Standard uses a learned bidder and a compact card-play network trained from public-information Trickster demonstrations. Search is an experimental alternative that compares legal plays over sampled unseen hands. Computation runs locally in a browser worker.
The AI receives its own cards, public bids, revealed cards, known void suits, and scores. It does not receive your hand or the shuffle seed.
Loading evaluation results…
The bidding test isolates a change to bidding: both teams use the same playing heuristic. Teams swap seats on the same seeded deals, and evaluation seeds are separate from training and calibration. Confidence intervals account for the paired deals. The original search test compares the earlier search policy with the earlier Standard policy over individual hands; it is a separate experiment from the full-match bidding test.
The external tests compare released, pinned bot code under the same rules. They do not establish a calibrated Elo rating or strength against expert human players. Search relies on heuristic rollout opponents and approximate hidden-hand sampling. Nil experiments include actual failed-contract scoring and cumulative bag penalties.
Download the evaluation report · Training and evaluation source · Archived original article