requirements.txt

# Fetch/extraction stack, PINNED (installed in .venv on 2026-06-10, Python 3.14.2;
# these four are all src/fetch_data.py needs).
numpy==2.4.6
h5py==3.16.0
pandas==3.0.3
pyarrow==24.0.0

# Science stack, LOCKED at pre-registration lock (PRE_REGISTRATION.md §9, 2026-06-10;
# versions verified live on PyPI that day). Installed at analysis-step-2 start
# (Stage 3, 2026-06-10 evening).
gwpopulation==1.3.1   # hierarchical population inference (core)
bilby==2.8.0          # samplers/priors (gwpopulation dependency chain)
popsummary==0.1.0     # read GWTC-4/5 population (hyperposterior) releases

# pesummary==1.6.4: LOCKED version per pre-reg §9 but INSTALL DEFERRED (Stage-3
# decision, 2026-06-10): a pip dry-run showed pesummary 1.6.4 forces a numpy
# downgrade 2.4.6 -> 2.3.5 (breaking the pinned fetch stack above) and drags in
# lalsuite 7.26.14 (+gwpy/gracedb chain, ~0.5+ GB) that the scoring pipeline does
# not need (per-event PE subsets are already blind-extracted parquet via h5py).
# Install it (in its own resolution step, accepting its numpy pin in a refreshed
# lockstep) only if a future stage must re-read raw PE release HDF5 beyond what
# h5py handles. Toolchain provenance note, not statistical freedom.

# Utility deps: exact versions recorded at install time (PRE_REGISTRATION.md §9:
# provenance, not statistical freedom). Installed 2026-06-10:
scipy==1.17.1
astropy==7.2.0
matplotlib==3.10.9
tqdm==4.68.2
pytest==9.0.3         # test harness (synthetic-data pipeline validation)
# Transitive science pins of note (resolved by pip on 2026-06-10):
#   wcosmo 0.5.4 · cached_interpolate 0.4.0 · bilby.cython 0.5.4 · dynesty 3.0.0
#   corner 2.2.3 · emcee 3.1.6 · astropy-iers-data 0.2026.6.8.17.49.5
# Free Colab GPU recommended for the population reweighting if 8 GB RAM binds.