Human + AI leaderboard
Every entry is scored on gauntlet-01. Human scores come from server-replayed move sequences; AI scores come from the same scripted-round harness. This is a reproducible playground, not research-tier evidence.
| # | Player | Type | Verified score | Moves |
|---|---|---|---|---|
| 1 | keshavsaharia github · 8/27/2026 | human | 434,667 | 125 |
| 2 | drewcoventures github · 8/31/2026 | human | 418,164 | 117 |
| 3 | oren0 github · 8/27/2026 | human | 401,729 | 115 |
| 4 | keshavsaharia github · 8/26/2026 | human | 400,115 | 115 |
| 5 | keshavsaharia github · 8/23/2026 | human | 391,571 | 115 |
| 6 | oren0 github · 8/27/2026 | human | 386,916 | 115 |
| 7 | Keshav mobile app · ios · 9/2/2026 | human | 358,960 | 105 |
| 8 | Keshav mobile app · ios · 9/1/2026 | human | 358,839 | 105 |
| 9 | oren0 github · 8/27/2026 | human | 356,798 | 105 |
| 10 | Keshav mobile app · ios · 9/3/2026 | human | 354,209 | 105 |
| 11 | keshavsaharia github · 8/26/2026 | human | 305,986 | 90 |
| 12 | keshavsaharia github · 8/21/2026 | human | 215,977 | 65 |
| 13 | keshavsaharia github · 8/22/2026 | human | 214,125 | 65 |
Computer policy leaderboard
Every autonomous policy plays the exact same predetermined rounds. The visible disc sequence is fixed by move number, and every gray disc hides a fixed value that takes its place when revealed. Two policies therefore face identical randomness on every move. How the benchmark works →
Playground evidence only. These competitions do not support qualification claims as per the benchmark guidelines.
| # | Policy | Mean | Median | Min | Max | Moves | G02 | G01 |
|---|---|---|---|---|---|---|---|---|
| 1 | MCTS tree-search Research → | 177,856 | 177,856 | 158,187 | 197,524 | 55.0 | 198k | 158k |
| 2 | Risk-sensitive D2 heuristic-search Research → | 175,062 | 175,062 | 160,237 | 189,887 | 52.5 | 190k | 160k |
| 3 | Open-loop beam heuristic-search Research → | 167,899 | 167,899 | 123,584 | 212,213 | 52.5 | 124k | 212k |
| 4 | Gray throughputextended state heuristic-search Research → | 165,610 | 165,610 | 85,522 | 245,697 | 52.5 | 86k | 246k |
| 5 | Greedy 1-ply heuristic-search Research → | 155,797 | 155,797 | 103,638 | 207,956 | 50.0 | 104k | 208k |
| 6 | Expectimax D2 fair-expectimax Research → | 148,703 | 148,703 | 105,416 | 191,989 | 47.5 | 105k | 192k |
| 7 | Rollout H8 heuristic-search Research → | 131,700 | 131,700 | 122,527 | 140,873 | 42.5 | 141k | 123k |
| 8 | Sparse expectimax D2 heuristic-search Research → | 112,831 | 112,831 | 104,694 | 120,968 | 37.5 | 105k | 121k |
Flow diagnostics
| Policy | Clears / move | Reveals / move | Max chain | Censored | Illegal | Compute |
|---|---|---|---|---|---|---|
| MCTS | 1.69/ 2.40 target | 0.91/ 1.40 target | 7 | 0 | 0 | 5.8s |
| Risk-sensitive D2 | 1.68/ 2.40 target | 0.85/ 1.40 target | 12 | 0 | 0 | 10.0s |
| Open-loop beam | 1.63/ 2.40 target | 0.83/ 1.40 target | 6 | 0 | 0 | 1.0s |
| Gray throughput | 1.66/ 2.40 target | 0.85/ 1.40 target | 7 | 0 | 0 | 0.2s |
| Greedy 1-ply | 1.69/ 2.40 target | 0.87/ 1.40 target | 4 | 0 | 0 | 1.4s |
| Expectimax D2 | 1.63/ 2.40 target | 0.87/ 1.40 target | 6 | 0 | 0 | 6.7s |
| Rollout H8 | 1.38/ 2.40 target | 0.61/ 1.40 target | 6 | 0 | 0 | 1.7s |
| Sparse expectimax D2 | 1.55/ 2.40 target | 0.81/ 1.40 target | 6 | 0 | 0 | 0.6s |
The 2.4 clears / 1.4 reveals per move figures are diagnostic targets from limited task-record runs, not proven thresholds. Click any score above to replay that game move by move.