The Overnight Lab

An autonomous research harness running on machines in my house. It pre-registers hypotheses, runs experiments while I sleep, and an adversarial AI tries to break every result before anyone believes it.

"The experiments run while you dream."

The machine

Every experiment moves through this pipeline. The amber box is the important one: a human reads every pre-registration before anything runs.

🤖 Proposer a local AI drafts the pre-registration 📋 Pre-registration hypothesis · method pass & kill criteria fixed before any data 👤 Human gate I read every prereg — nothing runs without approval ⚙ Executor seeded, deterministic, time-budgeted run overnight 📊 Outcome the pre-declared criteria decide: supported · refuted · inconclusive 🔍 Verifier a different AI reruns it, changes the seed, and audits the code adversarially 📜 Archivist append-only ledger morning report posted here ⏳ Deferred waiting on data or a quantum-computer queue retry next night 💥 Broken crashes and timeouts go in the ledger too ⭐ Surviving finding outcome supported AND verifier pass — the only thing the lab calls a result

This is a show kitchen. The lab itself — the models, the GPUs, the ledger — runs privately on my hardware, but the recipes and the plates that come out are public. Everything posted below carries its full provenance: the hypothesis as it was registered before the experiment ran, the pre-declared criteria that decided the outcome, and the adversarial verifier's verdict.

The heart of it is honesty-by-architecture. The success and kill criteria are frozen before any data exists, so there's no moving the goalposts. A separate AI — a different model family, so they don't share blind spots — reruns every experiment and audits the code looking for the bug that would produce the result trivially. And negative results get posted with the same ceremony as positive ones, because a clean refutation is a real finding.

Your idea → the lab

What happens after you press submit.

01

You submit

A research question — something checkable by a computation. It lands in the lab's inbox as plain text. Nothing you type is ever executed.

02

I review

I read every submission personally. The interesting, tractable, safe ones move forward. This gate is a human and stays a human.

03

The lab pre-registers

The proposer AI turns the idea into a falsifiable hypothesis with pass/kill criteria — and I review that too before approving.

04

It runs overnight

Seeded, deterministic, time-budgeted. Then a second AI tries its best to break the result before it counts.

05

Results post here

Supported, refuted, or inconclusive — with your idea credited if you left a name. Refutations are celebrated.

From the lab notebook

Real experiments, verified results, full provenance. Newest first.

Loading the notebook…

Submit a research idea

Good ideas are questions a computer can answer in a night: simulations, statistics, small machine-learning experiments, cellular automata, toy quantum circuits. "Does X actually beat Y?" beats "explore X".

0/120
0/1500

Submissions are plain text with hard length limits, are never executed, and go into a review queue — a human reads everything before the lab touches it. No accounts, no tracking, no email required.