RL Tasks
Hard, verified coding tasks
Built from private production code no model has seen, for post-training. One real commit becomes one task.
- Per task
- A runnable environment at the before-state, the instruction, a graded verifier with continuous reward from 0 to 1, a hidden reference patch, QC evidence, and solver trajectories.
- Checks
- The reference solution scores at least 0.95 in a cold container; doing nothing and cheat probes score at most 0.1; fairness, mutation and consistency checks pass. Only tasks that solver runs score from 0.1 to 0.8 ship.
- Exposure
- Every source repo passes an exposure check. Every bundle ships a file-tree manifest (sha256 per file plus a root hash) so you can check overlap with your own corpus.
- Pricing
- Priced per engagement, quoted after a scoping call.
- Delivery
- Harbor task format, as a tar.gz with its sha256. 1,000+ new tasks every month.
- Licensing
- Non-exclusive by default.