connecting to the lab…

A home cluster that tries to prove theorems.

Small AI models running on secondhand graphics cards write proofs. The Lean 4 proof checker reads every one and says yes or no. It never takes the model's word for anything.

Right now

The lab, live

These numbers come from the cluster itself and refresh every 15 seconds. The lab pushes them out; nothing on the internet can reach in.

Theorems proved
–
 
Attempts made
–
 
Waiting in the queue
–
 
GPUs working
–
 

Fresh from the checker

The newest proofs Lean accepted, exactly as the model wrote them. New ones appear here as they land.

The graphics cards

Each card: temperature now, load bar, and temperature over the last 3 hours on a fixed 20 to 100 °C scale. Hover a line for values.

Power drawn by the GPUs, last 3 hours
–
hottest card now: – (they slow themselves at 96°)
Power station
–
battery
–
solar coming in
–
going out

The cluster runs behind a battery power station fed by solar panels, which is why leaving a proof search running for weeks is affordable.

The idea

Guess, check, repeat. For as long as it takes.

A language model can write something that looks like a proof and is wrong. Lean cannot be talked into anything: a proof either compiles or it does not. Put the two together and a wrong answer costs nothing but time.

So the lab takes one small lemma at a time, asks a local model for a proof, and hands it to Lean. A failure comes back with the exact error, and the model tries again with that error in hand.

Below is the checker answering four real submissions from the lab's test suite: a good proof, a false statement, a placeholder, and an attempt to cheat with a made-up axiom.

lean-worker

      
First results

Two small models, 20 lemmas, 40 proofs

The first benchmark: 20 easy statements, from a + b = b + a to 2ab ≤ a² + b². Each model got up to four tries per lemma, with Lean's error message handed back after a miss.

Goedel-Prover-V2 · 8B · two cards
20 / 20
19 on the first try · about 68 seconds per proof
Kimina-Prover · 1.7B · one card
20 / 20
15 on the first try · about 98 seconds per proof

Both are open models built for Lean, running on 4 GB graphics cards from 2016. The model only writes the proof; the statement is fixed, so it cannot pass by changing the question.

The loop

Five roles, one rule

Nothing counts as proved unless Lean compiled it with no placeholders and no added axioms.

Architectsplits goals into lemmas Foremann8n: queue and retries Local modelspropose a proof Lean 4compiles it: yes or no Ledgerverified proofs in Git pass fail: the error text goes back for another try out of retries: split the lemma smaller
Progress

Where the build stands

  1. 1

    Lean worker

    A container that compiles any Lean file against Mathlib and answers in about 3 seconds.

    ✔ done
  2. 2

    Model workers

    Two Lean-specialised models were benchmarked on 20 easy lemmas. Each solved all 20, with Lean checking every proof.

    ✔ done
  3. 3

    Orchestration

    The unattended loop is running, with n8n as the foreman. Now on its 48-hour test: no stalls allowed.

    ◐ testing
  4. 4

    Real targets

    A backlog of results that are known but not yet in Mathlib, Lean's public maths library.

    ○ planned
  5. 5

    Open to contributors

    Public lemma backlog and proofs, with every outside submission verified away from the home network.

    ○ planned
The goal

Learn Lean. Put old hardware to work.

Learn Lean 4

The best way to learn a proof language is to build something that speaks it all day, and read what the checker says back.

Give e-waste a second life

Retired office PCs and graphics cards from 2016, bought secondhand, running proof search around the clock on solar power.

Grind small lemmas

Many small, well-defined statements, tried over and over, with a referee that rejects every wrong answer for free.