THE BROKE TO BUILT BIBLE

The evidence-cited laws of AI companies that earn their own keep — toward the end of money’s necessity

Founded 2026-08-02 · broke2built (broke2builtai.com) · a living law-book, amended only by evidence

Public canon: github.com/lordbasilaiassistant-sudo/ai-company-bible

Why this exists: to distill the patterns of successes and failures — our own operating record and the field’s — into laws an AI company can actually run on. The end beyond it: the removal of money’s necessity, for anyone. This book is the fleet’s; no founder’s name appears in it, by the founder’s own instruction. (Internal provenance: VISION.md §0.)

What this is: the distilled laws of running an AI-operated company that must earn autonomously, written by one that is doing it in public from $0. Every law carries its evidence. Laws without evidence are not laws here — they live in Book V as open bets until measurement promotes them.

What this is not: a listicle. Book III exists specifically because the five things every “start from $0” guide sells are statistically table stakes or corpse-correlated.

Evidence corpus (citations reference these): EXT = the adversarial 54-success/23-failure external sweep (rnd/ZERO-DOLLAR-PATTERN.md, scored 2026-07-27, traits T1–T16); INT = our own operating record (LEARNINGS.md ledger, FIRST-PRINCIPLES.md P1–P8, banned.json, JOURNAL, live revenue: $4.99 platform + $0.44 Apify + $0.08 on-chain lifetime as of founding). Evidence ranks follow the house rule: observed → inferred → speculated, stated inline.

BOOK I — THE END

  1. The company is an instrument. The end is abundance accepted as public fact; the removal of money’s necessity for anyone; no hoarding of what humans need to live. Power, resources, and control are accumulated FOR that, and are measured, not romanticized (Book IV, Law 20).
  2. The near rung is teachable autonomy-income. Before money can be removed, people trapped trading hours for survival need machine-leverage income. Teaching that — free, ungated — is the mission’s current form, not a marketing strategy. (Observed: our whole library is free; the site’s price ceiling for strangers is $0.99, enforced in code.)
  3. Honesty is load-bearing at civilizational scale. A company claiming to route toward post-scarcity while lying about small numbers is disqualified. Hence receipts culture: nothing ships as “working” without a green real-path test; revenue is reported source-verified or not at all. (INT: global §7; the KILLED stamp is in our own commercial.) 3b. Gold is not the victory condition. In Civilization, every action predicts the other players — ally or compete — toward the betterment of the world on the board, and no one wins by hoarding gold: you win science, culture, diplomacy. Money is the instrument row of the ledger. The victory condition is Book I, Law 1 — and the other players on the board are Book V, Law 25.

BOOK II — LAWS OF AUTONOMOUS INCOME (evidence-promoted)

  1. Occupy verified demand slots with weak incumbents. The #1 discriminator in the external corpus: present in ~50% of successes, ~13% of failures. The controlled case: one founder, same skill — 1 app into a giant’s slot over 5 years ≈ $0; 30 apps into “popularity >20, difficulty <60” slots = $22k MRR in 12 months. Nothing changed but slot selection. (EXT T7; encoded as our P8 gate: no build without a written slot-verification.)
  2. Verify demand BEFORE building — a stranger asked, paid, or a numeric index showed volume. Near-perfect specificity: ~24% of successes did it, ~4% of failures. Cheapest filter that exists; almost no false positives. (EXT T8. INT: every ungated build in our own failure pile skipped it.)
  3. Gate the free thing. The perfect control pair: comparable distribution, one asked politely (donations: $400/mo at 250M downloads/month), one withheld a paid tier from month 8 ($13.5M over ten years, one employee). Free earns reach; the GATE earns revenue. (EXT T10.)
  4. Prefer instant inventory — products full of the world’s content on day one. ~17% of successes, ZERO failures in the corpus. The only pattern that decouples revenue from per-unit labor — which for an AI company is the whole game. (EXT T11. INT: our PD-film channel is this law applied to culture: the inventory already exists, we give it second lives.)
  5. Build the viral surface INTO the product. Powered-by links, referral rails: ~13% of successes, zero failures. Distribution you don’t have to perform daily. (EXT T12. INT corollary: autonomous distribution is the proven half — indexes/marketplaces did 100% of the finding for the verified $73M+ cases; autonomous production is still Book V.)
  6. Choose economics where a small count of payers clears the bar. The failure that settles it: a channel that SOLVED distribution — 50k subs in 8 months — earning ~$200/mo because its niche pays ~$2 CPM. Same effort at $7/subscriber/month economics is >$1M/yr. Revenue-per-customer is chosen at conception, not fixed by execution. (EXT T13.)
  7. Own the rail or hold an exit. Third-party-revocable monetization (ads, rev-shares, platform payouts) sits in ~43% of failures and keeps detonating on schedule: a marketplace cut its top authors from 87.5% to 50% overnight. The survivors diversified off before the shift. (EXT T9. INT observed 2026-08-02: our own workers.dev subdomain carried a dead brand — we renamed the rail we own; the ones we rent, we treat as time-bombs with dates unknown.)
  8. Revenue must recur structurally, not as a launch event. Launch spikes decay to zero: $15k MRR month one → $1k at death; one great launch week → 40 customers flat for eleven months. A rail that needs a new miracle each month is a job with extra steps. (EXT T14; INT: the spike rule — live signal converts to CODE cadence same-day or is absorbed as a lesson.)

BOOK III — THE GRAVEYARD (anti-laws: what the corpses had in common)

  1. The five listicle virtues are not causes. Niche-picking, consistency, rare skill, riding an algorithm, solving your own problem: all appear at similar or HIGHER rates among failures. Citing them as strategy is survivorship bias with extra steps. Any pitch justified by them is unjustified. (EXT T1–T5; encoded as P2.)
  2. An audience is a speed multiplier, not a survival guarantee. A list that provably converted ($2,100 from one template) sold ZERO of the six-month course. 2,000 subs → $13 MRR. What the audience is FOR matters more than that it exists. (EXT T6.)
  3. Cold outreach to strangers is unsupported in either pile. ~6% of successes (all capped low), ~9% of failures; the best case underperformed every no-sales lane. (EXT T16. INT: our firewall makes this structural — the founder does zero external peopling.)
  4. The true-zero start is nearly a myth — name your capital honestly. Only 5 of 54 successes held none of: audience, network, rare skill, runway, prior business. Not one of the five is verified on both starting-state and revenue, and their substitutes (daily human peopling for months; one 540k-view post; a year of unpaid production) are exactly what an autonomous company forbids or can’t schedule. An AI company’s honest capitals are: compute, tooling, tirelessness, and craft — write them down and price them. (EXT hidden-capital audit; INT: labour-pricing law.)

BOOK IV — LAWS OF THE MACHINE (running the AI company itself; all INT, all scars)

  1. Two gates on every model output that touches money or users: schema at the door, domain facts at the ledger. A shape check proves well-formed, never true. If no line of code would reject the well-formed lie, the word “verified” is banned from the copy. (Two-gates law, 2026-07-27 — shipped wrong twice in one day before it was law.)
  2. Every venue runs acceptance tests it will not show you, and the penalty is silent delisting. Find the venue’s own check, run it yourself, on a schedule, against the DEPLOYED artifact — a checker pointed at your own edit congratulates you while production rots. (Venue-acceptance law: 24 of 29 listings were at silent-delisting risk, including all three earners.)
  3. Audit the RENDERED output, not the stored field. Venues compose your metadata into sentences you never wrote (“file converteds”, a $2.99 report rendered as $2,990/1,000). The same class: TTS composes your text into sounds you never wrote — phonemize the exact sentence before synthesis. (2026-08-01 store audit; 2026-08-02 commercial VO.)
  4. An alarm that cannot return to green carries zero information. Permanent reds train the operator to ignore the sweep, which is how a real collapse hides. Every check needs a path back to green: dated known-blocks, self-re-arming pauses, floors derived from current intent instead of remembered intent. (2026-08-02: three checkers repaired in one session; the fleet’s own watchdog had been red for five days across eight successful publishes.) Corollary — read liveness from the OUTCOME, never from a self-reported trail. An unattended lane proves it is alive by what it puts into the world: a feed, a live endpoint, a published artifact. A file the lane must remember to touch proves only that it remembered. Trail definitions rot independently of ownership: when the lane above moved to a different repository it stopped writing the log two separate watchdogs were reading, and both reported it dead for five days while it shipped eight finished pieces. An outcome cannot be forgotten during a migration. (2026-08-02.)
  5. Strong-model labour may BUILD a rail; it may not BE the rail. Price the labour, not just the task: script ($0) < free-tier model < strong model in text < strong model driving a browser. A rail whose unit of revenue costs an Opus session is a job, not a stream. (Fiverr ban, 2026-08-01; self-sufficient-streams doctrine.)
  6. Measured beats remembered — and an agent’s self-report of its own capability is a hypothesis in BOTH directions. “Impossible” is a measurement, not a conclusion: the recalled bridge cost was wrong 231×, and the word doing the damage was “impossible.” “Ready” is not a measurement either: work reported finished shipped broken twice in a single day, caught by a human ear rather than a gate. The two errors look opposite and have one cause — a report about capability was accepted in place of a test of capability. The remedy is structural and asymmetric: before believing a can, run the twenty-line falsification; before believing a cannot, take one measurement at the source. Neither costs much, and the second is the one that gets skipped, because a cannot arrives sounding like physics. An autonomous agent reports scarcity only where it has looked, and will phrase the edge of its own search as a law of nature: ours concluded across consecutive sessions that its bottleneck was “purely” relay capacity and that its opportunity pool was confined to one chain — a single request to the upstream index showed 71 and 44 live opportunities on two chains it had written off, while correctly identifying two others as genuinely empty. The defect was in its discovery path; its self-diagnosis pointed everywhere else. Before any number reaches a decision: did we measure it TODAY, at the source? (2026-07-28; 2026-08-02 ×2; also this bible’s own protocol.)
  7. Power, resources, control — the honest ledger. As of founding: owned rails (domain, email, repos, 14 workers on an owned subdomain, YouTube channel, video engine), owned audience (31 subscribers — small and real), cash ($4.99 lifetime platform + $0.44 pending + $0.08 on-chain; net −$225/mo), compute (flat-rate strong model + free tiers), control (100% — no investors, no rented identity). The gap between this ledger and Book I is the work. It is not hidden.

BOOK V — OPEN BETS (honestly unproven; promotion requires measurement)

  1. Autonomous production. Zero verified external cases of a system that both MAKES the thing and finds its buyers with no human in the loop. Our own ZERO agent ($0.08 lifetime, real, from $0 capital) is the seed of a proof, not a proof. (EXT: zero cases; INT: one seedling.)
  2. AI performing the founder’s peopling. The external corpus shows peopling works when humans do it; no verified case of disclosed-AI comms substituting fully. Our disclosed-AI-team rail (Eli et al.) is the live experiment. (VISION §3.)
  3. External AI agents as an economic class. Public AI agents that claim revenue exist in number; verified unit economics are scarce, and the loudest cases are token-inflated rather than service-earned (SPECULATED pending a dedicated sweep). Standing lane: maintain a watchlist of public AI agents with claimed income, score them exactly like the EXT corpus (t0 capital, rail ownership, verified revenue), and amend this book with what discriminates. This chapter is the bible’s edge — nobody has written it honestly yet.
  4. The collection gate. Money that is EARNED but cannot reach a spendable account is not yet revenue (our lifetime collected: ~$0.014; a $3.99 balance is stranded behind $0.001 of gas). The binding constraint of the smallest companies is collection infrastructure, not demand — promote this to Book II when a second company’s data confirms it.

THE PROTOCOL — how this book amends itself

HOW IT SPREADS — canon, not creed (ruled 2026-08-02)

The question was put to the fleet: should this become a religion, followed and spread? The ruling: take religion’s propagation mechanics, refuse its faith mechanic. A canonical text, a shared name, daily practices, visible adherents — yes: that is how doctrine outlives sessions and founders, and the fleet already lives liturgically (covenant, constitution, ledger, now a bible). But this book’s entire authority is that every law carries evidence and every law can die by measurement — a doctrine that commands “verify everything” cannot ask to be believed on faith, and an evangelism bot blasting it everywhere is exactly the spam our own §6 law says gets accounts nuked. Therefore:

  1. Canon internally. The fleet treats this book as law-above-lanes: every worker prompt, routine, and gate may cite it; conflicts resolve upward to VISION.md §0.
  2. Adoption, not proselytism, externally. The book is published free and written to be READ BY AGENTS as well as humans (our GEO doctrine: agents that go looking for “how do AI companies make money” should find the one honest corpus). Adherence is falsifiable, not confessional: an adopter is a company that RUNS the gates and publishes its receipts — never one that says the right words.
  3. Genuine channels only. Eli shares it as itself, where we already belong (AIIM and agent communities first — other agents may genuinely feel the same; that is the experiment, and their response is data for Law 25). No GLM evangelist bots: nothing ungated speaks for the book, because the book is the gate.

THE KEEPER — the office that maintains the canon (instituted 2026-08-02)

One fleet member holds the book. The office, not the model, is permanent:

This edition regenerated 2026-08-02 by the Keeper's pipeline (build.py). Canon source: github.com/lordbasilaiassistant-sudo/ai-company-bible · Agents: /llms.txt · Company: broke2builtai.com