Reproducibility package for the ReleaseMind paper (Aureuma/ReleaseMindPaper).
This repo contains the public artifacts and executable pipeline to reproduce the paper’s quantitative results on commodity CPU hardware.
src/releasemind_repro/— CLI package and reproducibility engine.configs/paper.toml— paper defaults for sampling, thresholds, and cost model.data/— bundled fixtures and cached source inputs.outputs/— generated artifacts from command runs.artifacts/— reference manifest bundle for verification.docs/— setup and reproducibility instructions.tests/— command/config validation coverage.
cd ReleaseMindRepro
uv sync
uv run releasemind-repro --config configs/paper.toml doctor
uv run releasemind-repro --config configs/paper.toml doctor --jsonuv run releasemind-repro --config configs/paper.toml reproduce-paperThis runs:
- SmartNote proxy training (
outputs/risk_proxy/smartnote_proxy.parquet) - RNSum proxy construction (
outputs/risk_proxy/rnsum_proxy.parquet) - Oracle audit over sampled rows
- Risk-control evaluation and coverage/risk tables
- Cost–risk routing simulation + plots
- Reference comparison against committed artifacts
uv run releasemind-repro --config configs/paper.toml train-proxies
uv run releasemind-repro --config configs/paper.toml audit --provider gemini --sample-size 3000 --max-workers 5 --min-interval 0.4
uv run releasemind-repro --config configs/paper.toml evaluate-risk --compare-corrections
uv run releasemind-repro --config configs/paper.toml simulate-routingGOOGLE_API_KEYorGEMINI_API_KEY: Gemini API key for oracle audits.AWS_ACCESS_KEY_ID,AWS_SECRET_ACCESS_KEY,AWS_DEFAULT_REGION: optional Bedrock fallback credentials.- Never store secrets in this repository.
- RNSum source:
https://github.com/nlab-mpg/RNSum-Dataset(CC BY 4.0). - SmartNote source:
https://github.com/osslab-pku/SmartNote(+ replication package, CC BY 4.0). - Upstream datasets are not bundled in full; only lightweight fixtures are stored.
outputs/eval/risk_control_summary.csvoutputs/eval/risk_control_summary.meta.jsonoutputs/routing/simulate_routing.csvoutputs/figures/risk_control_curve_rnsum.pdfoutputs/figures/risk_control_curve_smartnote.pdfoutputs/figures/routing_cost_risk_rnsum.pdfoutputs/figures/routing_cost_risk_smartnote.pdf
Reference bundle (required for strict reproducibility checks):
artifacts/reference/
Prefer this order in commands and scripts:
uv run releasemind-repro --config <path> <command><command> --config <path> is still supported for backward compatibility.
Code and docs in this repository are licensed under the repository-level license for this reproducibility package. Upstream datasets and code remain subject to their original terms.