MIDAS Data portfolio

Midas · a portfolio of real operating work

Sixteen years of real operating history. Four collections.

Almost everything a model trains on today was written with nothing at stake. This is the opposite: the data, the code, the paperwork and the operating practice of two businesses that carried real customers and real money for sixteen years — plus a live environment you can point an agent at today.

16Years of production history
1.1MRides, joined to the money
3.4 TB+Raw company record
850K+Lines of shipped code
47Graded agent tasks
The Go X OpsGym environment site, showing 1,111,327 rides and 699,996 riders
Go X OpsGym — the operations environment, live now
The Go X rider app showing a live scooter map, range and unlock rates
Go X rider app — eight years in the App Store
The Go X fleet: electric scooters and vehicles
Go X — 5,500+ scooters launched across 15 cities
01 — The sources

Go X, Virool, and the OpsGym.

Everything here was produced by one of these three. Two operating companies and the environment built out of one of them — with what each business ran, what it recorded, and which parts of it you can take.

02 — How this plugs in

Where each part lands in your stack.

This is not a document set to read. Each part maps to a specific stage — corpus, supervision, evaluation, environment — and arrives in the form that stage expects.

01

Pretraining corpus

3.4 TB+ of raw business record — mail, files, chat, calls, ledgers, telemetry, logs — and 850K+ lines of production code with full commit history. JSONL shards, per-document provenance, scrub report attached.

02

Reasoning traces

Decisions with the deliberation still attached: redlines against the negotiation, counsel memos against the correspondence, incident threads against the fix. Input, judgement, and what the operator did next.

03

Evals with ground truth

Every arc here already ended. Score a model against the outcome the business actually reached — the settlement signed, the deal closed, the unit recovered — not against a rubric written afterwards.

04

A live RL environment

OpsGym is running today: 47 graded tasks behind 47 tools, deterministic resets, and reward computed from the world state and the P&L an agent leaves behind. Point a model at it and read the score.

03 — The portfolio

Thirty works, each with its numbers.

Operational data, codebases, legal and corporate, and one live environment. Open any work for what is inside it, what it is good for, how it was prepared and what it cannot do. Scope is set per engagement — one work, one collection, or all of it.

04 — Provenance & preparation

Where it came from, and what was removed.

Every work carries a rights file, a scrub report and a manifest, so provenance is a document your team can check rather than a claim you have to take.

05 — Request access

Tell us what you are training or evaluating.

We come back with a sample manifest, the rights and scrub documentation, and an NDA. The diligence material comes first — none of it sits behind a commitment.

If it is the environment you want, OpsGym can be pointed at a model this week.

Sent straight to us. No mail client, no attachment.