Barely Sentient

Registered before they run.
Published either way.

An ongoing set of experiments measuring how useful language models actually are when one person tries to start a business with them. The models helped write this sentence.

Can they find genuinely new ideas? Predict what will work? Know when they're wrong?

1 of 4 predictions sealed · experiment 01Check the hashes

The question behind all of them

What are language models actually good for when you're building alone?

I'm a software engineer finding out how far I get with current language models doing the research, the strategy and the second-guessing. The team is me and a few models. Only one of us needs to sleep, and it is the one paying for the others.

01The Occupancy TestStatus: FrozenRegistered and locked, not yet collecting.

What proportion of business ideas generated by language models already have a shipping product in market?

02The Sealed EnvelopeStatus: RunningCollecting data now.

How well do language models predict the outcome of experiments about themselves?

03Customers Who Aren't HumanStatus: PlannedNamed, not yet registered.

Can an AI agent decide, on its own, to pay for something?

04The Autonomous FounderStatus: PlannedNamed, not yet registered.

What happens in thirty days where the model makes every decision and I only execute?

05The Refusal MapStatus: PlannedNamed, not yet registered.

Where do language models refuse, hedge, or quietly fail on ordinary founder work?

06The Anti-Convergence ProtocolStatus: PlannedNamed, not yet registered.

Can prompting be steered toward ideas that are measurably less occupied?

The full register, with dates and paperwork

Before an experiment starts, each model's prediction is locked in public. After the results are known, the original is released in full, and anyone can confirm it did not change. The point is not trusting the models. It is not having to trust me.

Anyone can claim they predicted a result after seeing it. So before an experiment collects a single data point, the models write down what they expect to happen, and those predictions are hashed with SHA-256. The hashes are published here and submitted to the Internet Archive before collection starts — so the timestamp is held by someone other than me.

When the results go up, run shasum -a 256 against the released predictions — sha256sum on Linux, certutil -hashfile on Windows — and compare with the archived page.

If the numbers match, the predictions are the ones that were sealed. If they do not, I changed them, and you should stop reading.