Public coding-agent RL environment

MathCode
Mini 🧮

Inspect coding and mathematical reasoning tasks through typed repository tools, public tests, and terminal rewards. This is an expired, visible compatibility release for researchers to inspect and reproduce.

No paid Space compute required

One task, real interaction

The public environment includes a workspace, action budget, test budget, state hashes, replayable events, and a visible contract grader.

Source on GitHub ↗ Dataset ↗

Explore the public contract

Taskbisect_repair_v1

Assumption-sensitive repair of a certified numerical solver.

Actions8 typed tools

Inspect, edit, test, inspect the diff, and finish.

Reward7 public checks

Semantic correctness, robustness, and integrity.

Reasoning track3 imported tasks

Proof-unit repair, theorem-scope auditing, and adversarial claim checks.

Source formatMathCode bundles

Normalized problems, dependencies, budgets, and editable assessment artifacts.

Research dataSanitized Dataset

Public projections without canonical proofs or private evaluator labels.

Click the button to display the sanitized public report.
Boundary: this static Space displays a public baseline report. It does not execute arbitrary Python or expose Ulam’s private evaluator, active tasks, or scorer credentials.