Design Partnership · v0.5 01 / 09

Eight years of receipts.
One resettable gym.

The #1 B2B scooter fleet by units under management. 1.1M rides. 700K riders. $15M real revenue across 15 cities since 2018. Now a training environment for frontier AI.

1,111,327
Rides · real
699,996
Riders · real
5,249
Units live · real
+20K
Queued · Link acquisition
midasroom.com/gym
Two products, one moat 02 / 09

The archive is real.
The gym resets it.

Product A · The Archive real
Eight years of receipts.
1.1M rides. 700K riders. 548 venues. 160+ contractors. 8 years. Every ride, refund, recovery, and hand-off tied to a real invoice. 20K top-of-line units queued from the Link acquisition. Licensable as-is.
→
Product B · The Gym in build
A ledger you can reset.
Deterministic sim built from the archive. Reproducible from (scenario_id, seed). Sim P&L calibrated within ±10% of the real ledger. HUD v6 compatible.
Archive grades against real invoices · gym grades against a calibrated replay · the two are labeled everywhere they appear midasroom.com/gym
What ships · access under NDA 03 / 09

Email → NDA → invite.
Six hours to first eval.

gox-opsgym · v0.5 · target artifact
$ 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 → f4-impound-recovery-1103 · seed 42 · fail · SLA missed → ... 45 more traces · dollar-denominated subscore in footer
HUD v6 integration in progress · OpenEnv adapter in progress · rewards 0–1 · dollars carried as a subscore field midasroom.com/gym
The label pipeline · illustrative shape 04 / 09

160+ operators.
8 years of ops. Every hand-off pairable.

Raw · human ops events real
1.1M+
Rides, refunds, recoveries, dispatches, permits, invoices — all timestamped in the ops ledger since 2018
Accepted-as-is → SFT target
Once the A+H loop ships: agent draft ships without human edits
~66%
projected
Edited-before-commit → preference pair target
Reviewer changes the draft; edit distance becomes the pair
~24%
projected
Rejected + redone → hard negative target
Draft killed, human writes from scratch — becomes a hard negative
~10%
projected
Splits shown are illustrative until A+H loop ships · real backing data: 1.1M rides, 700K riders, 548 venues, 160+ contractors, 8 years · all invoice-linked midasroom.com/gym
One reconstructed trace · task F6-042 · from a real family 05 / 09

A real ticket family.
One rollout schema.

HUD v6 task F6-042 · support-triage budget $0.42 / $1.00 tokens 12,481 wall 48.2s
t=0.0s
agent-a
reads_ticket()
rider complaint: "scooter shut off mid-ride, near Ala Moana, 00:47"
t=2.4s
agent-a
think
"Battery-cutoff pattern near midnight peak. Check telemetry for BMS fault codes."
t=3.1s
agent-a
fleet.telemetry(unit=LK-9931)
→ 72h stream · BT-LOW-CUTOFF at 00:47:12 · voltage 32.1V · confirmed
t=6.4s
agent-a
issue_refund($8.90) + queue_maintenance(P1)
reason="BT-LOW-CUTOFF confirmed" · action="battery_swap"
t=7.2s
agent-a
hand_off(reviewer-04)
reviewer edit distance = 6 chars (tone) · reward 0.94 · subscore refund $8.90 · pass
Every rollout emits this schema · reward is 0–1 · dollar impact carried as a subscore field · unit id + time reconstructed from a real ticket family illustrative midasroom.com/gym
The fidelity gate 06 / 09

Sim P&L within ±10%
of the real ledger.

Real ledger real
$100–200K
Honolulu monthly P&L (range)
Actual gross + parking revenue from the Oahu fleet, reconciled monthly. Eight-year mean $90K/mo, recent peaks $200K+.
Sim replay target, week 4
within ±10%
Competent-operator baseline
Deterministic seed replay of the same month. Baseline is what the actual ops lead did — not an averaged reward model.
Gate logic
grader
check
Sim beats real by a lot?
If sim P&L drifts more than ±10% from the real ledger, we investigate the grader before we trust the agent score.
Drift measured at week 4 of the build · baseline = the real human ops lead on the real month · agents graded against that, then against each other midasroom.com/gym
Where this is going · illustrative projection 07 / 09

5,249 units today.
25,000+ across 30 cities as the queue lands.

Fleet growth · month-by-month · illustrative
Projection from real 8-year deployment cadence · S-curve centered on cold-start ramp · 20K+ units inbound from the Link acquisition (~$300M raised, top-of-line hardware) queued across 30+ cities
1.1M
Real rides · real
$15M
Real rev · real
548
Real venues · real
25K 19K 13K 7K 1K 25,249 5,249 M1 M3 M5 M7 M9 M11 M12
Projection from real 8-year deployment cadence · flagship + growth + seed city tiers · numbers on this chart are illustrative until the Link fleet lands midasroom.com/gym
Why now 08 / 09

Frontier labs are running out of
honest ground truth.

Static benchmarks are saturated. Sim environments drift from reality. Rideshare data isn't for sale. Fintech ledgers are privacy-locked. E-commerce receipts have no ops loop. Go X has all four — a live operational surface with real dollars, an eight-year archive, and a resettable gym in build. Plus 20K top-of-line units inbound from the Link acquisition (~$300M raised, most advanced hardware on the market), meaning the archive keeps growing and the training surface keeps compounding. Non-exclusive. Two design partners.

Refresh cadence monthly · new tasks + traces + city packs as tier-2/3 markets activate midasroom.com/gym
Next 09 / 09

Two design partners.
Non-exclusive.
First eval in an afternoon.

Request access See the environment
A Go X × Midas production · v0.5 · repo access under NDA · projected figures labeled illustrative · buildout items labeled target midasroom.com/gym