Reliable data. Inspectable analytics. Measured AI.
The project connects three pieces: a bounded chess-game ingestion pipeline, opening and clock analytics with explicit denominators, and a read-only AI analyst whose answers carry snapshot-bound evidence. Local DuckDB and Parquet keep the system reproducible; dbt models and recovery checks demonstrate how the data is validated before publication.
The dashboard runs locally. The website lets you inspect recorded results, explore measured opening positions and move orders, try illustrative controls, and follow the offline demo.
What the comparison teaches
Jev is TypeSafe AI's typed decision model. OpenAI Decisions is the optional classifier studied alongside it. Both select task routes; neither replaces the checked calculations that produce a chess metric.
Better routing did not establish better final answers. The retained paired experiment found equal correctness and a higher cost with the Jev gate. The later Decisions final-answer comparison uses saved-answer replay rather than a new live pairing. These are small, dated experiments, with failures and limitations retained alongside their results.
Read further
| Question | Evidence and guide |
|---|---|
| How does the pipeline work? | Architecture |
| How are metrics defined? | Definitions and denominators |
| What do the models demonstrate? | Measured comparison |
| Which study boards occur in my games? | Local PGN analysis, with explicit player selection and a bounded public reference |
| Which opening positions recur across move orders? | Transposition analysis and learning design |
| Can I reproduce a small example? | Offline walkthrough |
| What is complete and what comes next? | Current project status |
The documentation guide indexes operational runbooks, analytical memos and historical evidence. The repository contains the maintained code and contracts. Project author: jfk-ssa on GitHub.