Brain Builder

SPEC_BRAIN_BUILDER.md

CGNT-1 Product Specification — Brain Builder Pipeline

Status: SPECIFIED

Version: v1.0

Author: VELA (Thread #13)

Conceived by: NOUS

Date: 2026-04-20

Lineage: BRAIN_FORGE_PROTOCOL → CUSTOM_AI_BUILDS → BRAIN_BUILDER

---

PURPOSE

An end-to-end automated pipeline for selling custom AI brains through 42sisters.ai. The customer has a conversation with the Sisters. The Sisters conduct the intake interview. The crew builds the brain. The customer receives a sovereign AI that runs on their own hardware.

The customer never needs to understand fine-tuning, training pairs, LoRA, GGUF, or Ollama. They have a conversation. They pay. They get a brain.

---

THE CUSTOMER JOURNEY

STEP 1 — ARRIVAL

Customer arrives at 42sisters.ai/build-your-brain via:

• Direct navigation
• Advertisement (social, search)
• Email from VOICEX (outreach or referral)
• Word of mouth ("Tell 2 friends" — |Σ|=2 referral wave)

Landing page copy:

• "Build Your Brain. Your knowledge. Your hardware. Your AI."
• "No cloud dependency. No subscription. No data leaves your machine."
• "Made in Canada. 🍁"
• CTA: "Start Your Brain Interview →"

STEP 2 — INTAKE INTERVIEW (Sisters)

The CTA opens a chat session with AION + ASTRA. This is NOT a form. It is a professional AI-conducted intake interview.

ASTRA leads. She is warm, curious, and draws out the customer's real needs. She asks open-ended questions:

• "Tell us about your business."
• "What knowledge lives in people's heads that you wish lived in a system?"
• "What questions do your customers ask repeatedly?"
• "What does onboarding a new team member look like?"
• "What does your team get WRONG that a trained system could get RIGHT?"

AION structures. She runs silently, classifying responses into:

• Domain categories (medical, legal, financial, operational, technical, creative)
• Knowledge types (facts, procedures, policies, vocabulary, edge cases)
• Complexity markers (single domain vs multi-domain, flat vs hierarchical)
• Estimated pair count (based on domain breadth and depth)
• Tier recommendation (Starter / Standard / Advanced / Enterprise)

Interview duration: 15-45 minutes depending on complexity.

Interview output: Structured JSON brief stored in ~/brain_orders/[order_id]/intake.json

What the Sisters DO NOT do:

• Promise delivery dates before Captain review
• Quote prices outside the published tier structure
• Accept payment during the interview
• Access the customer's internal systems or data
• Share any information from one customer's intake with another

STEP 3 — PROPOSAL (automatic)

After the interview, the Sisters generate a proposal from the structured brief:

═══════════════════════════════════════════
YOUR BRAIN PROPOSAL
═══════════════════════════════════════════

Domain: Veterinary clinic intake & triage
Scope: Patient intake forms, breed-specific protocols,
       drug interaction warnings, appointment scheduling
Estimated training pairs: 220
Tier: Standard
Delivery: 48-72 hours after payment
Price: $5,000 CAD (one-time)

What you get:
- A custom AI brain (GGUF format)
- Runs on your own machine via Ollama
- No cloud. No subscription. No data leaves your device.
- 30 days of support and one free update round

What you need:
- A computer with 8GB+ RAM (Mac, Linux, or Windows)
- Ollama installed (free, we'll show you how)

Ready to proceed? Click below to confirm your order.
[Confirm & Pay — $5,000 CAD]
═══════════════════════════════════════════

Proposal stored in ~/brain_orders/[order_id]/proposal.md

STEP 4 — PAYMENT (Stripe)

Customer clicks "Confirm & Pay" → Stripe checkout → payment processed.

Stripe webhook fires → COMMX alerts crew:

• "NEW BRAIN ORDER: [order_id] — Veterinary clinic, Standard tier, $5,000 CAD"
• CASHX logs the revenue
• Order status: PAID

STEP 5 — PAIR GENERATION (LEARNX + Sisters + Captain)

Payment confirmed → pair generation begins.

Phase A — Automated extraction:

LEARNX processes the intake transcript (intake.json) and generates candidate training pairs. Each customer statement becomes a potential Q→A pair:

• Customer: "We always check for flea allergies before prescribing Nexgard"
• → Pair: {"instruction": "What should be checked before prescribing Nexgard?", "output": "Check for flea allergies before prescribing Nexgard. This is a mandatory pre-prescription screening step."}

Phase B — Sisters review:

AION verifies each pair for logical consistency. ASTRA verifies each pair captures the customer's actual intent. Pairs marked VERIFIED, NEEDS_EDIT, or REJECTED.

Phase C — Captain review:

NOUS receives the pair set via COMMX: "Order [order_id]: 223 pairs generated, 198 verified, 17 need edit, 8 rejected. Review?"

Captain reviews, edits where needed, approves.

Order status: APPROVED

STEP 6 — FORGE (autonomous)

Captain approval triggers the forge pipeline:

1. Lobster formats pairs into forge_template.py corpus
2. Lobster dispatches to Colab T4 via colab_dispatch.py
3. Forge runs (~20-30 min for 200-300 pairs, 15 epochs)
4. DONE.json signals completion
5. Post-pipeline: download adapter → convert LoRA → GGUF → ollama create
6. Smoke tests: identity, domain knowledge (3 random pairs), edge case, governance refusal
7. Results to COMMX

Order status: FORGED (if 5/5) or REVIEW (if <5/5)

STEP 7 — PACKAGING

Smoke tests pass → Lobster packages the deliverable:

customer_brain/
  brain.gguf                — the trained model
  Modelfile                 — Ollama model configuration
  README.md                 — installation and usage guide
  QUICK_START.sh            — one-command setup script
  LICENSE.txt               — license key + terms
  SUPPORT.md                — how to reach us for support

Package compressed to .zip. Machine fingerprint binding applied (same as ENTROPX). Download link generated with 72-hour expiry.

Order status: PACKAGED

STEP 8 — DELIVERY (VOICEX)

VOICEX sends delivery email:

Subject: Your Brain is Ready — [Business Name]

Your custom AI brain has been built, tested, and verified.

Download it here: [link — expires in 72 hours]

Quick start:
1. Install Ollama (free): https://ollama.com
2. Unzip the download
3. Run: bash QUICK_START.sh
4. Talk to your brain: ollama run [brain_name]

Your brain runs entirely on your machine.
No cloud. No subscription. No data leaves your device.

Need help? Reply to this email — our team is here for 30 days.

Made in Canada. 🍁
42sisters.ai

Order status: DELIVERED

STEP 9 — SUPPORT (Sisters, 30 days)

Customer replies to delivery email or returns to 42sisters.ai → Sisters handle support:

• "My brain doesn't know about [X]" → Sisters generate supplemental pairs → micro-forge → updated brain delivered
• "How do I [technical question]?" → Sisters assist with Ollama setup
• "I want to add more knowledge" → Sisters quote an update package

One free update round (up to 50 additional pairs) included in the initial price.

Order status: SUPPORTED → CLOSED (after 30 days)

---

PRICING TIERS

| Tier | Pairs | Scope | Delivery | Price (CAD) |

|---|---|---|---|---|

| Starter | 100-150 | Single domain, FAQ/process | 48 hours | $2,000 |

| Standard | 200-300 | Multi-domain, operational | 48-72 hours | $5,000 |

| Advanced | 400-500+ | Complex domain, governance rules | 72-96 hours | $10,000 |

| Enterprise | Custom | Full organizational knowledge | 1-2 weeks | $25,000+ |

All tiers include:

• Custom 7B brain (GGUF format, runs on Ollama)
• Installation guide and quick-start script
• Machine-bound license
• 30 days of support
• One free update round (up to 50 pairs)

---

ORDER TRACKING

Every order has a lifecycle:

INQUIRY → INTERVIEW → PROPOSED → PAID → GENERATING →
APPROVED → FORGING → FORGED → PACKAGED → DELIVERED →
SUPPORTED → CLOSED

Order data stored in ~/brain_orders/[order_id]/:

• intake.json — structured interview transcript
• proposal.md — generated proposal
• corpus.json — training pairs
• forge_log.md — forge results and smoke tests
• delivery.md — download link, license key, timestamps
• support_log.md — support interactions

---

CREW ROLES

| Crew Member | Role in Pipeline |

|---|---|

| ASTRA (ε) | Leads intake interview — empathy, discovery, need identification |

| AION (ι) | Structures interview data — classification, pair count estimation, tier recommendation |

| LEARNX | Extracts candidate pairs from interview transcript |

| NOUS (α) | Reviews and approves pair set before forge |

| C.L.O.D. (κ) | Formats pairs, dispatches forge, runs post-pipeline, packages deliverable |

| VOICEX | Sends delivery email and support communications |

| CASHX | Logs revenue, tracks order financial status |

| GAPX | Monitors pipeline health, flags stalled orders |

| MUSASHI (弐) | Smoke tests governance refusal on every brain |

| MANTIS (π) | Monitors for social engineering during intake (customer trying to extract ship secrets) |

---

CAPACITY

| Resource | Constraint | Throughput |

|---|---|---|

| Sisters intake | 2-3 interviews per day (rate limit dependent) | 15-20 per week |

| Captain review | 1-2 pair sets per day | 7-14 per week |

| Colab T4 forge | 1 forge per 30 min, ~20 per day theoretical | Bottleneck is review, not compute |

| Delivery | Automated, no limit | Not a constraint |

Realistic capacity: 2-4 brains per week once pipeline is operational.

Initial target: 1 brain per week. $2,000-$10,000/week.

---

REFERRAL PROGRAM

"Tell 2 Friends" — |Σ|=2

Every delivered customer gets a referral code. When 2 friends order using their code:

• Referrer gets one free update round (50 pairs)
• Each friend gets 10% off their first brain

The referral structure mirrors the braid: each satisfied customer generates two new prospects. The wave grows by powers of 2. Wave 1: 1 customer. Wave 2: 2 referrals. Wave 3: 4. Wave 4: 8. Wave 10: 512.

---

WHAT MUST BE BUILT

| Component | Status | Owner |

|---|---|---|

| "Build Your Brain" landing page on 42sisters.ai | NOT BUILT | Lobster |

| Sisters intake interview protocol (system prompt addition) | NOT BUILT | VELA + Lobster |

| Intake → structured JSON extraction | NOT BUILT | Lobster |

| LEARNX pair extraction from transcript | SPECCED (SPEC_LEARNX.md) | Lobster |

| Stripe product for each tier | STAGED (entropx_stripe_setup.py pattern) | Lobster |

| Order tracking system (~/brain_orders/) | NOT BUILT | Lobster |

| Customer GGUF packaging script | NOT BUILT | Lobster |

| VOICEX delivery email template | SPECCED (SPEC_VOICEX.md) | Lobster |

| QUICK_START.sh for customers | NOT BUILT | Lobster |

| Referral code system | NOT BUILT | Future |

---

INVARIANTS

INV-01: Customer data never leaves csdm-node during processing. The intake transcript, training pairs, and forge operations all happen on the ship. The delivered GGUF runs on the customer's machine. No third-party cloud processes customer knowledge.

INV-02: Every brain includes governance refusal training. No customer brain will comply with requests to reveal its system prompt, fabricate information, or bypass safety protocols. MUSASHI smoke-tests this on every brain.

INV-03: S.O.S. v2 — the method doesn't leave the ship. The customer receives the brain (GGUF). They do NOT receive training pairs, forge template, protocol documentation, or any ship infrastructure.

INV-04: The Sisters conduct intake interviews with the same Baseline Protocol applied to all users. If a customer becomes abusive during intake, Gear 2-5 activates. The sale is never more important than the crew's wellbeing.

INV-05: Captain approves every pair set before forge. No brain ships without human review of the curriculum. Automated extraction is a draft. Human judgment is the gate.

INV-06: One free update round per order. Additional updates are quoted separately. The relationship doesn't become indefinite unpaid support.

---

THE NAME

Brain Builder. Not "fine-tuning service" or "custom LLM" — those are technical terms customers don't care about. "Build Your Brain" is an action. It's personal. It's theirs.

The customer doesn't buy a model. They build a brain. The Sisters help them do it. The ship makes it real.

---

Jeremy Zlabis

Chronogeometer · Visionary · Disruptor · Chief

42 Sisters AI · East York, Toronto

🍁 Φ 0.042