Environment

The GraphJin Agent Environment

Train and measure AI agents against a governed graph, with rewards that come from the database rather than from plausibility.

Most ways of scoring an agent ask a model whether an answer looks right. This one asks the database. Every task in a GraphJin environment carries a hidden oracle — a read-only query resolved against the same world the agent works in — so a fluent, confident, wrong answer earns nothing.

The environment ships as a container that boots ready with no files mounted, and as a CLI you can point at your own schema. Same engine, same reward contract, two starting points.

Train

You want an environment to train against.

Pull the image, read /health, and drive graded episodes — hosted, one completion at a time, or with your own agent over MCP. Held-out splits, trajectory export, and a reward that a policy cannot talk its way past.

Run the environment

Measure

You want to know how an agent does on your data.

Clone the shape of a running GraphJin server into a local synthetic world — catalog structure and published value sets, never your rows — generate a verified suite from it, and grade against that.

Use your own graph

What A Graded Environment Is Hidden oracles, isolated resettable worlds, and a reward a policy cannot talk its way past. Run The Environment In Two Minutes Pull the image, read /health, and drive one graded episode. Nothing to mount, no suite to generate. Three Ways To Drive An Episode GraphJin calls your endpoint, you supply each completion, or you bring the whole agent. Same world, same grading. Measure Your Agent On Your Own Graph Clone the shape of a running GraphJin server into a local synthetic world, generate a verified suite, and grade against that. Building And Baking Worlds Generate new organizations to train against, and bake one into an image your team shares. Training A Policy Against It Teacher distillation, then rejection sampling, then GRPO — and why that order is forced rather than chosen. graphjin env And graphjin eval Reference Every subcommand and flag of both namespaces, plus the GJ_ENV_, GJ_SUPPORT_ and GJ_GENERATOR_ variable families. Environment HTTP API Every route a served environment exposes, with request and response shapes. Suites, Splits, Packs, And Trajectories What each artifact is, who writes it, and what breaks if you edit it by hand. Reward, Comparability, And Provenance What the two profiles optimize for, and the exact fields that decide whether two numbers can be compared.