Exact, best found, or high accuracy
Short runs are faster. The default is the setting we recommend. Deeper runs return a tighter EV.
| Result | Label | What that means |
|---|---|---|
| Street and opening EV | High accuracy | Default 5,000 samples. Short is faster. Up to 50,000 is tighter. Not every remaining card. |
| Taiwanese arrangement | High accuracy | All 105 sets vs a policy. Same sample choice, 1,000–50,000. Not a solved game. |
| Variance paths | Projection | v1 self-play. Expected winnings stay 0 unless you set an edge. Not a recorded hand. |
| 2-7 Fantasyland, known cards | Exact optimum | Every legal middle is checked. |
| High-middle Fantasyland | Best found | Best board from the search. A better one can still exist. |
Street EV is high accuracy
The Trainer deals the rest of the hand many times and averages the score. It does not enumerate the rest of the deck.
You set samples per candidate. 500–2,000 is a quick look. 5,000 is the recommended default. 10,000–50,000 is for a close call. Every line sees the same deals. Pruning can drop a line that is already far behind.
Future streets use a playout policy:
- v4 Quick prune — default.
- v5 hybrid — slower, closer than v4.
- v3 Fast terminal — shortlist, then shared endings.
- v1 Heuristic — what Variance uses.
- v2 Full lookahead — deepest, about 15–20× slower per sample.
Three-player Board Setup is experimental: one shared deck, pairwise scoring, no model of turn order or collusion.
Taiwanese EV is high accuracy
All 105 sets, then one shared simulation. Samples run from 1,000 to 50,000. Short is faster. 5,000 is the default. 50,000 is tighter. Balanced is the recommended opponent. The number is against that policy, not a solved game.
A screenshot of seven dealt cards with an empty arrangement is the screen before that rollout. It is not a solved set.
Fantasyland is exact, or the best board found
Thirteen cards placed. Discards returned. Not a street-by-street simulation.
- 2-7 middle is an exact optimum. Every legal middle is checked.
- High-middle is the best board found. A better one can still exist.
High precision is the default. Standard is shorter. Maximum is deeper.
| Precision | Calibration hands | Bottom beam | Middle beam |
|---|---|---|---|
| Standard | 4,000 | 60 | 20 |
| High (default) | 8,000 | 100 | 35 |
| Maximum | 15,000 | 160 | 55 |
2-7 uses the same calibration counts. The middle is fully checked, so those beams do not apply.
Variance charts are a self-play projection
The Variance page estimates a standard deviation from v1 self-play, then draws sample paths from that distribution. The screenshots on the home page are that projection after 2,000 three-way self-play hands, with expected edge set to 0 and 20 displayed paths. They are not a recorded user session.
On those screens the model shows 37.64 points of standard deviation per hand, 0 expected points, and a total standard deviation of 11,902 points over the 100,000-hand horizon it was given. One drawn sample shows a maximum downswing of 7,315 points, a longest downswing of 23,288 hands, and a current downswing of 3,282 points.
Two screens you can re-read
These are photographs of one installed session, not a promise about your next hand.
- Candidates. Complete, 5,000 samples, 232 candidates. The best line on that table is +26.65 EV vs field. Royalties are a column.
- Finished Trainer hand. Street 4 is 0.00 EV lost. The line is +16.33 at 10,000 samples. Earlier streets read −5.75, 0.00, 0.00, and −4.43.
How the engines are checked
The Rust engines are accepted when they match the JavaScript engines on checked-in fixtures: the same cards, seed, and sample count. Hand evaluations and chosen rows must match. Reported EV must match within 1e-6. A sharded run must match the single-thread result. The fixture pack covers the five OFC formats, policies v3 and v5, Fantasyland at standard, high, and maximum, and Taiwanese from two to six players on the 52-card and 54-card decks.
That check is a port test. It does not turn a 5,000-sample street estimate into an exhaustive EV.