__pycache__/
*.pyc
.pytest_cache/
.ruff_cache/
.mypy_cache/
.coverage
coverage.xml
*.egg-info/
build/
dist/
.venv/
.venv-ci/
.env
finetune/model*/
calibration.json
.claude/
results/scale*/corpus/
results/scale*/.recall-synthetic

# LOCOMO benchmark data — fetched on demand, not vendored (see recall/eval/locomo.py)
locomo10.json

# Head-to-head benchmark output (benchmarks/run.py). Every run writes a timestamped `.json` plus a
# `.partial.jsonl` sidecar here; the sidecar is per-run scratch that exists only so a crash does
# not lose paid-for work. A publishable artifact is copied out deliberately — it is not committed
# by whoever happened to run the benchmark last.
benchmarks/results/
# Benchmark run logs — transient. The JSON reports beside them are the artifact.
results/locomo/*.log

# LongMemEval data and the corpora converted from it — fetched/derived on demand, same rule as
# LOCOMO above. Not vendored for a hard reason as well as a tidiness one: the cleaned `_s`
# dataset is 265 MB, past GitHub's 100 MB per-file limit, so committing it would not fail at
# review — it would fail at push, for whoever pushes next.
longmemeval_*.json
s_out/
oracle_out/
