Make the model useful
within an operating budget.
Control the plant without breaking its limits. Detect a weak plume without flooding the operator with indications. Raise timely well alarms without occupying normal operation.
Every operating group stays visible
Run your controller, then inspect the complete result
The kit sends delayed observations and the current target to your program. Late or invalid actions are counted and replaced with a hold. A fresh grader reconstructs each request and the nonlinear plant from the saved commands.
python control_runner.py --participant 'python my_controller.py' --out my-run python grade_saved.py --directory my-run --out my-replay.json
Files are processed in this browser. A preview checks membership and operating gates; it does not authenticate traces, run a model, submit an entry or establish rank.
Why these tasks are demanding
Hidden states, changed dynamics and delayed sensors expose nominal-control weaknesses. Plume recall must coexist with per-date indication limits. Well coverage must coexist with low false alarms and low alarm occupation.
Measured failures of named references support a development challenge. Independent attempts are still needed to establish community difficulty. Research bars were chosen by the organizer; feasibility and site requirements need separate evidence.
What remains outside the evidence
A committed, unopened synthetic control reserve is separate from public development. It is not a physical trial, and no hidden scoring service is operating.
Sensor calibration, chemical gas/no-gas truth, independent field acquisitions, target-device timing and site acceptance remain open. The original reserved well groups retain their separate prospective protocol.
Read the reserve commitmentSource rights and reproducibility
The control equations are an independently written adaptation of Johansson’s adjustable-zero laboratory benchmark. The process additions and operating gates are disclosed adaptations. Plume source-derived statistics retain CC BY-NC 4.0; 3W-derived statistics retain CC BY 4.0. Authored control code is MIT and the plume scorer component is GPL-3.0. Per-artifact rights · Frozen protocol · Scientific source hashes