Dual-Pool Adversarial Review

🧩 Digital Twin OS · 对抗审查核心模块 — 自我净化闭环的纠错引擎. AI-assisted code & content review using real people's documented philosophies, not abstract roles. Fixed pool for depth, random pool for surprise. Manager-curated per task. Principle-first with confidence levels — never fabricate a quote.

Status: Methodology + Reference Implementation — This repo describes and demonstrates the dual-pool adversarial review methodology. It includes a reference Claude Code skill and pre-sourced persona principles. It is NOT a standalone tool.

What This Is

This repository is three things:

1. A methodology paper — METHODOLOGY.md describes the dual-pool adversarial review approach: why it works, how it's structured, and what problems it solves. This is the primary artifact.

2. A reference implementation — SKILL.md is a Claude Code skill that implements the methodology. It produces structured, confidence-tagged code reviews using named personas with sourced principles. Use it as-is, adapt it, or treat it as a worked example.

3. A sourcing backbone — persona-principles.md documents the engineering, product, design, and management principles of 20+ named figures, each with verifiable sources and confidence levels. This is the anti-fabrication foundation — every attribution carries a confidence tag (high/moderate/low). If a persona's principle cannot be sourced, drop the persona.

What's Missing for Tool Status

This is honest about what it is and isn't. The following would be needed to call it a tool:

1. Standalone CLI — The review currently runs inside Claude Code as a skill. There's no pip install-able package with a dual-pool-review command.
2. Programmatic validation — Confidence levels and principle sourcing are manual. A tool would verify attribution against persona-principles.md automatically.
3. CI integration — No GitHub Actions, pre-commit hooks, or bot accounts that run reviews automatically on PRs.
4. Test suite — The methodology is validated through self-review (16 issues found and fixed) and community feedback (PR #866, #867 on alirezarezvani/claude-skills), but there's no automated test suite.
5. External validation — One-team validation on one repo's PRs. Broader validation across different codebases and teams is needed.

Why This Exists

AI review systems typically use abstract roles — Saboteur, New Hire, Security Auditor. Abstract roles produce generic findings. "Consider adding error handling" isn't actionable.

This system uses named real people with sourced, citable principles. Linus Torvalds doesn't say "check for code quality." He says "eliminate the special case entirely" (TED 2016 — confidence: high). That difference turns "fix this" into "delete this."

v1.1 (2026-07-01): Anti-fabrication discipline. Every attribution carries a confidence level (high/moderate/low). Cite the principle, not an invented verbatim quote. Drop a persona rather than fabricate. See persona-principles.md for every persona's pre-sourced principles + sources + confidence levels.

Architecture

Task enters
  |
  v
Triage Review (v1.0, 1 round)  <-- determines depth
  |
  +-- quick/standard --> v1.0 (simple, fast)
  |
  +-- deep --> v2.3 Dual-Pool Orchestration
                  |
        +---------+---------+
        v                   v
   Fixed Pool           Random Pool
   (digital-twin       (web-searched
    matched)            fresh each time)
   stability            surprise
        |                   |
        +---------+---------+
                  v
        Cross-pool rounds
        explore<-->exploit

The Two Pools

Fixed Pool (Convergence)

• Matched to the user's digital twin (expertise, personality, goals)
• 2 managers + 9 workers across 5 roles
• Strengths: depth, consistency, personal relevance, pre-sourced principles
• Risk: echo chamber — needs periodic audit

Random Pool (Divergence)

• Fresh personas via web search each session
• No preset list — manager defines search keywords
• Strengths: surprise insights, blind-spot coverage, diversity
• Risk: quality variance, principles not pre-sourced (confidence ceiling: moderate)

Orchestration Modes

• exploit-only: all fixed pool (repeatable tasks)
• explore-only: all random pool (inspiration tasks)
• explore->exploit: random first, fixed verify (strategy decisions)
• exploit->explore: fixed baseline, random challenge (portfolio review)

Per-Round Process

1. Randomly draw a manager from current pool
2. Manager analyzes task → decides depth + required roles (2 engineers + 1 product + optional auditor)
3. Manager recruits specific workers from the worker pool
4. Team reviews — each person applies their documented philosophy. Every finding: sourced principle + confidence level. Zero-finding burden: 3+ principles the code satisfies + how
5. Output: [Manager] picked [A,B,C] because [reason]. Found N issues. Verdict: BLOCK/CONCERNS/CLEAN
6. Next round: new manager + keep ≤2 members from previous round

Note: earlier versions required each reviewer to find at least one issue. Community feedback correctly identified this as counterproductive — forced findings produce exactly the generic noise this method exists to escape. The refined approach: principle-first review with symmetric burden (findings need sourced principles, non-findings need equal justification).

Key Innovations

1. Real people, not abstract roles — searchable, citable, sourced philosophies
2. Principle-first, not quote-first — cite documented stances with confidence levels, never fabricate verbatim quotes (v1.1)
3. Dual-pool cross-orchestration — depth + surprise, not repetition
4. Manager curation per task — dynamic team, not fixed template
5. Pool evolution — promote (random→fixed), demote (stale→removed)
6. Digital twin matching — fixed pool curated to user's profile

Validation

Tested 2026-06-27 on PR #866 (alirezarezvani/claude-skills, 18.7K stars):

• R1 Fixed (McCord): 10 findings — structure, format, adoption
• R2 Fixed (Catmull): 8 findings — clarity, edge cases, UX
• R3 Random (Spolsky+DuVander): 3 findings — positioning, first impression, output destination

Random pool found things both fixed-pool rounds completely missed. Cross-pool works.

Anti-fabrication discipline (confidence levels, drop-persona rule) hardened 2026-07-01 based on alirezarezvani/claude-skills PR #867.

Limitations

• More complex than simple review (needs triage judgment)
• Fixed pool needs periodic maintenance
• Random pool depends on web search quality
• Not all tasks need deep mode
• No standalone tooling (CLI, CI integration, automated validation)

Files

File	Purpose
README.md	This document — methodology overview and honest positioning
SKILL.md	v2.3 skill — installable in Claude Code, principle-first with confidence levels
persona-pool.md	v2.3 fixed pool design — managers, workers, roles, evolution, anti-fabrication
persona-principles.md	Sourcing backbone — every persona's documented principles + sources + confidence levels
METHODOLOGY.md	Full methodology paper — architecture, rules, validation, risk analysis
LICENSE	MIT

Related

• checkgrow — Full AI quality toolkit
• PR #866 — original skill implementation
• PR #867 — anti-fabrication hardening (confidence levels, principle grounding)
• PR #2365 — delivery-gate

Acknowledgments

This methodology was shaped by direct feedback from reviewers and maintainers across three open-source communities. Each person's contribution is described as specifically as the public record allows — no endorsement beyond what they actually wrote is implied.

xg-gh-25 (SwarmAI) reviewed the self-audit skill PR (anthropics/skills, 2026-06-29). Their detailed review suggested a Step 0 mechanical artifact check and requirement traceability in the Completeness dimension, referencing their T-CBB pattern from SwarmAI. Both became core features of the self-audit methodology.

alirezarezvani (maintainer, alirezarezvani/claude-skills) reviewed and merged the named-persona skill (2026-07-01). Beyond repo integration, his review directly added the anti-fabrication discipline that is now foundational to this methodology: cite principles with confidence levels (not fabricated verbatim quotes), drop a persona rather than invent, and ground every persona in verifiable sources via persona-principles.md.

daltino (maintainer, affaan-m/ECC) reviewed and approved the delivery-gate PR (2026-06-26), validating that a reasoning-quality gate (distinct from code-quality verification) is a legitimate complement to existing CI pipelines.

Any errors or shortcomings in this methodology are the author's, not the reviewers'.

Part of gategrow

Repo	What
checkgrow	Unified quality framework — start here
delivery-gate	Stop hook for Claude Code
self-audit	pip install-able four-dimension audit
session-cost	L0→L3 layered cost tracking
dual-pool-review	Named-persona adversarial review methodology

License

MIT