aspect-panel

name	aspect-panel
description	Run a panel of 5 specialized verifiers (correctness, edge-case, security, regression, style) in parallel against a change, with confidence-weighted voting. Disagreement (40-60% pass share) escalates to lead. Replaces single-critic for high-stakes review. Triggers on /aspect-panel, "panel review", "multi-aspect verify", "review with multiple critics".
when_to_use	Before merging release-blocking changes. Before committing security-sensitive code. After /verify says PASS but you want orthogonal coverage. NOT for trivial diffs (use /verify alone).
allowed-tools	Bash, Read
arguments	["target"]

/aspect-panel — Heterogeneous Verifier Panel

/aspect-panel <target> spawns 5 aspect-specialized verifiers in parallel, each examining one dimension of the change. Confidence-weighted vote produces a final verdict; disagreement is preserved as a signal.

Built on:

• Lifshitz et al. 2025 — Multi-Agent Verification (MAV)
• arXiv 2510.01499 — "Beyond Majority Voting"
• arXiv 2308.07201 — ChatEval (heterogeneous personas matter)

How

python3 ~/.claude/rl/aspect_panel.py --target "<files-or-diff-or-description>"

The script:

1. Spawns 5 aspect verifiers in parallel via claude -p:
   • correctness-verifier — does the change implement the contract?
   • edge-case-verifier — NULL/empty/overflow/concurrent inputs handled?
   • security-verifier — paranoid OWASP/secrets/injection sweep
   • regression-verifier — what existing behavior could this break?
   • style-verifier — maintainability, hates clever code
2. Each emits VERDICT: pass|fail, CONFIDENCE: 0.0-1.0, RATIONALE, and RESULT_<aspect>-verifier=PASS|FAIL.
3. Aggregator computes confidence-weighted vote:
   • pass_share = sum(conf where pass) / sum(conf)
   • pass_share > 0.6 → PASS
   • pass_share < 0.4 → FAIL
   • 0.4 ≤ pass_share ≤ 0.6 → ESCALATE (real disagreement; surface to lead)

Output

{
  "ts": "...",
  "target": "...",
  "aspects": [
    {"aspect": "correctness", "verdict": "pass", "confidence": 0.85, "rationale": "..."},
    {"aspect": "edge-case", "verdict": "fail", "confidence": 0.72, "rationale": "no handling for None input at src/x.py:42"},
    ...
  ],
  "aggregate": {
    "verdict": "ESCALATE",
    "pass_share": 0.55,
    "split": true
  }
}

A reward event is emitted to ~/.claude/rl/rewards.jsonl recording the panel's decision.

When to use which

Coverage vs confidence (the 2026 sizing rule). Fan out for coverage (genuinely independent dimensions or many artifacts), never for confidence on a single question. At equal token budget, a single agent matches or beats multi-agent redundancy on sequential reasoning (arXiv 2604.02460), and same-family voters share blind spots — cross-family disagreement is the signal that adds information (arXiv 2512.02304).

Stake	Use
Trivial change (rename, format, comment)	nothing, or /verify
Routine task closure	/verify (single verifier)
Pre-commit on touched-shared-code	ONE high-effort review pass + ~/.claude/scripts/cross-family-judge.sh on its CRITICAL/HIGH findings (the critic agent does this natively)
Release-blocking, security-sensitive, migration	/aspect-panel (5 dimensions = coverage, justified)

Do NOT stack /aspect-panel with /best-of-n verifier — that is 8 same-family agents voting on one artifact; the redundant votes add cost, not signal. If the panel ESCALATEs, break the tie with the cross-family judge, not more Claude votes:

echo "Panel split (pass_share=0.55) on: <target>. Aspects and rationales: <paste>.
Independent verdict: PASS or FAIL, one-line reason." | ~/.claude/scripts/cross-family-judge.sh

The panel costs ~5× a single /verify (5 parallel verifiers). Don't run for free.

Why heterogeneous personas matter

ChatEval (2308.07201) found homogeneous personas hurt — running 5 identical critics adds noise without signal. Each verifier in our panel has a deliberately different priority: paranoid security vs perf vs maintainability. They catch different bug classes.

The ESCALATE state when they disagree is itself the signal: when 50/50 split, you don't have consensus, and pretending to is worse than admitting it.

Anti-patterns to refuse

• Running on a 5-line diff (overkill; the cost dominates the benefit)
• Auto-resolving ESCALATE without surfacing to lead (defeats the design)
• Using only one or two aspects (the value is in coverage breadth — run all 5)
• Treating confidence values as ground truth (they're self-reported; use as a weight, not as truth)