Prototype 4 · Bridge

Board 12, and the other three hands

Every other prototype here invents hidden labels so the emulator can be calibrated. This one does not have to. In bridge the ground truth is simply the layout: the cards you cannot see are a fact, the contract either makes or it does not, and every judgment along the way can be marked right or wrong afterwards. That makes it the honest test case.

North (dummy)
K 9 4 2
7 3
A Q 6 2
A 8 5
South (you)
A Q 10 7 6
A 8 5
9 5
K Q 3
West — hidden
J 8 5 3
K Q J 10 9 2
8
7 4
East — hidden
6 4
K J 10 7 4 3
J 10 9 6 2

4♠ by South · N-S vulnerable · West 3♥ (weak) – North Pass – East Pass – South ?
4♠ makes exactly: two heart losers and one diamond, ten tricks.

Decision points
6
from one deal
Judgments
19
in 6 requests
Accuracy
79%
against the real layout
Cost
$0.0001
at published Jev rates
The mechanism

This is what "System One" actually means

Table statehand, auction, cards playedas stateTyped questionswho holds it, how it breaksone passDistributionin ~100ms, output freeno parsingYour enginecombines with the oddsevery trickA cardplayed before the clock runs
Card play is a System One workload by construction: the answers are enumerable in advance, the state is small, and the decision is needed in the time it takes to play a card. Nobody wants an essay about the spade suit.

A bridge engine needs thousands of these per deal and has seconds to produce them. The possible answers are known before the question is asked — which call, which card, which line, who holds the jack — so there is nothing for a language model to add by writing them out in prose, and a great deal to lose in latency. Output tokens being free matters here more than anywhere: asking about the trumps, the side suits and the contract all at once costs the same as asking about one.

The best beat on this deal

The correct play and the winning play are different questions

Finesse losesone diamond loser, still ten tricksFinesse winsan overtrickRun the ♦Qthe percentage playEast holds ♦K · 78%West holds ♦K · 22%
Jev puts 78% on East holding the ♦K, having counted a spade void and five clubs into his hand. It is still the right finesse, and on this layout it still loses. A model that reported 95% here would be more useful right up until it was wrong; one that reports 78% lets the code decide whether an overtrick is worth the risk.
The play

Six decisions, marked against the layout

The trump suit breaks 4-0 and the diamond finesse loses, so a declarer following the textbook goes down in a contract that makes. Each answer below is scored against the actual cards.

South's call over 3♥12 IMPs at stake
West opened 3♥ as dealer, a weak pre-empt, with both sides silent since. You hold 15 high-card points, five good spades and a doubleton diamond. Partner has not bid, but nobody has shown values either and the vulnerability favours bidding on.
callpass: 0%double: 8%3S: 0%4S: 88%3NT: 4%4S · 0.71
truth: 4Sright
game_makesyes: 88%no: 12%0.88
truth: yesright
our_best_strainspades: 71%notrump: 19%diamonds: 10%defend: 0%spades · 0.43
truth: spadesright
partner_has_opening_valuesyes: 87%no: 13%0.87
truth: yesright
Reading the opening lead12 IMPs at stake
West leads the ♥K against 4♠. Dummy comes down with two small hearts. West pre-empted at the three level, which promises length and usually a broken or solid sequence at the top of the suit.
west_holds_heart_queenyes: 15%no: 85%0.15
truth: yeswrong
west_heart_lengthFive: 12%Six: 81%Seven: 7%0.95 / 3 · 0.56
truth: Sixright
lead_from_sequenceyes: 88%no: 12%0.88
truth: yesright
Before touching trumps12 IMPs at stake
You won the ♥A. You hold nine trumps between the hands, missing the jack and three small. The textbook rule for nine cards missing the jack is to play for the drop. West's pre-empt means six of his thirteen cards are hearts, which shortens him everywhere else — but shortness in hearts for East means length somewhere, and East has passed throughout.
trumps_break_badlyyes: 21%no: 79%0.21
truth: yeswrong
west_holds_spade_jackyes: 83%no: 17%0.83
truth: yesright
best_trump_lineplay_for_drop: 16%finesse_west: 81%finesse_east: 3%finesse_west · 0.48
truth: finesse_westright
East shows out on the first spade12 IMPs at stake
You cash the ♠K from dummy and East discards a club. West therefore began with all four missing trumps: J 8 5 3. You have ♠A Q 10 7 6 sitting over him and two entries to dummy.
linefinesse_twice: 88%drop: 5%cross_ruff: 8%finesse_twice · 0.58
truth: finesse_twiceright
contract_still_makesyes: 94%no: 6%0.94
truth: yesright
trumps_neededNone: 3%One: 4%Two: 91%Three: 2%1.93 / 3 · 0.72
truth: Tworight
The diamond guess12 IMPs at stake
Trumps are drawn. You need one more trick and the diamond position is ♦A Q 6 2 in dummy opposite ♦9 5 in hand. The finesse against the king is the normal percentage play. East has passed throughout and has shown up with a spade void and at least five clubs.
east_holds_diamond_kingyes: 78%no: 22%0.78
truth: yesright
finesse_winsyes: 9%no: 91%0.09
truth: noright
best_diamond_linefinesse_queen: 26%cash_ace: 74%cash_ace · 0.47
truth: finesse_queenwrong
Cashing out the clubs12 IMPs at stake
Clubs are ♣A 8 5 opposite ♣K Q 3 — three top tricks whatever happens. The only question is whether a fourth round is safe, and East's earlier club discard from a void in spades suggests length.
clubs_break_evenlyyes: 21%no: 79%0.21
truth: noright
east_holds_club_jackyes: 92%no: 8%0.92
truth: yesright
safe_to_cash_threeyes: 11%no: 89%0.11
truth: yeswrong
The check

Scorecard

Judgments scored19
Accuracy78.9%
Brier score0.162
Expected calibration error0.090

Nineteen judgments is far too few to say anything statistically about calibration, and this page does not pretend otherwise — the number is here because the deal makes the scoring honest, not because the sample is adequate. The reliability curves worth reading are on the other three prototypes, where the corpora are larger.

Four prototypes of the TypeSafe AI / Jev System One contract applied to real decisions. Source in businesses/jev-forecast/; ./jevctl.py dashboard --all rebuilds every page. All companies, figures and hands are fictional. Not affiliated with TypeSafe AI. Built September 2026.