Play rubber bridge in your browser → A full rubber against three robots. Free, no sign-up, nothing to install.

How the Robots Work

What your three opponents are actually doing, and what they can see.

Three Opponents, Two Engines#

North, East and West are all played by the same software, and it is built from two quite different pieces:

• A rule engine — a large body of bidding and card-play knowledge written out explicitly. It knows what a 2/1 response promises, when a takeout double applies, that you draw trumps unless you need them, and which card to lead from K J 10 9. It answers instantly and it can explain itself, which is why every robot call comes with a reason.
• A Monte Carlo search — it deals out complete hands consistent with everything the robot legitimately knows, solves each one exactly with a double-dummy solver, and picks whichever choice does best on average. It knows no bridge at all; it just counts.

The two are used together. The rule engine proposes the sensible candidates and provides the bridge judgement; the search checks them against the actual cards where it has the time to do so. When the search cannot finish in the time available — which happens far more often early in a hand than late — the rule engine's answer stands.

Both run in a background thread, so the interface never freezes while a robot thinks.

How the Robots Bid#

Bidding starts with the rule engine. It reads the auction so far in your chosen system, applies the convention switches you have set, and produces a short list of candidate calls — each with a reason attached. That reason is what you read when you tap a call in the auction grid, and what appears in the Analysis screen's auction review.

The search then tries to referee between the candidates. For each one it deals out plausible hands for the three players it cannot see, plays the auction out to its end with the rule engine sitting in every seat, works out the resulting contract, solves it double-dummy, and scores the outcome using rubber scoring — including the current vulnerability and part-score. The candidate with the best average score wins.

In practice, a full 52-card double-dummy solve is expensive, and the search frequently runs out of time before it has scored enough deals to be trusted. When that happens the robots defer to the rule engine rather than act on a half-finished sample — deliberately, because a confident answer from bad data is worse than a sound conventional bid.

Two things the robots genuinely do take into account, and which most simple bridge programs do not:

• The part-score. The below-the-line position is passed into every bidding decision, so the robots compete harder when a small contract would finish their game, and harder still to stop yours.
• Vulnerability, as rubber bridge scores it — not duplicate's fixed schedule. A save is judged against what the opponents' contract is actually worth at this point in this rubber.

How the Robots Play#

Card play runs the same two-engine pattern, weighted the other way round.

The Monte Carlo search leads. For the seat to play, the app builds a picture of everything that seat may legitimately know, deals the unseen cards into the other hands in a way consistent with that picture, and solves each resulting deal exactly. Every legal card is scored on two things: how often the contract makes after it, and how many tricks declarer takes on average. Declarers maximise the first; defenders minimise it.

Where two cards come out exactly equal — which is common once the fate of the contract is settled — the robot plays the cheaper one. That is why they do not throw kings away on dead tricks.

The search now reaches trick one. It did not always. A full 52-card solve is far too slow for a playable game, and the app used to fall back to a rule engine for the first few tricks — which is where contracts are decided. It no longer does: a position part-way through a trick is now broken into the handful of ways the trick can finish, each of which the solver answers quickly. Measured across 148 deals a side, the search returns an answer for 95–99% of decisions in tricks one to five, on both defence and declarer play.

The rule engine covers the rest. Standard opening leads, second hand low, third hand high, cover an honour, draw trumps, keep honours, discard sensibly, and read partner's signals in whichever carding system you have chosen. It is consulted on roughly a quarter of decisions — but mostly on positions where the search has just reported that every card is worth the same, and someone still has to pick one.

The robots play techniques nothing in them is named after. There is no rule in the app that mentions holding up, unblocking or ducking. They were tested anyway, by setting the position and reading the card, and on those positions the engine finds the play 20 times out of 20. It works them out at the table rather than looking them up — which is also why the rule engine on its own, asked the same questions, gets two of them wrong.

As cards disappear the position shrinks further. By the endgame the robots are effectively playing double-dummy — which is to say, perfectly. If you are going to beat a contract, do it early — though "early" buys less than it used to.

What the Robots Know#

The robots do not cheat, and the design makes that structural rather than a promise.

Before any robot decision, the app builds a knowledge envelope for that seat: a self-contained snapshot of what that seat may legitimately know. It contains the seat's own cards, dummy once dummy is down, the full history of played cards, and nothing else. It holds no reference to any concealed hand at all — which is what allows it to be handed to a background thread without any possibility of a hidden card leaking into the decision.

What the robots deduce from that envelope is exactly what a good human deduces:

• Voids — a player who failed to follow a suit holds none of it. Pure inference from public information.
• Auction constraints — point ranges and suit lengths implied by every call made, read through your system and convention settings. A robot that hears you open a weak 1NT will not deal you a 19-count.
• Card counts — how many cards each unseen hand still holds.

Sampled deals must satisfy all of that. The one allowance is the standard one: declarer and dummy see each other's cards, because declarer plays both.

Hints and Analysis are different, and openly so. They do look at all four hands — that is their job. A hint during the play is double-dummy advice, and the Analysis screen's judgements are made with complete information. That is why the hint is a better teacher than an opponent.

Where the Robots Are Weak#

Knowing the shape of your opponents' weaknesses is a legitimate part of playing them, and worth stating honestly.

• Some techniques are checked; most are not. Three have been tested by setting the position and reading the card — holding up to keep the danger hand out, unblocking to keep a suit running, and taking your ace when the trick is the setting trick. The engine plays all three, every time. Suit combinations, safety plays, endplays and defensive planning have not been tested, and nothing here claims they are played well.
• Their opening leads are unverified. Twice measured, twice inconclusive — not because the robots did badly, but because most opening leads barely change the outcome, so there is often no right card to be judged against. Treat the first card of each hand as the least examined part of their game.
• The fallback rule for second hand is wrong on some holdings. "Second hand low" is right far more often than not, and it is what the app plays when the search finds nothing to choose between. On a long suit headed by a sequence it is the losing card. Three replacements were built and measured and none was an improvement, so it stands — knowingly.
• They are still strongest at the end. The search now answers almost every early decision too, so the old advice to attack in the first two or three tricks has weakened. What has not changed is the endgame: by then they are playing perfectly, so a plan that needs them to err late will not work.
• They do not deceive. The robots play the cheapest card among equals, which is honest and readable. A human would sometimes falsecard; they will not.
• They read your bidding literally. Their sampling assumes your calls mean what your convention settings say they mean. A psychic bid — a deliberate lie — will be believed completely, which is very effective and quite unsporting.
• The bidding is more conventional than the play. Because the bidding search so often runs out of time, most auctions are the rule engine's: sound, systemic, and less imaginative than a strong human's.
• They do not model you. There is no memory across deals and no adjustment to how you play. Each deal starts fresh.

None of this makes them easy — in the play, from about halfway through a hand, they are considerably better than most club players.

Pace of Play#

A robot that answers instantly reads as a machine, so each move waits a beat before it lands. Relaxed, Normal and Fast in Settings set how long that beat is — roughly 0.8–1.2 seconds, 0.5–0.85, and 0.2–0.4.

This is presentation, not thinking time. The pause and the computation run together, so the wait you see is the longer of the two rather than their sum, and choosing Fast does not make the robots play worse. On a slow device a genuinely hard decision may occasionally take longer than the pause you chose; that is the search finishing, not the pause.

The pace also applies to leaving a deal: if you go home mid-hand, pending robot moves are cancelled, and the deal is exactly where you left it when you resume.