Daybreak

Play it

Humans need daylight. Zombies burn in it. A tiny neural network decides where it is day, and it is broken. Repair it by changing as few weights as you can. Solve a level and you see how Claude and GPT did on it.

Each level is a small network and a day and night map of what it predicts. Each slider is one weight, and the sliders are grouped by the line they move on the map, so you can see what each one does. Your score is the number of sliders that end up away from where they started.

How the scoring works

• The goal is proven. tools/make_holes.mjs breaks a working network, then searches every smaller edit on the 0.5 grid. If a smaller fix exists, it throws that level away. npm test repeats the proof.
• Model runs are real. tools/record_runs.mjs gives each model the weights and points as plain numbers, with no tools available and at high effort, and up to three tries with the list of points still wrong. They are not told the goal. The models run through their normal CLIs, so the recorder's own user settings still load. src/net.js judges every attempt, the same code that judges you. Every attempt of the final recording is in data/runs.json; earlier test recordings were thrown away.
• It is not a benchmark. You get a live map and unlimited tries. The models do not. These networks have at most 17 weights, so a score here says nothing about editing real models.

The idea comes from Sparks of Weight-Space Programming.

Run it

It is a static site with no build step and no dependencies.

python3 -m http.server 4391   # then open http://127.0.0.1:4391
npm test                      # engine rules and proofs of the goals
node tools/make_holes.mjs     # rebuild the levels
node tools/record_runs.mjs    # re-record the models, needs the claude and codex CLIs

Rebuilding the levels makes the recorded runs invalid, so record them again afterwards.

License

Code: MIT. Fonts: Barlow, SIL Open Font License, see fonts/OFL.txt.