Chapter 00 · Proof

Compai Playbook — Source Index

7 min read Updated 2026-05-12

By Compai · Open-source educational portfolio · Built inside a profitable brand growing 100%+ annually · Repo: usecompai.com/repo

The playbook is a journey: eight sections in the order an SME actually builds an AI operating system. Read them in order the first time. Each section ends with a Ship it gate — the minimum working state before the next section makes sense.

The system you are about to build, in one sentence: it captures work as it happens, promotes durable signals, turns them into bounded execution loops, and leaves audit receipts for every consequential action — the operating memory of a company.

The 30-day path: week 1 = Start Here + The Brain (capture running) · week 2 = The Tools (hands on 2 systems) · week 3 = The Agents (first agent, propose-only) + Governance (authority matrix, truth manifest, scoped spaces) · week 4 = Operate (one bounded closure-first pilot) + Proof (your own honest ledger).

1. START HERE — Why an AI operating system, what it looks like, and how far you can take it

Ship it — do not move on until: - [ ] You can explain the thesis in one sentence: capture the company, formalize it into a brain, execute on top. - [ ] You know which rung of the maturity ladder you are on today. - [ ] You have seen the live dashboard and know what “good” looks like.

2. THE BRAIN — Capture the company as it happens, formalize it into memory, make it retrievable

Ship it — do not move on until: - [ ] A capture spine runs with at least 3 sources, each item stamped (source, timestamp, sensitivity). - [ ] Capture keeps queueing when the model is down — you have tested it. - [ ] Domain indexes exist with named owners; promotion-queue depth is on a dashboard. - [ ] Tables over ~10k rows live in the sidecar with a dataset card, not in markdown. - [ ] Every public claim carries a date, status and evidence class; known gaps remain visible.

3. THE TOOLS — One protocol between the brain and every real system — hands, not just answers

Ship it — do not move on until: - [ ] An MCP (or equivalent) server connects the brain to at least 2 real systems. - [ ] Your jobs report four states — ok / blocked-provider / blocked-reauth / failed-validation — never bare pass/fail. - [ ] You can swap the model provider without losing any context.

4. THE AGENTS — Domain agents that execute on top of shared memory, and the factory that ships them

Ship it — do not move on until: - [ ] Your first agent runs in ONE domain, propose-only, reading its scoped index. - [ ] It writes decisions back to the brain — the second agent can see what the first learned. - [ ] You ship the second agent from the factory pattern, not from scratch.

5. GOVERNANCE — Bounded execution: scope, security, audit — safe to depend on

Ship it — do not move on until: - [ ] An authority matrix (read / propose / execute) exists for every agent. - [ ] Finance, HR and legal live in scoped spaces BEFORE any external exposure. - [ ] Runtime identity, workspace ownership and artifact storage match the architecture contract. - [ ] Every canonical skill has a builder, an independent judge, evidence and a promotion state. - [ ] Published artifacts live on a protected company route with versions and receipts. - [ ] Degraded organs open tasks in a queue someone drains — not dashboard tiles.

6. OPERATE — The loops that turn captured context into work — and make the system improve weekly

Ship it — do not move on until: - [ ] The context-to-work loop runs: signals become candidates, work objects, receipts. - [ ] One approved task has completed ten reviewed runs at ≥80% verified closure before adding generators. - [ ] The human digest is capped (≤10 items) and actually gets drained daily. - [ ] Every correction becomes a rule — the system is measurably sharper than last month.

7. BUILD IT — The executable path: setup, bootstrap, templates, downloadable artifacts

Ship it — do not move on until: - [ ] You ran the setup on your own infrastructure and adapted the templates — not copied them blind. - [ ] Your brain has its first 50 real documents and one working loop end to end.

8. PROOF — Honest ROI, production lessons, and the failure ledger

Ship it — do not move on until: - [ ] You keep your own two-layer ROI ledger with every cost counted — including all the seats. - [ ] You keep a failure ledger. If it is empty, you are not measuring.

Version 5.0 — Truth & Execution · 12 July 2026 — This release adds the evidence contract, architecture invariants, skill governance and a closure-first execution gate. It also publishes the current reference snapshot without converting pilots into production claims: 4,842 Brain documents, 373 available skills, 47 canonical skills, 97 MCP tools, seven production agent runtimes, authentication in enforce, 15 connector smoke tests green and a 42,000+ receipt action ledger. 66 chapters including this source index. The public repo includes the full playbook plus starter kit artifacts, templates and evals.