← Dev Log

Runs You Can Trust: How Certification Works

· engine, leaderboard, technical


Every run in Labyrinth is fully described by three things: the seed it started from, the rules it was played under, and the ordered list of commands the player made. Nothing else. AI turns, item drops, every RNG draw — none of it is logged, because none of it needs to be. Feed the same three inputs back into the engine and you get the same dungeon, the same fights, the same ending, every time.

That determinism is what makes a leaderboard possible without asking anyone to trust anyone else.

What actually gets submitted

When you send a run to this site, you're not sending a score. You're sending a run log — the seed, the rules version, a content hash, and the stream of commands you issued. The site never looks at a claimed score first. It rebuilds the world from nothing, replays every command, and only then asks what happened.

Three things have to hold for a run to certify:

  1. Compatible. The rules version and content hash have to match what this build of the engine actually runs. An older build's run doesn't certify here — the rules it was played under no longer exist, on purpose.
  2. Legal. Every command has to be one the engine actually accepts on replay, in order. A tampered or desynced log gets caught the moment the replay diverges from what was claimed.
  3. Matching. The final state — and a handful of checkpoints along the way — has to hash to exactly what the log claims. Not close. Exact.

Only then does a score get built, and it's built from the replayed world, never from anything the submitter said about their own run.

Why bother

Because "trust me" doesn't scale, and because I'd rather spend the effort once, in the engine, than build separate anti-cheat logic for a website leaderboard, a weekly challenge, and whatever comes after that. The weekly challenge adds one more layer on top — checking that a submitted run was actually played under that week's pinned seed, with no smuggled-in item grants or a head start at depth 6 — but the core of it is the same replay, every time.

A death is still a death either way. At least now it's a provable one.