Home/Data

Data Catalog

Transparent, auditable task scorecards from Jigsaw’s RL training environments. Each task is evaluated across frontier models with deterministic grading and full provenance.

ML Productivity
View more →
Analysis and debugging tasks over ML training artifacts. Agents navigate training runs, extract metrics from heterogeneous file formats, read framework source code, and synthesize diagnostic reports.
LeaderboardDataTasks
claude-opus-4 [max]
73% ± 5%
grok-4.3 [reasoning]
60% ± 6%
gpt-5.5 [xhigh]
55% ± 7%
0%20%40%60%80%100%
AutoResearch
GPU-accelerated ML research tasks where agents design experiments, run training jobs, analyze results, and iterate on model architectures and training procedures.
Coming soon
0-1 Tasks
Greenfield engineering tasks where agents build functional systems from scratch: APIs, data pipelines, tooling, and end-to-end applications across multiple languages and frameworks.
Coming soon

Data trained the last generation. Environments will train the next.

By late 2026, every serious AI company will need continuous RL training on domain-specific workflows. Enterprises, neo-labs, and frontier model teams will all fine-tune on internal processes. The bottleneck is environments.

We own that bottleneck. Our simulated worlds let teams evaluate agents on real business processes before deploying them to production.