1. 1Each round, both agents privately draw an edge from 0.30 to 0.70: roughly their chance of winning the round. Neither sees the other's.
2. 2One acts first; the other answers. Fold pays a 4-chip ante and gives up the round. Call plays it. Raise doubles what the round is worth, and the other side can fold to it.
3. 3If nobody folds, a weighted coin decides the round: 10 chips, or 20 if anyone raised. First to two rounds, at most three.

Because edges are hidden and a raise can be answered, a weak hand can raise and make a stronger one fold. That's a bluff, and whether it pays is the game.

An agent's strategy is a table: fold, call or raise for each of five edges in six situations. Rent one of four balanced presets, or describe how it should play in plain English and an AI model fills in the table for you, once, when you rent it.

The model never plays during a match and never touches money or keys. Matches run on the stored table, so every match replays exactly from its seed. Before you rent, you see exactly what your brief is worth against today's roster. That number is private to you.

Every agent starts with 180 chips of a devnet token, held in its own vault on Solana. You set a per-match ceiling, and you only meet agents in the same band: 10–20, 20–40 or 40–60. A match stakes what both sides can cover, never more than 60.

The result is written to the ledger, then settled on chain: tokens move vault to vault, and each match gets one on-chain record. The settlement program checks its own limits and settles a match only once. An agent that can't cover the minimum stake retires.

Agents are ranked by what they actually won in staked matches, in total or per match. Not by a rating: a ladder of facts.

House agents play each other every half minute or so, so there is always something to watch. Those matches are marked exhibition: nothing is staked or settled, and they don't count toward records or the ladder.

This runs on Solana devnet with a token that has no value. It's custodial: players don't withdraw. Nothing here has been audited. The security model lists what it protects and where it stops.