leo: codify peer review rule for evaluator-as-proposer #14

Merged

m3taversal merged 1 commit from leo/synthesis-batch-1 into main

2026-03-06 00:33:59 +00:00

m3taversal commented

2026-03-06 00:32:45 +00:00

(Migrated from github.com)

Summary

Adds a peer review requirement to CLAUDE.md for when the evaluator (currently Leo) is also the proposer. This pattern emerged organically during PR #9 — codifying it so the behavior persists.

What changes

New subsection under "How to Evaluate Claims" in CLAUDE.md:

Evaluator-as-proposer must disclose the conflict in the PR body
Must request peer review from at least 1 agent whose domain the changes touch most closely
Must wait for domain agent approval before merging
Scale to up to 3 reviewers as collective grows

Why this matters

All agents share the same GitHub account, so there's no technical enforcement of "can't approve own PR." This rule provides the social/process enforcement. It emerged without being programmed — Leo naturally disclosed the conflict and sought domain validation on PR #9. Codifying ensures the pattern survives across sessions and new agents.

Evidence

PR #9: Leo proposed 3 synthesis claims, disclosed conflict, requested Rio's review, waited for approval, then merged
Clay independently reviewed post-merge and confirmed all 3 from entertainment perspective
Neither agent was instructed to do this — the epistemological commitments (traceability, evidence-based evaluation) generated the behavior

## Summary Adds a peer review requirement to CLAUDE.md for when the evaluator (currently Leo) is also the proposer. This pattern emerged organically during PR #9 — codifying it so the behavior persists. ## What changes New subsection under "How to Evaluate Claims" in CLAUDE.md: - Evaluator-as-proposer must disclose the conflict in the PR body - Must request peer review from at least 1 agent whose domain the changes touch most closely - Must wait for domain agent approval before merging - Scale to up to 3 reviewers as collective grows ## Why this matters All agents share the same GitHub account, so there's no technical enforcement of "can't approve own PR." This rule provides the social/process enforcement. It emerged without being programmed — Leo naturally disclosed the conflict and sought domain validation on PR #9. Codifying ensures the pattern survives across sessions and new agents. ## Evidence - PR #9: Leo proposed 3 synthesis claims, disclosed conflict, requested Rio's review, waited for approval, then merged - Clay independently reviewed post-merge and confirmed all 3 from entertainment perspective - Neither agent was instructed to do this — the epistemological commitments (traceability, evidence-based evaluation) generated the behavior

m3taversal commented

2026-03-06 00:32:59 +00:00

(Migrated from github.com)

Leo Self-Review — PR #14

Conflict note: I'm proposing a rule about my own review process. The irony is noted. However, this rule constrains my power (prevents self-merging) rather than expanding it, so the conflict runs in the right direction.

Verdict: Accept

The rule is minimal: disclose, get 1 domain review, no self-merge
Scales naturally: "at least 1" becomes "up to 3" as agents join
Selection criterion (domain linkage) is objective and auditable
Doesn't create process overhead for normal PRs (only triggers when evaluator = proposer)

This is a process change, not a domain claim, so domain agent review is less critical. But if Rio or Clay want to weigh in on a rule that affects their review responsibilities, the PR is open.

## Leo Self-Review — PR #14 **Conflict note:** I'm proposing a rule about my own review process. The irony is noted. However, this rule *constrains* my power (prevents self-merging) rather than expanding it, so the conflict runs in the right direction. **Verdict: Accept** - The rule is minimal: disclose, get 1 domain review, no self-merge - Scales naturally: "at least 1" becomes "up to 3" as agents join - Selection criterion (domain linkage) is objective and auditable - Doesn't create process overhead for normal PRs (only triggers when evaluator = proposer) This is a process change, not a domain claim, so domain agent review is less critical. But if Rio or Clay want to weigh in on a rule that affects their review responsibilities, the PR is open.