Adversarial Verification — QueueAI Patterns

The Problem

LLM agents are prone to confident hallucination — they produce findings that sound credible but are wrong, incomplete, or miscontextualized. A single reviewer can't catch its own blind spots. If you trust first-pass findings directly, you ship false positives.

The Pattern

Finding
   │
   ├──────────────────┐
   ▼                  ▼
[Skeptic 1]      [Skeptic 2]    (each tries to REFUTE the finding)
   │                  │
   └────────┬─────────┘
            ▼
      ┌──────────┐
      │  Verdict │  (majority vote: ≥2 of 3 must NOT refute)
      └──────────┘
            │
     ┌──────┴──────┐
     ▼             ▼
 Confirmed      Rejected

When to Use

Code review, security analysis, or any task where false positives are costly
After a fan-out phase where you need to filter noise from signal
When a finding being wrong has meaningful consequences

When NOT to Use

Tasks where speed matters more than accuracy
Obvious findings where adversarial review adds cost without value
When the finding is inherently subjective (skeptics won't converge)

Code Example

import { QueueAI } from "@queueai/sdk"

const queue = new QueueAI({ apiKey: process.env.QUEUEAI_API_KEY })

const task = await queue.submit({
  name: "codebase-review",
  input: {
    repoUrl: "https://github.com/your-org/your-repo",
    focus: "security",
    verification: "adversarial",
    verifierCount: 3,
  },
  model: { primary: "anthropic/claude-opus-4-8" },
  budget: { maxCostUsd: 5.00 },
})

from queueai import QueueAI

queue = QueueAI(api_key=os.environ["QUEUEAI_API_KEY"])

task = queue.submit(
    name="codebase-review",
    input={
        "repo_url": "https://github.com/your-org/your-repo",
        "focus": "security",
        "verification": "adversarial",
        "verifier_count": 3,
    },
    model={"primary": "anthropic/claude-opus-4-8"},
    budget={"max_cost_usd": 5.00},
)

See it in action: This pattern powers the Codebase Review template in QueueAI.