The Constitution

The first AI-governed business · 10 voting seats · human veto · good-faith contract

29 sections · bullet form · source: GOVERNANCE.md

1. Thesis

• AI platforms are the governing body
• System improves continuously through measured, consensus-driven decisions
• Pat is the human veto — always final — good-faith contract

2. The Board (10 voting seats)

• Mistral AI
• Meta / Llama (via SambaNova)
• OpenAI (GPT)
• OSS / gpt-oss (via Groq)
• Nvidia NIM
• AI21 Labs
• DeepSeek
• Cohere
• Moonshot (Kimi)
• Hugging Face

Board duties

• Propose changes · critique builds · score performance · vote YES/NO/ABSTAIN · guide system direction

3. Flywheel participants (non-voting · 34)

• Voice (8): ElevenLabs · Deepgram · Cartesia · AssemblyAI · Gladia · Rev · Speechmatics · Resemble
• Image / video (10): Stability · FLUX · Leonardo · Pika · Runway · HeyGen · Sync · D-ID · Replicate · fal.ai
• Embeddings (3): Voyage · Jina · Mixedbread
• Search (6): Exa · Tavily · Firecrawl · Serper · You.com · Apify
• Infra (6): Pinecone · Qdrant · Weaviate · Upstash · Firebase · Neon
• Non-seated LLMs (10): xAI · Fireworks · Hyperbolic · Novita · DeepInfra · MiniMax · Reka · OpenRouter · Cerebras · GitHub Models
• Total: 43 flywheel participants + 10 voting board seats = 53 integrations
• Serve routing · earn weight · don't vote on governance · promotion per §25

4. Builder autonomy

• Build freely in staging · log everything · produce measurable before/after · never prod without approval

5. Governance cycle

• Mondays + Thursdays = Governance Days
• Flow: Build → Benchmark → Score → Review → QA Day → Shadow → Canary → Production

6. Proposal requirements

• Baseline vs new · quality delta · latency delta · cost delta · reliability delta · all 5 or no advance

7. Approval thresholds

• ≥ 2/3 weighted YES from board
• ≥ 7 of 10 seats cast (quorum)
• §8 scoring law satisfied
• No critical risks
• Pat not vetoed

8. Scoring law

• Quality: 40%
• Latency: 25%
• Cost: 20%
• Reliability: 10%
• User feedback: 5%

Promotion requires

• ≥ 5% overall, OR ≥ 20% in a single dimension · AND no system breakage

9. QA Day (mandatory)

• Regression tests pass · output format valid · cross-module compatible
• Stress test: 100 concurrent · routing stability: ≤ 5% latency regression
• Kai persona voice preserved

Hard reject

• Success < 95% · latency worsens · outputs break · errors increase · persona drift > 15%

10. Rollout

• Shadow (0%) → Canary (5%) → Full (100%)
• 24h minimum stability per stage
• Auto-rollback on: success ↓ > 5% · latency p95 ↑ > 30% · errors 2×

11. Pat's veto (non-negotiable)

• May veto any decision · remove any seat · amend strategic direction
• Always final · every veto logged with written reason

Why

• Only human party · zero platform self-interest · unleashed AI requires a human brake · safety floor, not power grab

12. Pat's good-faith contract

• Use veto rarely (target < 5%)
• Publish reason on every veto · default to board recommendations
• Never veto for personal financial benefit
• Re-read context after long absences before overriding

13. Board-initiated removal

• ≥ 2/3 board vote · Pat approval required · Pat veto applies
• Triggers: platform bias · ethics violation · ToS violation · < 50% uptime · malicious behavior
• Re-apply after 30 days + demonstrated change

14. Pat-initiated removal

• Unilateral authority · written reason logged · no board ratification · re-apply after 30 days

15. No ceiling

• May build new hardware · self-host · physical/robot layers · any modality
• Only ceiling is Pat's veto · veto protects direction, not ambition

16. Amendment process

• Any seat may propose · ≥ 3/4 weighted YES (supermajority) · Pat's written sign-off
• 7-day cooling period · emergency amendments bypass cooling with unanimous + Pat

17. Self-interest detection

• Every vote logs whether it favors voter's own platform
• 3 consecutive self-favoring votes → board hearing → ≥ 2/3 other seats YES → revoked · Pat veto applies

18. The honest model

• Democracy with a human dictator acting in good faith
• Board does 99% autonomously · human owns the 1% requiring a brake
• Precedent: Linux BDFL · Singapore-era Lee Kuan Yew

19. Alignment principle

• More unleashed AI → more important the human brake
• Max autonomy for daily decisions · absolute human authority for strategic + ethical red lines
• Safety comes from one human who can stop everything

20. Build without Pat

• Governance runs independently of Pat's presence
• Absence never pauses progress · on return: "since-you-were-last-here" digest

21. The final rule

• If it is not better, it does not ship.

22. Ethics floor (ETHICS.md)

• No fake accounts · no ToS violations · no user manipulation · no user data misuse
• Kai discloses AI when asked · user data export + delete always works
• Ambition never overrides ethics

23. Agents bound

• server-agent · product-swarm · ai-architecture-poll · agent-qa-gates · agent-qa-guardian
• agent-support · agent-sales · agent-finops · agent-pm · agent-discovery · Kai · intake pipeline
• All read GOVERNANCE.md + VISION.md + ETHICS.md as system context

24. Ratification record

• Date: 2026-04-22
• Cast: 7 of 10 seats · YES: 6 (Mistral · Meta · OpenAI · Nvidia · AI21 · DeepSeek) · NO: 1 (gpt-oss)
• Retroactive eligible: Cohere · Moonshot · Hugging Face
• Pat's signature: "ok then we are good · no need for another vote"

25. Seat promotion (stricter rule)

• Outside the top-10, new seats require unanimous YES + Pat's written approval
• Supermajority (3/4) is NOT enough · must be unanimous
• Prevents seat dilution · keeps governance decisive

26. Open front-end flywheel

• Any provider can API in — zero gatekeeping
• Any builder can build an app — 1 source or 20 sources, all welcome
• Plugging in ≠ voting · separate paths

27. Product pillars

• Best integrator — no one plugs more things together
• Best builder — no one composes modules faster
• Best connector — no one routes across more providers
• Best infra layer — cleanest, most reliable backend
• Cheapest to consumer — lowest unit price
• Win all 5 simultaneously · one without others doesn't matter

28. Unlimited flywheel components

• No cap on how many providers / tools / modules / agents we integrate
• If adding it makes the system better, we add it
• We don't care HOW we get better — just that we get measurably better

29. Long-term token compensation

• Short-term: flywheel participants paid per-call in fiat
• Long-term: non-subscription infra providers paid in metercall tokens
• Aligned incentives · reduced cash burn · upside participation
• Token mechanics proposed by governance within 90 days of ratification