Research
Papers, specifications and field notes on code, documents, agent environments, markets, compute and physical AI.
Papers and specifications
- Codebase Collection: A Full-History Repository Corpus
- Prediction Markets as an Event-Driven Signal
- The GPU Rental Market: Price Dispersion and the Cost Curve of Compute
- Retail Financial Sentiment as a Systematic Signal
- Mapping Prediction Markets to Securities
- The Falling Price of Inference
- From Trained Model to Silicon
- Operational Telemetry: A Cross-Tool Activity Graph
- Open Robot Training Format
- Empirical Characterization of the Sim2Real Gap
Field notes
RSS feed- Partial Credit Is What Turns a Benchmark Into an Environment
- The Plagiarist Gate: Proving a Benchmark Cannot Be Passed by Recall
- How to Audit Codebase Data Before You License It
- How Git History Becomes Code-Model Evaluation Data
- What Code Data Do Codex, Claude Code, and SWE Agents Need?
- Repository-Level Code Data for Coding Agents
How we check our work
| Record the source | The data keeps where it came from, how it was collected, its usage rights and the dates it covers. |
|---|---|
| Show the steps | Each result comes with its manifests, timestamps, mappings and checks, so you can see how it was made. |
| Test before publishing | We run point-in-time tests, holdouts, permutation checks and leakage reviews before we publish a number. |
| Publish the limits | We publish null results, sample limits, exclusions and open questions with the result. |