Apppublic

ProCreations/repro-efficient-privacy-loss-accounting-for-subsampling-and-random-allocation

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes

App README

Reproduction bundle

Deterministic CPU audit of the six exact registered claims for OpenReview HDBpda5Vih and arXiv 2602.17284v1.

bash

python3 -m pip install -r requirements.txt
PYTHONPATH=official PYTHONWARNINGS=error python3 reproduce.py \
  --output replay --official-root official \
  --pdf source/paper.pdf --source source/source.tar
python3 build_manifest.py --verify

All 23 scientific gates and all 283 upstream tests pass in two byte-identical warning-strict runs. The audit covers the exact subsampling transformations, finite-support random-allocation identities, binary convolution schedule, geometric-grid accuracy, exact t=1,000/10,000 bounds, the Bernoulli privacy comparison, and three PREAMBLE block mappings.

Claims 5 and 6 are honestly falsified as literally registered: the substantive results are supported, but the paper places them in Figures 5 and 3, respectively, rather than the registered Figures 4 and 5.