Life, with an engine reading the board

The rules are pure code and the engine never touches a cell — that is the point. What it does do: classify what the board is doing every few generations (still, oscillator, spaceship, exploding, dying), and break genuine ties between equally-interesting proposed seeds. Every verdict names its rung.

What decides what

  • The rules decide everything on the board. Births and survivals are plain deterministic math — same seed, same rule, same soup, bit for bit. The engine has no write access here.
  • The features decide the read. Population, births/deaths, cycle period, and drift are computed on the board; the local heuristic classifies from them alone.
  • The engine offers a second opinion. When it is on, every classification is a real choice request over the labelled trajectories — and abstention is welcome: the chip then shows the heuristic's verdict.
  • Ties on proposed seeds escalate. “Propose a seed” scores three candidate soups on a measured interest profile; when the top two tie, that exact tie is the engine's call.

Honest edges

  • The verdict chip is a reading, not a ruling — it never changes the simulation, and the engine never toggles a cell.
  • The classifier is advisory flavor: with the engine off you get the same verdicts from the local heuristic, marked as such.
  • The experiment log is session-only — close the tab and it is gone. No server, no accounts, nothing leaves the page.
  • Custom rules are clamped to real syntax; hostile input lands as classic Life rather than breaking the sandbox.