Skip to content

Independent reproduction protocol

This protocol is designed for a person or organization that is independent of the VertiMosaic author/maintainer.

Clean environment

Create a new environment and install a released wheel from PyPI, not an editable checkout:

python -m venv .venv
. .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install vertimosaic==<released-version>
vertimosaic-reproduce --expected-version <released-version> --output evidence

On Windows, activate the virtual environment with the platform-appropriate command.

The command writes:

  • formal_comparison.csv, .json, and .md;
  • environment.json with OS, CPU, RAM, Python and dependency versions;
  • installed-distributions.txt;
  • reproduction.json with a stable result digest; and
  • SHA256SUMS covering the evidence files.

Publishing independent evidence

Open the Independent reproduction report issue template and attach or link the complete evidence directory. State your affiliation and relationship, if any, to the project.

A maintainer-run GitHub Action, a maintainer machine, or an author-generated evidence bundle is useful reproducibility evidence, but it is not described as independent external validation. The repository only marks evidence as independent when an unaffiliated reproducer publishes it.

Expected comparison

The formal table contains four explicitly separated protocols:

  1. centralized all-feature non-federated baseline;
  2. Bank-only/single-party non-federated baseline;
  3. vertical federated logistic regression; and
  4. vertical federated histogram GBDT.

Utility, runtime, RSS, protocol communication accounting and an empirical confidence-membership attack are recorded together. The optional full benchmark additionally records 50% overlap behavior and documented dropout behavior.