AI Humanizers for Student Writing: 29-Text Academic Subset

Original tests: October 22, 2026, America/New_York.
Academic-genre analysis: November 7, 2026, America/Los_Angeles.

This is a new exploratory analysis of frozen public results, not a new competitor
service run. We selected source genres before calculating their subset scores.
All 29 matching source texts, all 11 humanizers, and the unmodified source baseline
are included. The source passages are synthetic English texts, not real students'
submissions. Humanized output texts are not supplied by the original archive.
Turnitin is not part of this six-detector panel.

Files:
- academic-subset-selection-rule.json: exact inclusion rule and method.
- academic-subset-membership.json: selected and excluded sample IDs/genres.
- academic-subset-scores.csv: 348 rows (29 samples x 11 tools plus source).
- academic-subset-summary.csv / academic-subset-results.json: derived aggregates.
- academic-subset-source-metadata.csv: selected sample genres/models/source hashes.
- original-public-data/: original full public source CSV, score CSV and summary.
- verify.py: recomputes membership, checks file/text hashes, recalculates all results.

Run: python3 verify.py
No API keys, network requests, packages or paid calls are required.

AI scores in the original CSV are percentages from 0 to 100. A strict pass requires
AI percentage <50 for a particular detector. The mean strict-bypass rate averages
six per-detector pass rates; it is NOT the rate of clearing every detector.
All-six count is recorded separately. Raw probabilities are not averaged.

Attribution / data license:
HumanizerEval, October 2026 review, October 22, 2026,
https://www.humanizereval.com/cycles/2026-10/
https://www.humanizereval.com/sources/
Data: CC BY 4.0, https://creativecommons.org/licenses/by/4.0/
Source texts were published by HumanizerEval with their prompts and model names.

Changes / analysis:
Academic genre selection, derived CSVs, result calculations and interpretation by
StealthGPT, November 7, 2026. Original full public data files are unmodified.
The original experiment was operated by the StealthGPT team using external detector
APIs. This analysis does not establish comparative rewrite fidelity because the
original humanized outputs and per-sample quality judgments are unavailable.

No account credentials, private account usage receipts, private raw API responses,
student submissions or customer data are in this package.
