Screw the dealer: what the rule actually costs

The most-argued house rule in this game, and the argument has never had a number in it. We played every game twice, once each way, with identical cards both times. The answer is not the one either side expects.

Short answer, and it is not the one either side of this argument expects: no. The rule does cost the dealer — 2.79 points in every round they deal, measured over 1,200,000 hands played both ways. But the dealer's seat does not start level. It bids last, which is worth more than the rule takes away.

With the rule switched on, the dealer's seat still out-scores every other seat at the table. The hook does not punish the dealer. It claws back part of an advantage the dealer already had.

The comparison that settles it

Mean points per round, for the seat that is dealing against the seats that are not, with the rule off and then on:

Read the last column. Dealing is worth 5.56 points a round with no hook in play, because the dealer hears every other bid before committing to one. Turn the hook on and that edge falls to 2.67 — smaller, and still an edge.

This is why the argument never ends. The people who say the rule is unfair are measuring the right thing — it genuinely costs the dealer points — against the wrong baseline. The dealer's seat is not the neutral seat.

What it costs, table by table

The same measurement per table size. "Cost per deal" is the paired difference: the same seed played both ways, so the two arms are dealt identical cards and the difference is the rule rather than the luck. The ± is one standard error.

The cost is remarkably flat across table sizes, and small — a few points in a game where hitting a bid of three pays 50. For scale, everyone else at the table scores +0.11 ± 0.02 points differently when the rule is on — a tenth of a point, which this many hands can detect and no table ever will. The hook is a tax on one seat, not a change to the game's weather.

It almost never bites — and when it does, it hurts

Here is where the two camps are each looking at a different number. The rule applies to every deal. It only bites when the number the dealer actually wanted to bid happens to be the forbidden one.

• It bit in 21% of deals (30,306 of 142,500).
• In a further 28% it could not apply at all: the other seats had already bid past the round number, so the forbidden bid was not a legal bid anyway and the dealer was free.
• When it bit, it cost the dealer 13.11 points (−13.11 ± 0.21) — the difference between a good round and a bad one.
• When it did not bite, it cost exactly +0.00 ± 0.00.

That last line is a check on the harness rather than a finding. If the dealer's preferred bid is not forbidden, the dealer bids it, every seat plays the same seeded cards, and the round is bit-for-bit the round that the other arm played. A result of anything other than zero there would mean the two arms were not comparable and nothing else on this page could be trusted.

So both sides of the forum argument are reporting honestly. One is averaging 13.11 points over the 21% of rounds where it lands; the other is remembering the rounds themselves.

It is mostly a round-one rule

The surprise in the data. How often the hook bit, by round, at a four-player table:

It bites hardest in round 1 (50%) and then fades away to almost nothing in the long rounds. The reason is arithmetic, not psychology: the forbidden number is the round size minus what everyone else has already bid, and bids systematically undershoot — a table bidding for 15 tricks will commit to maybe twelve. So in a long round the forbidden number sits well above anything the dealer wanted, and the rule is a dead letter. In a one-card round there are only two numbers to choose from and one of them is banned.

Practical consequence for anyone adopting this rule: it does not make the endgame tense. It makes the opening rounds tense, which is the opposite of what most groups think they are buying.

The deal does not come round evenly

"It evens out, everyone deals the same number of times" is the usual rebuttal, and it is false — for a reason that has nothing to do with this rule and that we have not seen written down anywhere. A game is 60 ÷ players rounds, and that is never a multiple of the number of players. At four players it is 15 rounds, so 3 seats deal 4 times and 1 deals 3. At six players it is 10 rounds: 4 seats deal 2 times, 2 deal 1.

Measured over whole four-player games, by where you sit relative to the first dealer:

The first dealer pays the most — 16.22 points over a game — and 3 seats along pays 5.10. Note which way round that is: the seat that pays most is not the one dealing the biggest hands, it is the one dealing the early, small rounds, because that is where the rule bites. The cards-dealt column is there to make the point that hand size is not what drives this.

The spread is small in absolute terms, and it is a real seat-order handicap decided before a card is dealt. If it bothers your table, the fix is free: draw for seats at the start of each game rather than keeping last game's order.

How this was measured

The harness is packages/server/dealer-stats.mjs in our own repository, run against the same engine that deals the live game. It plays 2,500 seeds at each of three, four, five and six players, and plays every seed twice — once with screwTheDealer on and once off. That is 20,000 complete games and 1,200,000 seat-rounds. Every seat is the engine's strongest bot.

The pairing is the part that makes the small numbers above meaningful. The engine's shuffle is seeded on the game seed and the round number and nothing else, so round seven of the on arm and round seven of the off arm deal the same 60 cards to the same seats, with the same seat dealing. Every difference on this page is therefore a paired difference, and its standard error is computed over those pairs rather than by comparing two independent averages — which at a round-score spread of this size would need orders of magnitude more games to say anything at all.

To separate "applies" from "bites", the harness asks each dealer twice: once with the forbidden number set, and once with nothing forbidden, from the identical hand and table position. Where the two answers differ, the rule bit. Where they agree, it did not.

What these numbers do not tell you

The dealer's edge is as big as the dealer's use of it. The 5.56-point advantage of bidding last is our bots' advantage. They shade their bid by what the table has already committed to, but only mildly and only in the last couple of seats. A sharp human exploits bidding last harder than that, which makes the edge bigger and the rule look even more justified; a table of beginners who ignore the other bids gets less of it, and for them the hook really is close to a straight penalty. That is the single number on this page most likely to differ at your table.

When blocked, our dealer bids low. 76% of the time a blocked bot stepped down rather than up, because that is the tie-break in legalizeBid. Humans who are blocked often go up instead, chasing. Bidding up when forced is the more aggressive error and would show a larger cost than we report.

It measures bots, not people. No signalling, no grudges, no reading the player who always overbids. And no table talk — which matters here specifically, because at a real table the hook changes the bidding behaviour of everyone else: seats bid to set the dealer up. Our bots cannot do that, so this is the cost of the rule's mechanics with none of its politics.

It is one scoring system. 20 plus 10 per trick bid, minus 10 per trick missed — the standard on our scoring page. Variants that pay a successful zero by round number would shift these figures.

Everything here is reproducible: the bots are seeded, so running the harness returns these numbers. The tables are generated from its output at build time rather than typed in, and so are the claims that could flip — which seat pays most, whether the dealer's seat is still ahead. Re-run it and the page restates itself or contradicts itself out loud.

Play it both ways

It is a checkbox, not a rebuild: open a table, and "the bids may not add up" sits in the room settings, off by default. Turn it on for a game and off for the next one with the same group — that is the experiment above at a scale of one, and the hand you remember will be a round-one deal. More on this and the other additions groups make on variants and house rules.