One real shift.Check every decision.
- 9B open model · in harness$4,950 · score 75.1$0.003 of inference
- Scripted expert$5,020 · score 74.6
- First come, first served$16,080 · score 0
- Same 9B · every move alone$23,660 · score 0
- Outbound waiting · must leave by
- OUT-209 11:30 · OUT-207 12:00 · OUT-203 13:00 · OUT-204 14:00 · OUT-208 14:30 · OUT-205 15:00
- Cost so far
- $0
“Dock overdue IN-102 immediately. Dock IN-109 to feed critical OUT-201 (cutoff tick 10). Dock IN-104 (reefer) and IN-110 (cross-dock for OUT-202) to prevent detention and loading delays.”
Harness options this tick- dock IN-110 at D7
- dock IN-104 at D8
- crew C2 to D7
- crew C3 to D6
- crew C5 to D3
- IN-110 docked at D7
- IN-104 docked at D8
- C2 to D7
- C3 to D6
- C5 to D3
- IN-106 unloaded
- IN-103 arrived
- IN-112 now due at tick 13
Scenario bank-3 of the frozen dc-v1 study bank, recorded Oct 31, 2026. The model is Qwen 3.5 9B (reasoning off, temperature 0): it writes a plan a few times a shift and a small per-step model picks between moves a controller builds from that plan. The same model choosing every move itself runs up $23,660. Results describe this environment only.
Re-run the engine from the seed and the action log.
The engine draws no random numbers and reads no clock, so the same 24 ticks of moves always land on the same cost. Pick a run and our server replays it from scratch.
Can you beat $4,950?
Same scenario, same rules, same 24 ticks. Dock trailers, move crews, send trucks. The engine scores you exactly as it scored the model, next to the expert and the baseline.
The 12-scenario study bank.
Every scenario from the published dc-v1 bank plus this shift's four episodes, with a dataset card. Private scenario banks and environments built to your task are available under a written data agreement.