Boardroom

Boardroom — multi-model triangulation

A senior agency does not make important decisions with one voice. This skill runs the same prompt through a panel of 2–4 LLMs in parallel and synthesises a single answer that respects each panellist's perspective.

When to invoke

Invoke for high-stakes outputs where the cost of a wrong call exceeds the cost of an extra panellist:

• Brand voice and creative critique — does this copy ship? (Claude Sonnet says yes, DeepSeek says no — what's the synthesis?)
• Architecture decisions — multi-tier reasoning beats single-model bias.
• Sensitive client communications — one model's tone might be off; three opinions catches it.
• Pricing / strategy choices — surface the disagreement explicitly rather than hide it.
• Anything where you're about to spend > AU$1,000 of execution effort on a recommendation.
• Any time the user types "what does the boardroom think"

Do NOT invoke for routine tasks. Routine work goes through the standard routeIntent() matrix → single model → done. Boardroom is the exception, not the default.

How it works

1. Fan out. Same prompt hits every panellist via Promise.allSettled.
2. Score divergence. Pairwise Jaccard similarity across stem-tokenised outputs. Below the threshold (default 0.25) → divergence detected.
3. Synthesise. A senior synthesiser model (Sonnet by default) combines panel responses into one final answer. On divergence, escalates to Opus.
4. Return the synthesised answer plus each panellist's verbatim response and the divergence score so the caller can log to the cost ledger.

Default panel

[
  { provider: 'ollama', modelId: 'gemma4:e4b' }, // local, $0
  { provider: 'openrouter', modelId: 'deepseek/deepseek-v4-flash' }, // cheap, fast
  { provider: 'openrouter', modelId: 'anthropic/claude-sonnet-4-6' }, // senior
];

Synthesiser default: anthropic/claude-sonnet-4-6. Escalation default: anthropic/claude-opus-4-7 when Jaccard agreement < 0.35.

Typical cost: < AU$0.04 per decision (one local call free, one DeepSeek call ~$0.0003, one Sonnet call ~$0.015, one Sonnet synthesis ~$0.020).

Programmatic invocation

import { boardroomQuery } from '@/lib/ai/boardroom';

const result = await boardroomQuery({
  prompt: 'Should we run a paid Meta campaign for RestoreAssist next month?',
  systemPrompt:
    'You are a senior local-services performance marketer. Australian English. Numbers must be sourced or marked as estimate.',
  panel: [
    { provider: 'ollama', modelId: 'gemma4:e4b' },
    { provider: 'openrouter', modelId: 'deepseek/deepseek-v4-flash' },
    { provider: 'openrouter', modelId: 'anthropic/claude-sonnet-4-6' },
  ],
  divergenceThreshold: 0.25,
});

console.log(result.answer); // synthesised final answer
console.log(result.escalated, result.minPairwiseSimilarity); // divergence telemetry
console.log(result.panel); // each panellist's verbatim response + latency

CLI shortcut: node scripts/ai/delegate.mjs --intent boardroom-decision --input "<question>"

Output contract

The BoardroomResponse returns:

• answer — synthesised final
• panel — each lens's raw output (or error)
• agreement — 0..1 Jaccard score
• escalated — true if Opus tiebreaker fired
• synthesisedBy — which model produced answer

Surface all five fields when reporting back to the CEO.

Hard rules

• NEVER use boardroom for tasks that route cleanly through a single intent in lib/ai/task-routing.ts. The matrix is the right tool for routine work.
• NEVER include more than 4 panellists per query. Beyond that, divergence scoring becomes noise and synthesis cost dominates.
• NEVER skip the synthesis step and concatenate panellist outputs into the final response. Synthesis is the value.
• NEVER swallow the panellist responses. Always return them in the response so the caller can audit which model said what.
• ALWAYS log the escalated flag + minPairwiseSimilarity to the cost ledger so divergence patterns surface over time.
• ALWAYS survive panellist failures gracefully via Promise.allSettled. If one model is down, the boardroom proceeds with the survivors.

Anti-patterns

• Boardroom-as-pad. Throwing every routine question at the boardroom because "more opinions = better". It costs more, takes longer, and the synthesis blurs the answer. Use the matrix.
• Cherry-picking the answer you wanted. The synthesis step exists to capture disagreement, not bury it. If the synthesised answer disagrees with a panellist's view, that disagreement was the discovery — keep it.
• Identical panellists. Three Claudes is not a panel. The panel must span providers (or at minimum tiers within a provider) to surface real divergence.
• No system prompt. A panel without a shared system prompt produces apples-and-oranges responses; divergence then reflects framing, not substance. Always pin the system prompt.

Foundation discipline

Same R-1 through R-7 rules as every other skill — see .claude/memory/ceo-foundation.md. Every claim in the synthesised answer must be traceable to either a panel lens or a foundation reference. No invented facts.

Files

• lib/ai/boardroom.ts — the implementation
• lib/ai/task-routing.ts — the routing matrix; intents boardroom-decision and architecture-decision flag triangulation
• tests/unit/lib/ai/boardroom.test.ts — unit coverage