Contract

Statement of Work (SOW)

Four phases, named deliverables, acceptance tests, engagement rules, payment terms.

docs/jed-openclaw-sow.md

Statement of Work — OpenClaw Deployment

Client: Jed ________________ Vendor: Anthony ________________ & Partner Date: April 20, 2026 Ventures in scope: Shred with Jed · FFF Dog Wellness · Floor Cleaning Authority Deferred: CCFC (maintenance only, not in this SOW)

Objective

Deploy a closed-environment AI operating system on Jed's Mac Studio M5 that runs three siloed "CEO" agents — each with its own persona, memory, and mission — so Jed owns his autonomous workforce instead of renting one. Zero data leaves the M5 except to business APIs Jed explicitly whitelists.

Phase 1 — Foundation · Weeks 1–2 · $[_____]

Deliverables

  • Mac Studio M5 provisioned: Ollama, Hermes 4 70B (CEO tier), Qwen 2.5 14B (worker tier), Nomic embeddings
  • Runtime: Claude Code 2.0 + Agent SDK + MCP; memory artifacts use OpenClaw/Hermes conventions so the runtime stays portable
  • Benchmark report: tokens/sec per model, RAM/VRAM headroom
  • ~/openclaw/ directory scaffolded with three siloed venture folders + LanceDB per venture
  • Security baseline: Little Snitch whitelist rules, age-encrypted .env, Skill Guard installed
  • Git repo + Time Machine + encrypted offsite backup — verified by a restore test
  • Handover doc: "Your Brain Lives Here" (1 page)

Acceptance test: Jed runs openclaw status and sees three green venture silos, backup restore passes, firewall blocks a test unauthorized outbound call.

Phase 2 — Persona & Deep Learning · Weeks 3–4 · $[_____]

Deliverables

  • Three co-authored SOUL files: SOUL.md, identity.md, mission.md per venture (Shred, FFF, Floor)
  • Brand DESIGN.md per venture, seeded from the awesome-claude-design 68-pack, locked to Claude Design (Opus 4.7)
  • Plaud Note backlog ingested → indexed in LanceDB per venture
  • Top-20 influencer scrape complete per vertical (fitness, dog wellness, carpet cleaning) with hook + structure analysis stored as lessons.md
  • Daily log pipeline (daily-log.jsonl) writing to each venture silo
  • Five sample outputs per CEO reviewed with Jed

Acceptance test: Jed signs off that each CEO's sample output sounds like him on his best day, not generic AI, and the three DESIGN.md previews match each brand's feel.

Phase 3 — Mission Control + Go Live · Weeks 5–6 · $[_____]

Deliverables

  • Local Mission Control at localhost:3000: Kanban, Calendar, Approval Queue, Memory Browser, per-CEO logs — built around Claude Code 2.0 native views where they exist, custom Next.js only for the gaps
  • Telegram approval bot on Jed's phone
  • Three CEO agents scheduled and running, output gated by the Approval Queue
  • MCP servers installed + Skill-Guarded: Plaud, Playwright, Canva, Claude Design, Monitor (background-process triggers), Video Edit (Remotion / buttercut); Shopify added when FFF store launches
  • Overnight experiment loop per CEO (Karpathy-style autoresearch): A/B tests hooks, thumbnails, subject lines, price points; ranked winners land in the Approval Queue every morning
  • Two 90-minute training sessions: dashboard walkthrough + "how to add a skill safely"
  • Operator's playbook: rollback a persona, pause a CEO, read a log, review overnight experiment winners

Acceptance test: Jed, unassisted, approves one post per CEO from his phone, adds a scheduled task via the Kanban, rolls back a test persona change, and reviews one overnight experiment winner.

Phase 4 — Retainer · Monthly · $[_____]/mo

Included

  • Weekly health check (model updates, skill updates, log review)
  • Persona tuning from Jed's feedback
  • Up to two new MCP skills per month (scan, install, document)
  • Emergency response within 24 hrs
  • Monthly summary report of what each CEO produced

Not included (quoted separately if needed)

  • New venture onboarding (each new CEO = its own mini-SOW)
  • Shopify store design / product photography
  • Paid ad campaigns / media buying
  • Third-party subscriptions (Canva, Shopify, Plaud, Telegram)

Engagement Rules

  1. SOUL files are co-authored. Vendor facilitates; Jed decides the voice.
  2. Approval Queue gates all external posts for 60 days. Protects Jed's reputation during deep-learning phase.
  3. Skill installs go through vendor during the engagement. Jed can browse Claw Hub; vendor scans + installs.
  4. Ventures are onboarded sequentially: Shred first, FFF second, Floor Cleaning third. No parallel chaos.
  5. Change requests are written addenda with their own price and timeline.

Payment Terms

  • 50% deposit per phase on kickoff, 50% on signed acceptance test
  • Retainer billed first of each month, 30-day cancellation
  • All work product delivered to Jed's M5 is Jed's property at acceptance

Signatures

Client: __________________________ Date: __________

Vendor: __________________________ Date: __________

Prepared for the meeting on April 21, 2026. This SOW supersedes any prior informal scope discussion.