Experiment outputs for SorryDB. Data only, no code:
the scripts that produce and consume all of this live in scripts/ in the main SorryDB repo.
Split out of the main repo because 68,000 files and 6 GB made every git worktree add slow.
| path | what |
|---|---|
intermediate_experiment_outputs_full_reservoir_3_months/ |
main experiment sweep, per model |
intermediate_experiment_outputs_full_reservoir/ |
earlier full-reservoir run |
intermediate_experiment_outputs/ |
earliest runs |
outputs/, output/ |
aristotle submit/collect output |
charts/ |
generated figures |
*.json (root) |
aggregated databases and analysis results |
Within an experiment run: result.json is the analyzable artifact, individual/ holds
per-sorry results keyed by hash, and logs.tar.gz holds the raw per-sorry logs.
Clone this repo alongside SorryDB, so the two are siblings:
~/some/dir/
SorryDB/
sorrydb-experiments/
Then, from a SorryDB checkout:
just link-experimentsThat creates gitignored symlinks so every script's existing relative paths
(intermediate_experiment_outputs_full_reservoir_3_months/..., charts/...) resolve
with no changes. Re-run it in each new worktree.
The 34,000+ per-sorry .log files are stored as one logs.tar.gz per run (4.6 GB -> ~400 MB).
Scripts that read logs need them extracted first:
just unpack-logs intermediate_experiment_outputs_full_reservoir_3_months/gpt/1000/2026-01-23_02-33-21_llmExtracted logs/ dirs are gitignored, so re-archiving is never needed; just delete them.
*.tar.gz and result.json are LFS-tracked. Run git lfs install once before cloning,
or git lfs pull after, or you will get pointer files instead of data.