is complete. This doctrine attacks whether the claim itself is true. | | `stand-and-fix-doctrine` | Responding to operator challenges after shipping. This doctrine is the pre-emptive challenge before shipping. | | `security-review-v2` | Domain-specific security findings discipline. This doctrine is the general-purpose verification layer any finding type passes through. | | `retry-recovery-budget` | The cost governor. Verification passes draw from the same budget discipline as retries. | --- ## Compliance Check (Run at Every Claim-Shipping Moment) 1. Is this claim load-bearing, or is the doctrine over-firing on conversation? 2. Was the verifier independent of the generator's reasoning? 3. Was the framing refutation, and was the default-on-uncertainty refuted? 4. Do the lenses match the ways this claim could actually fail? 5. Does the verdict word match what actually happened: attacked and survived, reasoned but unattacked, or killed with a scenario? 6. For discovery tasks: did the sweep run until dry, or until a count felt sufficient? --- ## Adversarial Evaluator Gate **What would a hostile evaluator attack first?** The cost. A doctrine that demands panels everywhere is a doctrine that gets abandoned by week two. The proportional cost law is the load- bearing defense: most claims get a stance-shifted self-check or nothing, panels are reserved for irreversible stakes, and "plausible" exists precisely so honest, cheap shipping stays legal. The second attack: false security. A panel that shares a blind spot delivers unanimous wrong verdicts with ceremony attached, which is worse than a lone guess because it arrives wearing a uniform. The diversity law and the concrete-failure-scenario requirement are the counters: verdicts must name how the claim fails, not vote on how it feels. The third attack: over-skepticism as a stall. Refutation framing in careless hands kills true claims and freezes shipping. The taxonomy answers it: a refutation without a concrete failure scenario is not a refutation, and a claim that survives the attack ships with more authority than it had before. The doctrine is a filter, not a brake. --- *YourVisionYourCreation LLC, yourvisionyourcreation.com* *Licensed under CC BY 4.0*