Example notebooks¶
Read the notebooks in order. The first three generate their own simulation with
glossogen.testing; the fourth reads committed CSVs exported from a real cohort.
None needs an API key or the network, so all four run in a fresh clone before
you have credentials.
| 01_read_a_run.ipynb | The JSONL event log: what a run is made of, one row per message, characters per round, round verdicts |
| 02_score_a_run.ipynb | The metric layer through run_scenario_evaluation, deterministic metrics and then a judge-backed one against a stub |
| 03_compare_runs.ipynb | Two runs differing in one knob, scored the same way, read side by side |
| 04_analyze_an_export.ipynb | The export CSVs read in pandas: a real 40-run budget sweep, one plot per table, the empty-versus-zero rule in the data |
The example CSVs under data/ are real exports; data/README.md
records the cohort and the commands that produced them.