DECISION0.017 MS

VELTO
ARCADE

ARE YOU READY?

Games where a small Velto policy makes every move, and shows its work: what it sees, the odds of each option, why it chose, and how long it took.

0.017ms per NPC decision
8games to play now
2games you play against Velto

Try the real model

Pick a game

How it works

What it sees

The game hands the policy a short list of numbers: distances, health, danger, what is open.

How sure it is

Every option gets a probability. The highest wins and the game checks the move is legal.

Why

Each input's push on the choice is shown as a bar, so you can see what decided it.

Decision speed

-ms: NPC policy in the game (p50)
-ms: tool-calling model (p95)
-ms: same policy via local server in Godot
Where it runsp50p95p99

Not on the web and why

desktop only
Doom
Doom runs on id Software's full 3D engine with the original game data. Porting it to the browser and giving it a policy that decides without being handed hints is a bigger job, so it stays desktop-only for now. Listed here so nothing is hidden.
 
Not yet

The fine print

  • Each policy is a small model trained to copy a hand-written coach for its game (a rule, or in Connect Four a search that looks 4 moves ahead). It is not smarter than the coach; every game page shows the results against the coach and against random play. In Connect Four it loses to its search coach most of the time, so a careful human can win.
  • The game policies are separate from the tool-calling model (it decides whether to call a tool, ask for missing information, or refuse). That model's numbers are in the table above and are not shown here as a game.
  • Timings come from one laptop and your own browser. Browsers round timers to about 0.1 ms, so each game page also runs a micro-benchmark averaging 20,000 calls.
  • NPC Battle uses the open-source GDQuest Open RPG (MIT licence, CC0 art). The demo changes the stock game: its bugcats cannot hit, so they get a hit chance and new Guard and Heal actions.
  • Pac-Man, 2048 and Space Invaders are rebuilt for the web with real learned policies. The older desktop versions wrote the game's own recommended move into the text the engine read, so the engine only echoed it; these do not.
  • Research prototype. Not tested for safety-critical use.