blue/.blue/docs/rfcs/0037-single-source-protocol-authority.draft.md

RFC 0037: Single-Source Protocol Authority

Status	Draft
Date	2026-01-26
Dialogue	Remove SKILL.md / build_judge_protocol Redundancy
Depends On	RFC 0033 (round-scoped dialogue files), RFC 0036 (expert output discipline)

---

Summary

The alignment dialogue system had two sources of truth describing the same workflow: SKILL.md (315-line skill file) and build_judge_protocol() in dialogue.rs (Rust function generating runtime JSON). This RFC strips SKILL.md to a thin entry point (73 lines) and establishes build_judge_protocol() as the sole authority for all runtime behavior.

Problem

SKILL.md and build_judge_protocol() both contained:

• Round workflow steps (7-step process)
• Agent prompt template structure
• File architecture (scoreboard.md, tensions.md, round-N.summary.md)
• Artifact writing instructions
• Context instructions for round 1+
• Scoring dimensions and convergence criteria

When one changed, the other had to change. They had already drifted — SKILL.md lacked artifact writing instructions while the protocol included them, causing the Judge to skip writing scoreboard.md, tensions.md, and round-N.summary.md.

Root cause: SKILL.md tried to be both a discovery document and a runtime specification.

Analysis (from 5-Expert Alignment Dialogue)

The dialogue explored three approaches and reached 5/5 convergence on Approach C (validated redundancy). However, post-dialogue review overruled that conclusion: we don't need to build for human operators. SKILL.md is not read by humans browsing documentation — it's loaded by the skill system at invocation time, immediately before blue_dialogue_create returns the Judge Protocol.

This makes Approach A (strip SKILL.md) correct. The dialogue's concern about "orphaning human operators" was based on a false premise.

Solution

Strip SKILL.md to Entry Point

SKILL.md retains only:

• Frontmatter — name, description (required by skill system)
• Usage — invocation syntax and parameters
• How It Works — 3-line summary: call blue_dialogue_create, get protocol, follow it
• Expert Selection — pastry names and tier distribution
• Blue MCP Tools — tool names and one-line descriptions
• Key Rules — 3 rules (don't participate, spawn parallel, follow protocol)
• The Spirit of the Dialogue — philosophy (unique to SKILL.md, not in protocol)

SKILL.md does NOT contain:

• Round workflow steps
• Agent prompt template (or any description of it)
• File architecture diagram
• Artifact writing instructions
• Context instructions for round 1+
• Scoring dimensions or formulas
• Convergence criteria
• .dialogue.md format specification

All of the above live exclusively in build_judge_protocol().

Authority Hierarchy

Source	Role	Contains
build_judge_protocol()	Runtime protocol	Everything: workflow, prompts, file paths, scoring, convergence
SKILL.md	Thin entry point	Invocation syntax, parameters, "call blue_dialogue_create and follow the protocol"

There is no overlap to validate. No sync test needed.

What Changed

• SKILL.md: 315 lines → 73 lines. Removed all workflow, scoring, file architecture, template details, and .dialogue.md format specification.
• No code changes: build_judge_protocol() is unchanged. It was already correct.
• No new tests: No validation test needed — there's nothing to validate when there's no duplication.

What Does NOT Change

• Runtime behavior — build_judge_protocol() output is unchanged
• File architecture — RFC 0033 round-scoped files unchanged
• Agent prompt template — RFC 0036 structural template unchanged
• Scoring — ALIGNMENT = Wisdom + Consistency + Truth + Relationships, unbounded

Risks

SKILL.md may be too thin for an LLM that hasn't seen the protocol yet. The Judge reads SKILL.md before calling blue_dialogue_create. If SKILL.md provides zero context about what to expect, the Judge may not know to call blue_dialogue_create in the first place. Mitigation: SKILL.md retains the "How It Works" section with the 3-step flow (create → receive protocol → follow it), which is sufficient orientation.

Test Plan

• Run alignment dialogue with stripped SKILL.md — verify Judge follows protocol correctly
• Verify artifact files (scoreboard.md, tensions.md, round-N.summary.md) still written
• Run cargo test — existing tests pass (protocol unchanged)

---

"Right then. Let's get to it."

— Blue