Open-sourced experiments from Cassis. Each experiment is self-contained in its own directory, with the code, the question sets, the published run data, and instructions to reproduce every figure and number in the accompanying write-up.
| Experiment | Question it answers |
|---|---|
derivation-distance/ |
How much does spelling things out in the context (rules → metric formulas → worked examples) help an LLM write correct analytics SQL, and how does that interact with model capability? |
Every experiment directory ships:
README.md— design, how to reproduce the figures offline, and how to re-run the benchmark itself.images/— the write-up's figures, regenerable from the published data.results/runs/— the published run files the write-up's numbers come from.requirements.txt+ self-testing modules insrc/.
Code and text are MIT-licensed (see LICENSE). Benchmark data
carries its own per-dataset licenses — notably the Olist dataset is
CC BY-NC-SA (non-commercial); see each experiment's datasets/ directory.