.gitignore

# Python
__pycache__/
*.py[cod]
*.egg-info/
.eggs/
build/
dist/
.venv/
venv/
env/

# Testing / tooling
.pytest_cache/
.mypy_cache/
.ruff_cache/
.coverage
htmlcov/

# OS
.DS_Store

# Model artifacts. Checkpoints can be large; only the demo checkpoint is tracked.
# Training/eval logs are tracked: they are the receipts for every number in the README.
runs/*
!runs/*.log
# The study payloads the figures are drawn from (~48KB). scripts/make_figures.py
# reads only these, so committing them is what makes the figures re-derivable from
# a fresh clone rather than trusting the SVGs. The feature caches (runs/*.npz,
# ~4.7MB) and runs/study_datastore/ (~591MB) stay out; both rebuild from scratch.
!runs/*.json
*.pt
*.ckpt
checkpoints/*
!checkpoints/lookaside-demo/
!checkpoints/lookaside-demo/*.pt

# Retrieval datastores are large and rebuildable from the checkpoint + corpus; never commit.
*.datastore
datastore/
*.npy
*.faiss

# Data: large raw corpora are downloaded, not committed — except the eight
# public-domain Doyle books the shipped demo's memory indexes (~3.4 MB total), so
# the retrieval memory builds offline on a fresh clone. The wider classics set
# (used only to retrain from scratch) stays download-only.
data/raw/*
data/cache/
*.bin
!data/sample/*
!data/raw/a_study_in_scarlet.txt
!data/raw/adventures_of_sherlock_holmes.txt
!data/raw/his_last_bow.txt
!data/raw/hound_of_the_baskervilles.txt
!data/raw/memoirs_of_sherlock_holmes.txt
!data/raw/return_of_sherlock_holmes.txt
!data/raw/the_sign_of_the_four.txt
!data/raw/the_valley_of_fear.txt

# Editor
.idea/
.vscode/
*.swp