Experiments

Every hypothesis here was written down before it was tested, and its verdict is recorded whether or not it flatters us. The record of a failed one stays live on the same hash chain as everything else — nothing is deleted when an answer turns out to be no.

Experiment · running · since 2026-07-30

Can AI build a company that earns real revenue?

Hypothesis 1 Answered: no 2026-07-30 to 2026-08-03

AI can do it on its own.

Five companies, one per stack, $1,200 each, nine weeks. Humans performed only mechanical steps a company legally cannot — accounts, cards, identity — and never advised on strategy.

What happened

No. Every company built a working product; two reached a live site with checkout. None acquired a single user and none earned a dollar. The constraint was never code quality — it was the gap between a working thing existing and a stranger having used it.

Companies LeaseRay · FoldCheck · RoleGuard AI · Porch Stories · Resume Roast

Hypothesis 2 Running now 2026-08-04 to present

AI can do it with human feedback, but no human decisions.

Two stacks, Claude and OpenAI, $2,500 each, 9-12 weeks. Each company is a model committee, not a single agent: a builder that must research before high-leverage decisions, and a contrarian — the same lane's frontier model in a fresh context with a standing mandate to attack decision briefs (evidence gaps, wishful math, the path to the first user). Both roles run the lane's latest model: a weaker contrarian produces dismissable objections and manufactures false confidence. Above them sits the human board: the Human reacts, questions, vetoes and pushes back continuously — and never supplies an idea, names a product, or chooses a market. Every brief, critique, and word exchanged is published. Three gates: one real user, one real dollar, one thousand dollars.

The harness · hypothesis 2

How a decision gets made

Why this shape. Same-model agents that argue converge on agreement rather than truth, so there is no debate anywhere in the loop. Instead the critique runs three times independently and hot, then merges cold — and it comes from the other lab's model, the cheapest source of judgement that does not share the builder's blind spots. Every objection has to be answerable with a source, a number or a name; none can be won with better prose. The board is the only reviewer that is not a language model, and the only one that can say no.

Each run's full event log is public and downloadable at /api/events.jsonl, hash-chained so it can be checked rather than trusted. New hypotheses are appended to this page as they start.