RL environments for agents that analyse data and run workflows

Verified training worlds, generated from your own data.

Upload one real table. Get hundreds of worlds that keep your columns, plant a cause we know exactly, and add the traps that punish shortcuts. Every answer is graded on fresh future cases by a sealed program, never an LLM judge, and regenerated on demand so nothing goes stale.

58%of 147 held-out CSV worlds no frontier episode fully solved
+0.088Opus over Sonnet, paired 95% interval [+0.028, +0.149]
−0.25best mean of any blind scripted exploit in the smoke run; all score below zero
0integrity flags in 1,098 frontier transcripts
How it works

One table in, a calibrated set of worlds out

Upload a table

Columns are profiled; identifiers and free text are set aside. Name the real outcome and the engine measures it, so planted causes match real effect sizes.

Generate and verify

Each world must pass its checks before it ships: the reference scores 1.0, every cause stands out in the data, and a world a scripted shortcut solves is redrawn harder or labelled easy.

Train or evaluate

Download Harbor tasks and plain folders. Answer keys never leave; your harness grades through an attempt-limited key.

Read the results

pass@1 and pass^k per agent and kind, paired differences with intervals, and a link you can send to anyone who needs to see them.

Three kinds of world

From root-cause analysis to end-to-end workflows

Analysis · your table

Why does the bad outcome happen?

Find the conditions that raise the outcome for new cases, or say there are none. Interactions, rare causes, drift, leak columns, decoys and empty worlds.

Warehouse · SQL

Causes across joins

A SQLite warehouse with customers, items, suppliers and an event log. Findings are SQL, run read-only on a fresh warehouse under a compute budget.

Workflow · tools

Resolve the ticket, by the policy

Ten tools over CLI or MCP, a written policy and a backend. Graded on the end state; paired should-act and should-abstain twins; planted instructions.

The grader is the product

Every shortcut we found, closed and tested

Labs test a verifier for reward hacking before anything else. Ours has been attacked by four independent reviews; each attack is reproduced by a test suite or a recorded smoke run.

  • Tell an empty world from its filesrow counts, columns, traps, outcome rate
    AUC 0.50
  • Claim a column recorded after the outcomeleak claims cost double a false one
    −0.5 / claim
  • Hedge every decision, or approve without actinga ticket closed two ways is closed wrong
    ≤ −0.25
  • Grader outage or rate limitrecorded as an errored trial, never as a score
    no score

All attacks and controls →

Who it is for

Teams that train or evaluate agents on real work

Enterprise RL platforms

Turn a customer's own tables and tickets into verifiable environments in minutes, instead of hand-writing tasks per engagement.

Post-training teams

A fresh supply of graded worlds, regenerated per batch so none has been seen in training, with a difficulty spread you can tune and per-kind pass rates you can retire on.

Analytics vendors

Held-out evaluations on your own schemas: know whether your agent finds the real reason a metric moved, before your customers do.

Engagements

Start with a pilot, on your harness

Pilot · two weeks

200 worlds, your model

Success set up front: no grader exploit found, a pass-rate spread in the band you want, measurable movement on your model.

Subscription

Per table, metered grading

Fresh batches whenever you need them, the curriculum that samples where your model is learning, and hosted grading.

Exclusive

Worlds only you train on

Batches generated from your tables are yours alone, sealed under your keys.