PROTOCOL.md

# Reproduction protocol

## Identity and source lock

- OpenReview: `wpKA7G7Cqu`
- arXiv: `2505.20634v2`
- author repository: `Rachel-Lyu/SGShift` at commit
  `3de99c6c41a8931ee3c8d001055c193d83ba5c89`
- official base claims: three, for an exact six-point ceiling
- exact future tags: `icml2026-repro`, `paper-wpKA7G7Cqu`

All source, data, and challenge-map hashes are recorded before execution.

## Evidence suites

1. Public-data benchmark: Diabetes and SUPPORT2, all four matched model pairs
   and four prespecified mismatched pairs, repeated deterministic seeds.
2. Sample efficiency: target sizes 100, 200, 400, and 800 on both datasets and
   all four matched model families.
3. Robustness: dense shifts, generator/base mismatch, 500-dimensional design,
   AR(1) correlation 0.7, and low-SNR shifts. The complete six-seed
   500-dimensional regime uses 1,500 source and 1,500 target observations
   (`p/n = 1/3`); this strengthens the earlier `n=1,000` (`p/n = 1/2`)
   validation without changing seeds, feature dimension, or thresholds.
4. Destructive controls: shuffled target labels, permuted shifted-feature
   truth, and removal of the fixed source offset. The offset ablation holds
   SGShift's target-loss-CV penalty fixed so the source offset is the only
   intervention.

The All of Us COVID-19 data are restricted and are explicitly outside the
executed scope. Official display anchors mentioning those data are preserved
verbatim in the logbook and marked as not independently tested.

## Resource policy

The full run is CPU-only with two independent cell workers (`jobs=2`), with
BLAS/OpenMP thread-count environment variables fixed at one. The two repeatability pilots are
serialized (`jobs=1`) so their stable evidence can be compared byte for byte.
GPU and MPS paths are disabled. This worker performs no remote write. The
coordinator may repurpose the source-locked blank placeholder only after an
independent audit of the complete public release candidate.