Fresh student humanizer comparison v2 — November 7, 2026 (Pacific)

Three fixed original synthetic academic passages, first outputs only.
Configurations:
- StealthGPT Super, quality mode through the API.
- WriteHuman API Standard, academic tone, English, enhanced=false.
- Undetectable.ai: University, Essay, More Human, v11sr.
- Unmodified original source baseline.

External detectors: GPTZero 2026-11-04-base; Originality 5.0 model-5.0; Pangram 4.1.
Each source/output was sent byte-identically to all three detectors.
Vendor-specific measures remain separate; raw probabilities are not averaged.
Threshold counts:
- GPTZero: human-class probability >50%.
- Originality 5.0: AI-assisted text percentage <50%.
- Pangram 4.1: human-text fraction >50%.
Classification labels are also included. The Undetectable research output meets the
Pangram fraction threshold but its headline is AI Detected. A threshold count does
not necessarily mean a vendor classification of Human Written.

This is a demonstration of these exact configurations, not a population pass rate.
Turnitin was not tested. Output selection was fixed before external scoring.
No rehumanize calls, rerolls, paid POST retries or manual output edits were used.

Package contents:
texts/: complete original, Super, WriteHuman Standard and Undetectable v11sr texts.
provider-evidence/: requests, final outputs and every Undetectable status poll;
account/document IDs and remaining balances removed. Text and models unmodified.
detector-evidence/: all 36 scoring records. Originality administrative scan properties
and billing fields removed. Authentication headers were never saved.
scores.csv / summary.json: factual results and qualitative editorial findings.
verify.py: verifies every text hash, observed model, score and aggregate count.
file-sha256.json: integrity hashes of package files.

Originality note: earlier source/Super requests selected Turbo but the provider
returned model-5.0. WriteHuman and Undetectable scans explicitly selected model-5.0.
All comparisons use the observed 5.0 AI-assisted text percentage, not Turbo confidence.

Fixture provenance and claim checklists were frozen before humanization. Editorial
notes identify individual source/output changes; they are not validated quality
scores or independent evaluator ratings. Inspect the complete outputs.

No credentials, private account usage receipts, real student submissions or customer
data are included. v1 files remained unchanged; this v2 package adds Undetectable
outputs and scoring evidence.
