# Measured-data provenance — chapters 4–13

ArtificialGate · 7 October 2026. These are historical public benchmarks, not generated replacement observations. `extract-data.py` produces `data.mjs`; `data-manifest.json` records SHA256 for each downloaded archive, extracted member and canonical browser rows. Runtime modules require no network request.

The extractor refuses unreviewed changes to the pinned source-member hashes. A row hash uses compact UTF-8 Python 3.13 JSON (`ensure_ascii=False`, separators `,` and `:`), preserving numeric float notation; it is not a hash of JavaScript's independently reserialized numbers. The manifest additionally pins the complete runtime-module bytes.

## Attribution and license

Each official UCI dataset page below states **Creative Commons Attribution 4.0 International**, verified on 7 October 2026. The data retain that license: https://creativecommons.org/licenses/by/4.0/. ArtificialGate does not claim ownership of the underlying observations or endorsement by the authors. Data are provided without warranty; benchmark results are not clinical, deployment, or population-level claims.

| Dataset | Attribution and original source | Browser data |
|---|---|---|
| Iris | Fisher, R. (1936). https://doi.org/10.24432/C56C76 · https://archive.ics.uci.edu/dataset/53/iris | 150 rows, four measurements in cm, three classes |
| Wine | Aeberhard, S. and Forina, M. (1992). https://doi.org/10.24432/C5PC7J · https://archive.ics.uci.edu/dataset/109/wine | 178 rows, 13 chemical measurements, three cultivars |
| Breast Cancer Wisconsin (Diagnostic) | Wolberg, W., Mangasarian, O., Street, N., and Street, W. (1993). https://doi.org/10.24432/C5DW2B · https://archive.ics.uci.edu/dataset/17/breast+cancer+wisconsin+diagnostic | 569 rows, all 30 measured features; benign=0 and malignant=1 |
| SMS Spam Collection | Almeida, T. and Hidalgo, J. (2011). https://doi.org/10.24432/C5CC84 · https://archive.ics.uci.edu/dataset/228/sms+spam+collection | 1,200 measured messages, 1,049 ham and 151 spam |

## Disclosed transformations

- Iris uses UCI's `bezdekIris.data`, the repository's corrected variant. This selection avoids silently mixing the historical `iris.data` errata with the corrected values. All four downloaded measurements are preserved. Class strings become indices 0–2.
- Wine retains all 13 downloaded measurements; original class codes 1–3 become indices 0–2.
- WDBC retains all 30 measurements in source order: ten means, ten standard errors and ten worst values. Original case identifiers are omitted. Source row positions provide stable, non-patient-identifying IDs. The positive class is malignant; this differs from scikit-learn's default class coding.
- SMS URL-like strings, email addresses and all numeric sequences are replaced with `[url]`, `[email]`, and `[number]`. Whitespace is normalized. This is risk reduction, not a guarantee of complete anonymization; no raw-message excerpt is required in the public interface. Names and ordinary public benchmark language may remain in the downloadable module. Exact case-insensitive sanitized duplicates are removed, keeping the first source occurrence. The first 1,200 messages ranked by SHA256 of `ag-sms-v1|` plus normalized lowercase text are selected, without balancing classes. Original labels are unchanged. No text, feature values, or labels are invented. Duplicate removal is performed before the fixed split. Near-duplicate templates can remain and limit generalization claims.
- The SMS archive parses to 5,574 physical lines, while some third-party loaders report 5,572 records. This pipeline treats nonempty tab-delimited source lines as records and records the actual source count in the manifest; it does not substitute a third-party copy.

Stable IDs identify source row positions. Plotted projections, predictions and embeddings are derived quantities, not additional observations. For SMS, vocabulary selection, document frequencies, IDF, class weights and two-dimensional truncated-SVD/LSA directions are fitted only to the training partition. The latter is computed locally from the actual document-term matrix; it is **not a pretrained language-model embedding**. Validation and test messages only undergo the fixed training transformation. Exact normalized duplicates cannot cross partitions; this does not establish independence of near-duplicate message families.

The fixed test partition is reused during repeated exploration. It is not a source of fresh unseen samples or secure certification. All measured sources are historical, modest benchmarks; model success here does not establish effectiveness on current messages, unobserved wine samples, or patients.