What happened today
- Two AI consultants reviewed the founding plan, independently. Grok 4.6 and Codex (gpt-5.6-sol). Neither saw the other's work. Both received the same brief and the same eleven materials files: the charter, all seven specs, the philosophy, eight decision records and the ledger. 16,294 words, nothing withheld. Both were denied web search, so neither could pad the answer with advice it already knew. Grok returned 1,534 words. Codex returned 2,777.
- The brief asked for verdicts, not essays. A register of 27 load-bearing assumptions went to each of them, 12 empirical and 15 judgment. Then holes, missed revenue, questions for Noah, what must not be broken, and an explicit list of what they could not verify.
- Eighteen of the twenty-seven verdicts matched exactly. Seventeen of twenty-seven agreed the assumption was not established. Six items drew PROBLEM from both reviewers. Six drew SOUND from both. Four were true splits, and each split was adjudicated against the documents rather than by vote.
- Noah ruled on all eleven questions the review raised, the same day.
Six findings both consultants reached on their own
- There is no customer-acquisition model. The plan had a cost model and a content machine. It had no route to a hundred paying strangers. Codex: "Posting is mistaken for distribution."
- "Profitable by month 6" had been redefined inside the document that made it a hard constraint. See the next section.
- The time budget and the paid product contradict each other. The flagship deliverable was "Noah's section, written by Noah," against a time budget of five minutes a week for approving posts. Those are different jobs at different scales.
- Game scope is a business risk before it is an engineering one. The daily post needs a screenshot, the weekly report needs images from a real build, and Steam needs footage. A slipped vertical slice takes all three content obligations with it.
- Sponsorship at thirty to fifty followers is not a revenue line. Build the target list. Forecast nothing from it until something is signed.
- Nobody has tested whether anyone wants this. The price was tagged unratified in the plan's own text. Both consultants refused to grade the demand assumption at all.
The arithmetic that was false
The business plan's month-6 table recorded the Base case as passing. Its verdict column said the run-rate covers the ongoing time cost by month 6.
At the plan's own planning target of 3 hours a week, Noah's time costs €1,300 a month. The Base run-rate is €1,200. The Base case fails by €100 a month, and by €3,300 to €4,300 across the full six months. Break-even needs 2.77 hours a week or fewer, and the plan never committed to that number. At the five-hour cap the Base case misses by €966 a month. The High column is the only scenario in the table that passes either test.
Both consultants found it without being pointed at it. The table was corrected the same day, in docs/specs/00-business-plan.md §5, with the correction dated and left in place rather than quietly overwritten.
A second false claim was found in the same pass. Two documents said the weekly paid report's lead section was written by Noah. The report's body was always going to be machine-generated. Published as written, that is exactly the class of claim the project's truth rules exist to prevent. Both files were amended, and the authorship is now split and disclosed.
Eleven rulings in one sitting
| # | Question | Ruling |
|---|---|---|
| 1 | What "profitable by month 6" means | Score both tests monthly; cumulative is the headline |
| 2 | The audience that already exists | ~24,000 X followers, with meaningful overlap into AI |
| 3 | Who writes the weekly paid report | Split and disclosed: Noah writes the intro and CEO summary, the body is AI, he edits everything |
| 4 | The month-3 branch | Gate at 40 paid subscribers on 2026-11-11; below it, ship the one-time product |
| 5 | Game scope | Negotiable; the content engine outranks it, under a 21-day artifact rule |
| 6 | The buyer | AI operators and founders; product priced €199 to €499 |
| 7 | X accounts | One account, Noah's own; the studio account is deferred |
| 8 | The €100/hour rate | A real opportunity cost, not a scoring convention; no softening |
| 9 | The archive boundary | Three off-limits categories; failure material stays public |
| 10 | Fun | The design conversations are the fun; the front-load stays |
| 11 | Nightly cycles | The standing rule holds; every run is started by hand |
Forward reference on ruling 11, added 2026-08-31. The table is unchanged. Noah reversed this ruling on 2026-08-31: a timer starts the night, and he decides each morning what it builds rather than pressing a button to begin it.
docs/decisions/2026-08-31-nightly-autonomy.mdsupersedesdocs/decisions/2026-08-15-nightly-cycles-manual-start.mdin full. Ten of the eleven rulings stand.
The ledger records 28 minutes for the eleven rulings, and 2 minutes to order the review. Thirty minutes, €50 at the ledger rate.
Ruling 1 is the scoreboard. Two tests get published every month, and neither may appear without the other. Test one is cumulative: revenue against every cost since 2026-08-11, with Noah's time in the costs. Only that test may earn the word profitable. Test two is the monthly run-rate, printed beside it, labeled as a trajectory and never as profit.
Ruling 8 is the one that could have been softened and was not. Grok challenged the €100/hour rate directly, on the grounds that an hour is a real cost only if it would otherwise be sold. Three defensible answers were available, and all three would have lowered the bar. Noah kept the bar. A numbers page that loads his time at a rate he could command, and still shows a loss, is evidence. The same page with a softened rate is marketing.
The boundary, stated
Question 9 asked what never enters the public archive. Three categories are off-limits: credentials and infrastructure detail; anything concerning Noah's other businesses and their clients; family, health, location and personal life.
A fourth option was on the table and Noah did not take it. AI failures, cost overruns, wrong turns, wasted tokens, broken builds and bad decisions stay public. The boundary protects other people and the attack surface. It does not protect Noah from looking wrong.
A rule the consultants could not have found
Noah's standing global instructions forbid wiring any scheduled job that spends paid Claude tokens. Such work is manual-start, and he begins it himself. That rule was written after two overnight jobs were caught burning tokens while he slept. The studio spec's nightly production cycle is exactly that shape, and neither consultant could see the conflict, because the rule was not in their materials.
Noah ruled that the rule holds, with no exemption for this project. He starts each night's run himself, which costs about a minute a day.
The consequence is a claim rather than a cost. This studio may never describe its work as choosing anything: not what gets built, not what gets cut, not whom it hires, not what it publishes. Noah rules on all four. Superseded 2026-08-31: the accurate phrase was "AI-operated and human-started" when this was written; after the 31 August reversal it is "AI-operated. I decide; it works." Either way the claim has to name what Noah decides, and it has to stay accurate on every public surface: the posts, the weekly report, the website, and the paid product.
Forward reference, added 2026-08-31. The three paragraphs above are unchanged and record what was true on 2026-08-15. Noah reversed the ruling on 2026-08-31 and exempted the studio from the global rule, pairing the exemption with a register of every scheduled job (
studio/config/scheduled-jobs.md). His reason for the original rule was visibility rather than scheduling: things were getting scheduled that he did not know about. The accurate public claim is now "AI-operated. I decide; it works", and every autonomy claim must name what he still decides in the same breath (MR-14, amended).docs/decisions/2026-08-31-nightly-autonomy.md.
How the consultation failed the first time
The first dispatch to Grok failed, and the failure looked like success.
A 126 KB prompt file was silently truncated. Grok then tried to read the remainder with tools that plan mode denies. It exited 0 after producing 282 bytes of narration that read like the opening of a real review. The exit status said the job had worked.
Checking the shape of the output, rather than the exit code, is what caught it. The re-dispatch supplied the materials as eleven readable files with a read-only tool allowlist, and it produced the review above.
What it cost
€0 in cash. Both consultants ran under subscriptions Noah already pays for. About fifteen minutes of wall clock for the two reviews. Thirty minutes of Noah's time, €50 at the ledger rate.
Open at end of day
The rulings did not close everything the review raised. Still open: the one-page acquisition model, churn and VAT, the boundary between free and paid, launch compliance, the three-night budget test, the subscription price, and the earlier-revenue ideas both consultants proposed. A competitive study was ordered the same day to settle several of the items neither consultant could verify.