PROVENANCE.md

# Provenance

The paper PDF and source archive were downloaded from arXiv. The authors’
repository and public preprocessed datasets were pinned by commit and SHA-256.
The reproduction imports the authors’ public module as a dependency check but
independently restates the convex v2 SGShift and SGShift-A objectives so they
can use a deterministic CLARABEL/SCS failover and equal source/target domain
weights. Knockoff construction uses `knockpy` with explicit seeds.

The public-data simulations follow the authors’ generator/base structure and
fixed sparse feature sets. The controlled complex suite uses the known source
link as an oracle offset to isolate recovery of the target correction; that
scope distinction is explicit in outputs and the logbook. Baseline attribution
uses exact linear/tree SHAP when supported and a declared bounded permutation
response fallback for kernel SVMs.

The complete six-seed `p=500` stress scope was rerun at `n_source=n_target=1500`
after the earlier `n=1000` scope exposed one high-variance destructive-control
cell. This disclosed strengthening changes `p/n` from `0.5` to `1/3` while
preserving every seed, the generator, the feature dimension, and all locked
science thresholds. The no-source-offset control reuses the correct arm's
target-loss-CV penalty, so removing the offset is its only causal intervention.

Poster copy and chart values are deterministically composed from frozen CSV
and JSON outputs. No desktop screenshots, synthetic chart numbers, or hidden
manual edits are used.