Result store as self-contained RecordBundles (8.1G total,
9 files). Every record and tensor is inside the .pb files:
the combined records bundle carries each experiment's report-cited dataset
digest in its metadata, capture bundles carry their tensors inline
(bfloat16 where the payload is losslessly representable), and large
payloads are split into .vNN volumes that any reader merges as a
directory.
Per-file SHA-256:
146e13796f0403a9ac208367f1be0e2f5ec8aac6dc85b07444c01e9dae646b36 quant-welfare-calibration-captures.pb
54e0f0ecd835af8ed0a2b1a8768c53133442555b601b799caa3ce5964527bd23 quant-welfare-records.pb
e577ca2fddc6aaa550d81b3146152d2072c31d570caaeffc4ce1e09597306ccc quant-welfare-s2-modea-1.pb
1ceaa42e9e9a72686c3cc760cb0c99ec3349783296a33b6db3c3ab97d0513585 quant-welfare-s2-modeb-1.pb
57515b7cbb00b1342c1cf26ac85691876fc005416d9697219e2aa1b9fba5760f quant-welfare-s2-modec-1.pb
1791b4b202685986682d87fc628630e7fd1377878d3d1d9875c05740489fe35b quant-welfare-s2-tokens.v00.pb
d3137c10597af4600b8565b9a417fd28b005d62e4271c4b600ed5e4cda0fcbeb quant-welfare-s2-tokens.v01.pb
c2c7e98b97a85ca2af62e70c43f05716c1bd6b834767f769f34802471d349047 quant-welfare-s2-tokens.v02.pb
e5c1e6fedd259dd49718e64b0ab65bc9a3c8b38517eb214d98bf73e6244e6f8e quant-welfare-s2-tokens.v03.pb
Decompose with the shipped CLI (no side files needed):
python3 -m modelwelfare.bundle inspect <file.pb>
python3 -m modelwelfare.bundle extract <file.pb> --out dir [--experiment E] [--condition C] [--uri NAME]
Reproduce from a download directory:
python3 experiments/quant-welfare/report.py --bundle <dir-or-file>
python3 experiments/quant-welfare/analyze.py --experiment study1/confirmatory --bundle quant-welfare-records.pb
python3 experiments/quant-welfare/tools/signature.py --experiment study1/confirmatory --bundle quant-welfare-records.pb