Tabletop MMORPG action mapper

Watch out

Every figure is self-reported in a post, with no methodology attached, so treat them as a developer's account rather than a benchmark. Agreement across repeated runs reached 96%, which measures self-consistency: the same input produces the same typed actions almost every time. That is stability, not correctness, and a consistently wrong action scores the same as a right one. The median latency of 317 ms is a median with no p95, no tail and no statement of what it includes; the 6.52-second previous setup is not named, so the comparison baseline is unknown. The $0.40 covers a test suite of 1,385 calls, which is not a session, and the capture gives no token counts or per-call breakdown. Not documented: the action schema, how much character context is sent, whether the action set is constrained, what happens at low confidence or on a malformed action, and whether the model can emit something the server rejects. The game is PT-BR only with a closed playtest waitlist, so this is a pre-release state. Copy reviewed against the linked page on 2026-10-04.