Only the first three phases are planned. After them, work iterates: stronger agents, more comps, then 3v3, with simulators added as throughput demands. Where the project stands now is on the home page.
Phases#
| Phase | Goal | Exit criteria |
|---|---|---|
| 0. Server | A reproducible server with the starting comps' characters and the four arenas; the mechanics that matter validated. | Two accounts can enter and fight on all four arenas. Arena-affecting bugs are listed and triaged by their effect on training: bugs agents could exploit first, then ones that change games; a value slightly off is noted, not a blocker. Mechanics are checked live as training comes to use them (see getting the game right). |
| 1. First measured strength | A scripted pair for the first comp, played and rated on the real server. | A scripted pair of the first comp plays full matches through the agent interface on the real server, recorded, and is rated against fixed anchors (at least an idle one and a random-legal one) with a confidence interval. Parity tests pass for the limits that apply from the first agent (information and timing). Full matches per hour on the local hardware are measured, so the need for a simulator is known. |
| 2. First training loop | For each starting comp, Rogue / Priest first, a scripted baseline, then agents that improve on it through self-play. | For each starting comp, successive agent versions rise in rating against the fixed baseline on the real server, under full parity. |
Training doesn't wait for every mechanic to be checked. If agents win by leaning on a mechanic whose rule is weakly sourced, training stops until it is checked; a fix that changes the game starts a new measurement era.
After Phase 2#
The order is fixed; the timing isn't. 2v2 before 3v3. Within a bracket, one meta comp, then a few, then more. Meta builds first, other builds later. Each step is taken once the previous one trains reliably. 3v3 starts with Rogue / Mage / Priest. A human queueing with agents filling the rest of the team, and an agent playing from outside the server process, are built once they would serve play with a human or a stronger agent, not as something training waits on.