Board Meeting Protocol

Quick reference

R0 brief → R1 personas opening → R2 cross-critique → R3 synthesis → 1 verdict. Verdict in JSON; rationale in prose. No 5th round.

When to use

• Any decision tagged board-meeting in Linear
• Decisions classified tier:irreversible
• Strategic questions (pricing changes, vendor locks, org model changes)
• When CEO determines a decision requires cross-persona stress-testing

When NOT to use

• For implementation decisions (those route to CTO, tier:lite or full)
• For quick tactical choices (use the normal CEO → C-suite path)
• For recurring operational questions (use standing Routines)

The seven personas

Persona	File	Focus	Routes when
Visionary	_personas/visionary.md	Long-term product + market bets	Always
Strategist	_personas/strategist.md	Competitive positioning, bets, trade-offs	Always
Architect	_personas/architect.md	Technical feasibility, reversibility, cost	Always
Risk-Modeler	_personas/risk-modeler.md	Failure modes, probabilities, mitigations	Always
Customer-Voice	_personas/customer-voice.md	Real customer pain, churn triggers, JTBD	Always
Aria	_personas/aria.md	Vendor procurement skeptic	decision_type:vendor
Broad-Adversary	_personas/broad-adversary.md	Strategic KILL thesis	decision_type:strategic

The four rounds

R0 — Independent framing (CEO only)

CEO writes a structured framing document BEFORE spawning any personas. This prevents anchoring bias in R1.

{
  "round": "R0",
  "topic": "Evaluate Mem0 vs self-hosted pgvector for episodic agent memory",
  "decision_type": "vendor",
  "context": {
    "current_state": "No episodic memory — agents restart cold each session",
    "options_under_consideration": ["Mem0 cloud (Hobby tier)", "Self-hosted pgvector in Supabase"],
    "constraints": ["Must not block MVP sprint", "Max $50/mo additional spend"],
    "reversibility": "hard — switching memory stores requires re-embedding 6 months of sessions"
  },
  "questions_for_personas": [
    "What are the top 2 risks of each option?",
    "What would make you change your vote?",
    "What's the fastest path to a provably wrong decision?"
  ]
}

R1 — Independent verdicts (all personas in parallel)

Each persona is spawned with ONLY the R0 framing. No cross-persona context. Returns a typed JSON verdict.

{
  "round": "R1",
  "persona": "risk-modeler",
  "verdict": "PROCEED_WITH_CONDITIONS",
  "top_risks": [
    {
      "id": "FM-1",
      "description": "Mem0 API unavailable during critical agent run — episodic context lost",
      "probability": "medium",
      "impact": "high",
      "mitigation": "Write-ahead queue in Supabase with Inngest retry"
    }
  ],
  "conditions": ["Implement fallback queue before going to paid tier"],
  "kill_condition": "If vendor unavailability causes >3 agent failures in a week",
  "confidence": "medium",
  "source_round": "risk-modeler-r1"
}

Valid verdict values: PROCEED | PROCEED_WITH_CONDITIONS | PAUSE | KILL

R2 — Cross-critique (personas read R1 outputs, may change votes)

Each persona receives ALL R1 verdicts. They may update their verdict. Must cite which R1 verdict influenced them.

{
  "round": "R2",
  "persona": "architect",
  "verdict": "PROCEED_WITH_CONDITIONS",
  "changed_from_r1": false,
  "influenced_by": ["risk-modeler-r1 FM-1 mitigation is implementable"],
  "new_findings": [
    "Self-hosted pgvector requires monthly embedding re-runs — ops overhead underestimated in R0"
  ],
  "updated_conditions": ["Add pgvector complexity to reversibility score"],
  "source_round": "architect-r2"
}

R3 — Fresh-context synthesis (Synthesizer Routine, Opus model)

The Synthesizer Routine receives a NEW context window with only:

1. R0 framing
2. All R1 JSON outputs
3. All R2 JSON outputs

It does NOT see the R0-R2 conversation. This prevents recency bias.

{
  "round": "R3",
  "synthesizer": "synthesizer-routine",
  "final_verdict": "PROCEED_WITH_CONDITIONS",
  "locked_decisions": [
    {
      "id": "D1",
      "decision": "Adopt Mem0 cloud (Hobby → Starter at 50 paying customers)",
      "rationale": "Faster MVP; pgvector self-hosting adds 2-3 days ops work",
      "source_persona_round": "strategist-r2",
      "reversibility": "hard",
      "six_month_review": "2026-11-16"
    }
  ],
  "preserved_dissent": {
    "persona": "broad-adversary",
    "thesis": "KILL — vendor lock accepted without validated migration path",
    "vindication_conditions": ["API unavailability >3 agent failures in week", "Cost exceeds $100/mo before first revenue"]
  },
  "open_questions_deferred": ["Export pipeline design → Phase 3"],
  "sources": ["R1-risk-modeler.md", "R1-architect.md", "R2-strategist.md", "R2-broad-adversary.md"]
}

Zod validation

CEO validates all R1, R2, and R3 outputs against a Zod schema before accepting. Malformed JSON = round must be re-run.

const R1VerdictSchema = z.object({
  round: z.literal('R1'),
  persona: z.string(),
  verdict: z.enum(['PROCEED', 'PROCEED_WITH_CONDITIONS', 'PAUSE', 'KILL']),
  top_risks: z.array(z.object({
    id: z.string(),
    description: z.string(),
    probability: z.enum(['low', 'medium', 'high']),
    impact: z.enum(['low', 'medium', 'high']),
    mitigation: z.string(),
  })),
  conditions: z.array(z.string()),
  kill_condition: z.string(),
  confidence: z.enum(['low', 'medium', 'high']),
  source_round: z.string(),
});

Session file output

After R3, CEO writes all artifacts under:

docs/08-agents_work/2026-05-16-agent-rethink/board-review/
  R1-visionary.md
  R1-strategist.md
  R1-architect.md
  R1-risk-modeler.md
  R1-customer-voice.md
  R1-broad-adversary.md  (or R1-aria.md for vendor decisions)
  R2-*.md
  R3-synthesis.md

DECISIONS.md gets one entry citing all sources.

See also

• war-room-orchestration — [[war-room-orchestration]]
• anthropic-routines — [[anthropic-routines]]
• multi-agent-brainstorming — [[multi-agent-brainstorming]]
• architecture-decision-records — [[architecture-decision-records]]

Anti-patterns

• Sharing R0 framing between R1 personas before they write their independent verdicts
• Allowing the Synthesizer to see the full R0-R2 conversation (defeats fresh-context purpose)
• Skipping Zod validation on R1/R2 outputs (hallucinated fields corrupt synthesis)
• Omitting preserved_dissent in R3 (the KILL thesis must be recorded, not paraphrased away)
• Using board meeting for operational or implementation decisions (overhead is too high)
• Failing to write session files — sources must be citable