# Cold-start reported-statistics check

Package version: 1.0.0 · prepared 2026-09-05.

## What this checks

Recomputes means from 20 reported DAVIS S2/S4 paired fold records, enumerates all 1,024 sign assignments per contrast for two-sided exact sign-flip p-values, and applies step-down Holm across the four S2/S4 by CI/RMSE contrasts.

Paper DOI: https://doi.org/10.1038/s41598-026-69754-2

The external paper-associated archive is https://doi.org/10.5281/zenodo.21805175. That archive is distinct from this small portfolio-authored package. Its installation environment and full experiments were not executed here.

## Run locally

Prerequisite: Python 3.10 or later, CPU only. No third-party packages, GPU, network, credentials, or dataset download is required.

Extract the complete ZIP, open a terminal in its folder, and run:

```sh
python check.py
```

An assertion failure exits nonzero. A successful check prints the numeric result and a scope-specific PASS message. The equivalent notebook contains the same Python code; open notebook.ipynb with its working directory set to this extracted folder. A notebook environment is optional and is not bundled.

## Expected output

Adjusted p-values: S2 CI 0.19921875; S2 RMSE 0.0390625; S4 CI 0.05859375; S4 RMSE 0.19921875. Only S2 RMSE is below 0.05.

Numeric reference values are in expected.json. The scripts check them deterministically; there is no random seed or stochastic training.

## Inputs and provenance

source.json is generated from content/cold-start-folds.json, the existing 20-row portfolio transcription of the reported paired-fold evidence. Model-score columns are retained for inspection; statistics use deltaCi and deltaRmse.

provenance.json records the source, scope, package version, and SHA-256 hashes of the executable check, input data, and expected output. README, Python, equivalent notebook, source JSON, expected JSON, and provenance are included together in the ZIP.

## Limits

Uses published delta columns because subtracting independently rounded scores can differ in the last digit. This does not retrain either model, recompute bootstrap intervals, infer incomplete or unrun cells, reproduce interaction-level predictions, or establish quantum advantage. Exact sign-flip arithmetic checks do not establish the experimental assumptions.

Fresh network verification of the external archive was unavailable during package preparation. These small checks use the portfolio's existing sourced transcription, not a newly downloaded archive. A successful check must not be described as reproducing the full paper.