Project directory

my-study.project/
├── project.json                  project manifest — name, author, date, version, fs layout
├── config/
│   ├── experiment.json           run settings (iterations, timing, seeds, parallelism)
│   ├── deployments.json          deployment arms
│   ├── endpoints.template.json   endpoint structure template (commit this)
│   ├── endpoints.json            endpoint URLs and auth config (never commit — gitignored)
│   ├── secrets.template.json     API key name template (commit this)
│   └── secrets.json              API key values (never commit — gitignored)
├── input/
│   ├── text.json                 text library
│   ├── queries.json              query templates
│   └── prompts.json              prompt bindings (text × query)
├── output/                       created automatically on first run
│   └── <timestamp>/
│       ├── runs/                 raw per-request JSON responses
│       ├── manifest.json         full run record (source of truth)
│       ├── manifest.csv          flattened CSV version of manifest.json
│       ├── summary_counts.csv    aggregated counts by text × query × deployment
│       ├── summary_pct.csv       percentage-normalised summary
│       └── run-config.json       snapshot of config used for this run
└── log/                          created automatically on first run

What to commit

File	Commit?
project.json	Yes
config/experiment.json	Yes
config/deployments.json	Yes
config/endpoints.template.json	Yes
config/secrets.template.json	Yes
config/endpoints.json	No — gitignored
config/secrets.json	No — gitignored
input/*.json	Yes
output/	Your choice — output files are large but reproducible
log/	Optional

endpoints.json and secrets.json are separated so endpoint config can be shared with collaborators while key values travel through a secrets manager.