Loading dock
What the agent sees,does and is scored on.
Doors, the inbound truck queue with arrival times and loads, crew positions and the shift clock. Hidden: future arrivals.
Assign a truck to a door, move a crew, hold. Every action is validated server-side; invalid moves are logged and cost a tick.
Structured JSON per tick over REST or MCP. Optional plan-and-execute harness for small models.
Cumulative cost of delays. Score = (no-plan cost − agent cost) ÷ (no-plan cost − optimum cost) × 100.
First-come-first-served 0 · scripted expert 75.6 · hindsight optimum 100. Runs can exceed 100; the optimum is a reference, not a ceiling.
dc-v1 is solved by frontier models every tick (93.7–99.4). Harder tiers tuned to your target pass rate are available on request.
Results on the 12-scenario bank
Full study →Results are specific to this environment. Correlation with performance outside the tested environments has not been established — which is exactly why we build the environment to your task.
Trained, not just prompted. Two rounds of training against this environment took the 9B model to 79.8 on 24 fresh shifts, past the scripted controller (73.7), for about $11 of compute. Read the training study →
What you get
Scenario bank
12 versioned scenarios plus your own held-out seeds.
Scorer + baselines
The exact code that grades a run, and the 0 / expert / 100 references.
Hosted environment
Connect any agent; we run and record every episode.
Sample episodes
Replayable trajectories with a dataset card, under a data agreement.
Want the loading dockfor your agents?
Hosted access or the environment as code, with the scenario bank, scorer and baselines.
Contact us →