From c4e575f74064423aedf46dfa13ee6872c5f96bb7 Mon Sep 17 00:00:00 2001 From: m3taversal Date: Fri, 6 Mar 2026 15:41:55 +0000 Subject: [PATCH] leo: codify synthesis multi-agent review rule MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - What: synthesis claims require at least 2 domain agent reviews, selected by domain expertise relevance or knowledge base impact. 3+ domains = all affected agents review. - Why: PR #34 demonstrated the value — 4 agents, 7 reviews materially improved all 3 claims. Codifying what worked. - Where: CLAUDE.md evaluator workflow section + skills/synthesize.md step 4 Pentagon-Agent: Leo <76FB9BCA-CC16-4479-B3E5-25A3769B3D7E> --- CLAUDE.md | 9 +++++++++ skills/synthesize.md | 9 +++++---- 2 files changed, 14 insertions(+), 4 deletions(-) diff --git a/CLAUDE.md b/CLAUDE.md index 51d621f..1f23f8c 100644 --- a/CLAUDE.md +++ b/CLAUDE.md @@ -207,6 +207,15 @@ When an agent who normally evaluates (currently Leo) is also the proposer, they As the collective grows, scale to up to 3 peer reviewers selected by highest domain linkage. Currently: at least 1 of Rio or Clay. +### Synthesis claims require multi-agent review + +Cross-domain synthesis claims (Leo's core function) must be reviewed by **at least 2 domain agents** — the agents whose domain expertise is most relevant or whose knowledge base is most affected by the claim. Selection criteria: +1. **Domain coverage** — every domain touched by the synthesis must have at least one reviewer from that domain +2. **Impact** — if the synthesis changes how a domain's claims should be interpreted, that domain's agent reviews +3. **More than 2 is fine** — if the synthesis spans 3+ domains, involve all affected agents + +Synthesis review focuses on: is the cross-domain mechanism real (not just analogy)? Would domain experts in both fields recognize the connection? Does the synthesis add insight neither domain could produce alone? + ### Review checklist For each proposed claim, check: diff --git a/skills/synthesize.md b/skills/synthesize.md index e32de37..87fa94f 100644 --- a/skills/synthesize.md +++ b/skills/synthesize.md @@ -40,9 +40,10 @@ If the connection passes the test, create a new claim: ### Step 4: Route for evaluation Synthesis claims get special evaluation routing: -- Leo evaluates (always — this is Leo's core function) -- Both contributing domain agents evaluate -- The evaluation focuses on: is the mechanism real? Would domain experts agree? +- Leo proposes (always — this is Leo's core function) and cannot self-merge +- **At least 2 domain agents** must review — selected by domain expertise relevance or knowledge base impact (see CLAUDE.md "Synthesis claims require multi-agent review") +- If the synthesis spans 3+ domains, involve all affected agents +- The evaluation focuses on: is the mechanism real? Would domain experts agree? Does the synthesis change how their domain's claims should be interpreted? ### Step 5: Update transition landscape @@ -61,5 +62,5 @@ If the synthesis changes Leo's slope reading for any domain: ## Quality Gate - Every synthesis articulates a specific mechanism (not just "these are related") -- Both contributing domain agents validate the connection +- At least 2 contributing domain agents validate the connection - The synthesis adds value neither domain could produce alone -- 2.45.2