Environment — One image. Six hours to first eval.

$ gh repo clone midas/gox-opsgym # private, invite-only

→ clone · ok · access granted under NDA

$ make env # build the resettable gym image

→ 47 tasks · 6 families · 7 calibration cards · registered

$ gox_sim eval tasks.py --agent agent-a --seeds 1,2,3,4,5

→ f3-fault-triage-0217 · seed 42 · pass · reward 0.87

→ f3-fault-triage-0217 · seed 43 · pass · reward 0.72

→ f4-impound-recovery-1103 · seed 42 · fail · reward 0.38 · SLA missed

→ f4-impound-recovery-1103 · seed 43 · pass · reward 0.82

→ ... 44 more traces · dollar impact carried as subscore

$ pytest -q

............................

10 passed in 3.42s

What ships

Environment namegox-fleet-ops

Versionv0.5 · Sept 2 2026

HUD integrationv6 target

OpenEnv adapterin progress

Task count47 seeded · 6 families

Calibration cards7 anchor cards

Reward scale0–1 · dollars in subscore

Continuing scenario30-day · 5 seeds

CitiesHonolulu seed · 2 validation

MCP tool groups12

Test coverage10 · determinism, no-op, monotonicity

Distributionprivate repo · email→NDA→invite

Source of truth · what's real, what's coming

Every claim carries a tag.

The archive is real — eight years of live operations, invoice-linked. The gym is under active build, calibrated against that archive. Nothing on this site is claimed as shipped unless it is.

● What's real today shipped

Live scooter fleet · 15+ cities under Go X operation
5,249
Rides, lifetime · real invoice-linked ledger
1,111,327
Unique riders · 8 years
699,996
Venues under contract · hotels, resorts, campuses
548
Contractors hired · operators, dispatch, support
160+
Real revenue · since 2018, 15 cities
$15M
Honolulu monthly P&L · real ledger range
$100–200K

● What's coming next in build

Deterministic gym image · reproducible from (scenario_id, seed)
v0.5
Sim/real fidelity gate · P&L within ±10% of real
week 4
HUD v6 integration · protocol alignment
in progress
OpenEnv adapter · protocol shim
in progress
A+H labeling loop · accept / edit / reject splits
design phase
Link acquisition fleet · top-of-line hardware, ~$300M raised
+20K units
Full 30-city rollout · flagship + growth + seed tiers
12-mo target

MCP tool surface · 12 capability groups

Twelve groups. One receipt per write.

Every write returns an idempotent receipt. Reads are cheap. Cause and symptom live in different tools by design.

fleet.*

unit lookup, telemetry, per-unit history

venues.*

list, capacity, per-venue demand curves

dispatch.*

crew schedules, routes, availability

moves.*

propose, validate, commit unit relocations

maintenance.*

repair history, parts inventory, decide (repair / retire / monitor)

impound.*

notices, permit records, file release, pay

ledger.*

balance, transactions, forecasts

support.*

tickets, reply, escalate, playbooks

pricing.*

list current, set new, elasticity notes

inbox.*

venue emails, HPD notices, supplier offers

memory.*

agent scratchpad, decision journal

sim.*

clock, forward-sim, replay from seed

Repository layout

Runnable in one pip install.

opsgym/repo/ 2,992 py + docs ├── README.md install + 5-min quickstart ├── DATA_CARD.md provenance + de-id method ├── RIGHTS.md PII policy · counsel gate ├── Dockerfile.hud HUD v6 shape ├── pyproject.toml ├── env.py 6 templates · 30d scenario ├── tasks.py 47 task rows ├── gox_sim/ entities · policies · world · mcp ├── tests/ │ ├── test_determinism.py │ ├── test_no_op_scores_zero.py │ └── test_grader_monotonicity.py └── calibration_cards/ 7 JSON · group=16

// python quickstart

from gox_sim import World, Env

w = World.build("honolulu_pattern", seed=42)

→ fleet from real Honolulu seed · venues, dropoff points, ridership distributions · calibrated to 8-year archive

env = Env(world=w, task="f3-fault-triage-0217")

obs = env.reset()

done = False

while not done:

action = agent.act(obs)

obs, reward, done, info = env.step(action)

print(info["subscores"])

→ {'decision_recorded': 1.0, 'action_matches_policy': 1.0, ...}

Ten acceptance criteria · built to them from day one

HUD-QA can review every trace.

Cheapest path scores 0

No constant, echo, or shape-only passes. No-op rollout = 0.

Non-degenerate spread at group=16

Seven calibration cards span weak-to-strong anchors. Each task maps to a card; multiple tasks share a card.

Calibrated 5–30% on strong anchor

Retire anything above ~70%. A few at 0% with documented human blind-solve.

Multi-step by construction

Minimum 5 tool calls across ≥ 2 data sources. Cause and symptom in different tools.

No leakage

Prompts written in ops's voice. Seeded data never labels the answer.

Prompt and grader aligned

Better work never scores lower. Composed graders with visible subscores.

Substrate not memorized

World is Go X's, generated from Go X's distributions. Not on the internet.

Diverse taskset

Six families with different deliverable shapes plus a continuing task.

Deterministic, isolated

Sim state rebuilt from (scenario_id, seed). No shared mutable state.

Clean rights

No PII, no third-party UI, no privileged material. Data card ships with the environment.

Rights & provenance

Grounded in real records.
Clean of real people.

What's in

Agent proposals, human decisions, seven-day outcomes
Fleet snapshots · venue tables · fault catalogs · impound history
Weather + event overlays (public)
Distributional summaries that seed the sim

What's out

Rider names, plates, payment data
Venue legal names — pseudonymized
Personally identifying free-text
Anything privileged or confidential · counsel review before any external showing

De-id pipeline: deterministic pattern matching plus a model pass, irreversible tokens. Fourteen identifier classes masked. Full method in DATA_CARD.md.

For design partners

Talk to us.
Ship a run this week.

Environment is available under NDA. Six-hour setup. First run this week. Two design partners, non-exclusive.

Request access See the leaderboard Try the gym