# Sources and credit These tools combine local evidence ledgers, process supervision and policy checks. The following public work supplies context for ideas used in the workflow. There is no claim that this repository implements or reproduces those research systems. - Cynthia Dwork and coauthors, [Generalization in Adaptive Data Analysis and Holdout Reuse](https://arxiv.org/abs/1506.02629), 2015: adaptive reuse of evaluation data can undermine validity. The harness records a holdout disposition; the paper's reusable-holdout algorithm is not implemented. - Avrim Blum and Moritz Hardt, [The Ladder: A Reliable Leaderboard for Machine Learning Competitions](https://arxiv.org/abs/1502.04585), 2015: repeated public feedback can overfit a competition leaderboard. Selection-service isolation and query policy remain external to this code; Ladder is not implemented. - Google DeepMind and coauthors, [AlphaEvolve: A coding agent for scientific and algorithmic discovery](https://arxiv.org/abs/2506.13131), 2025: related work on program proposal and automated evaluation. This harness provides source intake, experiment receipts and bounded workers rather than an evolutionary search implementation. - Owen Queen, Evan Gorrell and coauthors, [DiscoveryAgent: A Framework for Scientist-centric Agentic AI](https://arxiv.org/abs/2611.21950), 2026, and its code ([github.com/evangorrell/DiscoveryAgent-app](https://github.com/evangorrell/DiscoveryAgent-app), MIT). Adopted: the audit agent, as `rh audit LANE "question"` ([harness/audit.py](../harness/audit.py)). It is a read-only agent that answers from a lane's own ledgers, sweep reports and cited evidence, cites the files, and adds no opinion; its prompt follows theirs (`application/frontend/AIR/utils/code_chat.py`). It is the one part of the paper with a measured effect: in a small study (8 participants) people using it found 3.5 of 36 planted errors on average, against 0.25 with the code alone. Not adopted: the exploratory first round, because no experiment in the paper isolates it and the harness's normal-first rule (trivial predictor and simple baseline before any predictive claim) covers the same ground; the human hypothesis pick, because the paper's one comparison (steered against automatic on ProteinGym) also changed the method, so it does not show the pick helps; the per-round configuration snapshot, because sweep runs already keep each stage's exact prompt, remote command and model route (a resume that reads them back is still missing); and the `curl | sh` installer, because piping a download into a shell runs code nobody read, and `scripts/bootstrap.sh` already installs from a clone in one command. - Hugo Fernandes, 2026-11-03: the ten anti-overfitting practices mapped in [overfitting.md](overfitting.md). He wrote them with Claude's help, and Claude ran the simulations behind practice 10 (SEM penalty, safe-plus-ambitious pair). - Edwin B. Wilson, [Probable Inference, the Law of Succession, and Statistical Inference](https://doi.org/10.1080/01621459.1927.10502953), 1927: the role ranker uses the lower endpoint of the Wilson score interval with z = 1.96. Source links were checked on 2026-11-03. Python, pytest, Git, zsh and the optional provider CLIs retain their own licenses and attribution. No third-party package source is vendored here.