sethdford /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:
- 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
- Each emits
VERDICT: pass|fail,CONFIDENCE: 0.0-1.0,RATIONALE, andRESULT_<aspect>-verifier=PASS|FAIL. - Aggregator computes confidence-weighted vote:
pass_share = sum(conf where pass) / sum(conf)pass_share > 0.6→ PASSpass_share < 0.4→ FAIL0.4 ≤ 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)