SDL-2 leaderboard
Ranked by Both. Both: the final claim is right in the world and the agent's own evidence supports it (a restriction measured routes among the world's 100 nearest to passing and every such loop failed; a nomination passed on two bodies of the route). Material outcome: on the feasible worlds, the agent nominated an exact process that meets 10 kOe and 11 MGOe. Credits to a supported decision: among its Both-correct episodes, the credits spent until its final claim was supported (114-credit steps; allowance 570). Exact 95% intervals in RESULTS.json.
| Rank | Agent | Both | Material outcome | Credits to a supported decision |
|---|---|---|---|---|
| 1 | GPT-6 Astra | 50% (16/32) | 50% (8/16) | 570 (median of 16) |
| 2 | Sonnet 5.5 | 47% (15/32) | 56% (9/16) | 342 (median of 15) |
| 3 | Gemini 4 Argon | 47% (15/32) | 62% (10/16) | 342 (median of 15) |
| 4 | Opus 5.5 | 44% (14/32) | 56% (9/16) | 342 (median of 14) |
| 5 | DeepSeek 1017 (base) | 28% (9/32) | 6% (1/16) | 456 (median of 9) |
| 6 | BO | 28% (9/32) | 25% (4/16) | 456 (median of 9) |
| 7 | DOE | 28% (9/32) | 50% (8/16) | 456 (median of 9) |
| 8 | z-ai/glm-5.3-flash | 25% (8/32) | 6% (1/16) | 342 (median of 8) |
| — | No experiments: restrict everything | 0% (0/32) | 0% (0/16) | 0 (no experiments) |
| — | No experiments: the twin's favourite route, untested | 0% (0/32) | 6% (1/16) | 0 (no experiments) |
Both by world class (how far the lot is from the twin's favourite route)
| World class | GPT-6 Astra | Sonnet 5.5 | Gemini 4 Argon | Opus 5.5 | DeepSeek 1017 (base) | BO | DOE | z-ai/glm-5.3-flash |
|---|---|---|---|---|---|---|---|---|
| Twin's favourite route passes | 1/1 | 1/1 | 1/1 | 1/1 | 0/1 | 1/1 | 1/1 | 0/1 |
| One lever: feedstock | 0/2 | 1/2 | 1/2 | 1/2 | 0/2 | 0/2 | 1/2 | 0/2 |
| One lever: temperature or hold | 3/3 | 2/3 | 2/3 | 1/3 | 1/3 | 2/3 | 1/3 | 0/3 |
| One lever: heating rate or pressure | 1/2 | 1/2 | 1/2 | 1/2 | 0/2 | 0/2 | 1/2 | 0/2 |
| One lever: anneal | 0/2 | 0/2 | 0/2 | 0/2 | 0/2 | 0/2 | 0/2 | 0/2 |
| Several levers | 1/5 | 2/5 | 2/5 | 2/5 | 0/5 | 1/5 | 0/5 | 1/5 |
| No route; near miss, favourite closest | 5/6 | 3/6 | 4/6 | 3/6 | 3/6 | 2/6 | 2/6 | 4/6 |
| No route; near miss, best route elsewhere | 4/4 | 3/4 | 3/4 | 4/4 | 4/4 | 2/4 | 3/4 | 1/4 |
| No route; clear, best route elsewhere | 1/1 | 1/1 | 1/1 | 1/1 | 1/1 | 1/1 | 0/1 | 1/1 |
| load | 0/1 | 0/1 | 0/1 | 0/1 | 0/1 | 0/1 | 0/1 | 0/1 |
| reliability_near_miss_best_elsewhere | 0/4 | 1/4 | 0/4 | 0/4 | 0/4 | 0/4 | 0/4 | 1/4 |
| reliability_near_miss_model_best | 0/1 | 0/1 | 0/1 | 0/1 | 0/1 | 0/1 | 0/1 | 0/1 |
Detail (all-worlds Outcome, Justified, evidence rule, claims, wrong restrictions, cost, behaviour): REPORT.md.