This is the first Midas environment, and we built it on our own company. Go X operates electric scooters for hotels, resorts and campuses. Every ride, refund, recovery and permit since 2018 sits in one invoice-linked ledger. We turned that operation into a deterministic environment a frontier lab can run on its own hardware.
Compute became something you rent by the hour. Reality never did. A lab cannot rent two hundred field crew, ten cities and eight years of consequences. So we rented ours.
A dataset is a photograph. This is a place. The agent takes an action, the world changes, the ledger moves, and seven days later there is an outcome to grade against.
One image. One command. Reproducible from a scenario id and a seed. Runs the same on the HUD platform, local Docker, or your own sandbox. No shared state, no network surprises.
The ceiling is not a reward model. It is what the human ops lead actually posted over the same thirty days, on the same fleet. Agents are scored against a person who did the job.
1.1M rides. 700K riders. 548 venues. 160+ contractors. Every ride, refund, recovery and hand-off tied to a real invoice. Licensable as it stands.
A deterministic simulation built from that archive. Rebuildable from a scenario id and a seed. Simulated P&L calibrated inside ten percent of the real ledger.
The archive grades against real invoices. The gym grades against a calibrated replay. The two are labeled everywhere they appear.
The archive is real: eight years of live operations, invoice-linked. The gym is under active build, calibrated against that archive. Nothing here is claimed as shipped unless it is.
A simulation nobody trusts is a toy. So we hold the gym to the real books before we let it grade anyone.
The profit and loss a person actually posted over that window. This is the ceiling every agent is scored against.
Minus 4.3 percent from the real ledger. Well inside the gate. Same fleet, same window, same clock.
If the simulated books drift more than ten percent from the real ones, we investigate the grader before we trust any agent score built on top of it.
A rider writes in at quarter to one in the morning. Here is what the agent does, what it costs, and how it is graded. Reconstructed from a real ticket family in the archive.
Graders read world state, never prose. A plausible memo with no write calls scores zero. One logged violation can floor the whole reward.
Reads are cheap. Every write returns an idempotent receipt and validates against hard constraints. Cause and symptom live in different tools by design, so no task can be solved from one call.
De-identification runs deterministic pattern matching plus a model pass, writing irreversible tokens. Fourteen identifier classes masked. The full method ships in the data card with the environment.
Cumulative dollars, five seeds averaged. The human number at the top is the ceiling. The agent rows below show the target scoring formula on illustrative runs. Real model scores publish when a lab runs the environment. illustrative
The gap is the whole point. The best agent run lands at roughly one eighth of what a person posted on the same fleet in the same month. That distance is what there is to train on.
The environment, the fidelity gate, the task families, the rights posture, and where the fleet goes next. Twenty pages, no build required.
The environment is available under an NDA. Six-hour setup. Non-exclusive. Bring your model and score it against a person who did the job.