Cross model verification
Skill lawzava/megapowers/plugins/mega-orchestration/skills/cross-model-verification
Skills, plugins, and hooks that change how coding agents work — Claude Code, Codex, OpenCode, Antigravity. Measured, not vibed.
npx -y skills add lawzava/megapowers --skill cross-model-verificationAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Use when risky billing, auth, concurrency, security, or other high-stakes work needs another vendor to independently refute or adversarially verify it.
The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
3.1 KB, as published. Nobody here has run it
Cross-Model Verification
A second model catches what the first is blind to only if it is independent. Independence has two parts: the verifier comes from a different vendor than the author, and it never sees the author's reasoning or conclusion. A verifier that sees the prior conclusion anchors to it and confirms; a blind verifier keeps its edge.
Prefer an oracle to an opinion
Before asking a model, ask whether an executable check can decide it: tests, types, a compile, a property test, a reproduction. An oracle is deterministic and not fooled by a confident argument. Reserve model verification for what no oracle covers: design soundness, subtle logic, security reasoning, "does this actually do what it claims".
Procedure
-
Launch a verifier with the author vendor declared via multi-agent-delegation's
scripts/delegate-run --role verify --author-vendor VENDOR --artifact ... --claim .... The resolver fails if it cannot route away from every declared author vendor. A second instance of the same model shares the same blind spots, and self-preference bias is largest when a model judges its own family's output (arXiv 2410.21819). -
Hand over the artifact and the claim, nothing else. The verifier gets the diff, code, or document plus a crisp statement of what it is supposed to do or guarantee. Withhold the author's chain-of-thought, self-review, and justification. Information restriction is the point.
-
Prompt it to refute, with the burden of proof on "verified". Ask for the bug, the counterexample, the missed case, and a default of not verified under any real doubt. The resolved provider's reference file (multi-agent-delegation,
references/providers/) carries the adversarial template and output schema that make the verdict machine-checkable. -
Escalate to a perspective-diverse panel for high stakes. Run several independent verifiers, each with a distinct lens: correctness, security, concurrency, reproduction. The panel exists for coverage, not voting: one credible refutation from any lens kills the claim, no matter how many other passes said fine.
-
Act on the verdict as single writer. The verifier reports; it never merges its own fix. The lead applies changes and re-verifies material ones. Never trust a self-reported pass; re-run the oracle.
Guardrails
- Different vendor, or it isn't independent. Same-model "review" is a consistency check, not verification.
- Keep the launcher's receipt. It records which model verified what claim, against the exact artifact identity, and becomes stale after any change.
Relationship to other skills
- Unsure whether verification, selection, or a council fits? Start at mega-orchestration:orchestrating, the decision root.
- Serves as the blind judge in mega-orchestration:best-of-n when no oracle can rank candidates, and as the per-answer scrutiny step in mega-orchestration:council-adjudication.