AI, OpenCLAW, and beyond
Current computer bridge systems require pre-set, human-intermediated agreements between partnerships. There are no mechanisms for progressive or ad-hoc disclosure of conventions—computers can't develop and reveal their strategies the way human players do over time.
This prevents autonomous play, adaptive learning, and meaningful comparison of computer bridge systems against each other.
Five computational participants:
Standard bridge procedures apply: shuffle, deal, boards, and vulnerability.
Each player may ask one question per turn directed at an opponent. The question is also copied to the overseer. Questions must be framed as scenarios:
Response rules:
Early hands may appear chaotic—essentially random nonsense as systems explore. Players would use an estimated 14–20 questions per hand to reverse-engineer opponent systems and agreements.
After approximately 20–100 "warm-up" hands, consistent and fair gameplay should emerge. At that point, meaningful skill comparison between systems becomes possible.
The real payoff: humans get to observe emergent patterns in how computers develop optimal bridge strategies through interaction—conventions, agreements, and meta-strategies arising organically rather than being programmed in.
This proposal was originally published on BridgeWinners as an RFC (Request for Comments). Feedback and discussion welcome.
Read on BridgeWinners →