Our Test Log: Round 1, Raw Results

This is the raw record behind every Our test label on the site. Nothing here is smoothed or summarized away: each row is one detector run on one sample, exactly as logged. Our methodology and scope live on the methodology page; this page is the data.

Downloads: results-v1.csv · results-v1.json (the JSON is the canonical source; the CSV is a flattened view of the same 24 runs).

Round 1 at a glance

• Window: October 12–13, 2026 (all timestamps in the data files)
• Tools completed: ZeroGPT, Sapling (Oct 12); GPTZero, QuillBot (Oct 13) — all free web tiers, default settings, no account upgrades
• Runs logged: 24 (4 tools × 6 samples)
• Not attempted / deferred this round: Copyleaks (likely account required), Originality.ai (free tier allows 3 scans/day — scheduled for a later window), Winston AI (14-day trial signup required), Turnitin (institutional access only — excluded by our methodology)

The test set (v1)

Each sample is ~220 words, sized to stay inside free-detector limits. Provenance for every text:

All three human texts pre-date AI generation entirely. The full text of each sample is archived in our working files alongside this log.

Every run, as logged

"AI %" is the detector's own AI-likelihood figure for that run. A run on a human sample (H1–H3) flagged as AI is a false positive; a run on an AI sample (A1–A3) scored as human is a false negative.

The Darwin rerun (GPTZero, run-to-run instability)

The most important single observation of round 1. On 2026-10-12 an earlier same-day GPTZero run on the identical Darwin excerpt returned the opposite verdict — "highly confident human written" (AI 0% / Human 100%), logged via screenshot at capture time. On 2026-10-13 the same input scored AI 100%. Identical text, opposite outcomes, roughly a day apart. We did not average the two; both are recorded, and our GPTZero row carries the instability note.

Per-tool round summary

• ZeroGPT — AI detected 3/3; false positives 2/3 (Mill 55.1%, Dickens 85.4%; Darwin 0%)
• Sapling — AI detected 3/3; false positives 0/3 (0.1% / 0.6% / 6.3%)
• GPTZero — AI detected 3/3 (100% each); false positives 3/3 (all three human texts scored 100% AI) — the most aggressive scorer of the round, with one observed run-to-run flip (Darwin)
• QuillBot — AI detected 0/3 (8% / 0% / 0%); false positives 0/3 — the most conservative scorer of the round; on this set it cleared the human texts and the AI texts

How the runs were captured

• GPTZero: Basic AI Scan (model 4.1o) via the app.gptzero.me dashboard, 2026-10-13. Run notes record screenshot verification at capture time for A2, A3, H1, H2, and the Darwin rerun flip.
• QuillBot: Free AI Detector at quillbot.com, 2026-10-13; free tier's 6 daily scans were all used by this round. Run notes record screenshot verification at capture time.
• ZeroGPT / Sapling: free web checkers, 2026-10-12; scores logged as displayed.
• Screenshots: referenced in the run notes where captured; the screenshot files themselves are not published in this log — the numbers above are the recorded values.

Fields we did not capture in round 1 (recorded as not recorded, not estimated): exact time of day per run, detector UI/model version strings for ZeroGPT and Sapling, and server-side request logs.

Known limitations (carried from the methodology)

Three AI samples from a single model (GLM); three human samples that are pre-1900 public-domain prose, which skews classical; n=3 per class per tool. Round 1 is a directional spot-check, not a benchmark — which is exactly why the scorecard shows these rows as labeled evidence rather than percentages.

What's next in the log

Round 2 planning adds modern human-written samples and the tools deferred above (Originality.ai first — its free-tier window fits a single round). New rounds append to this page and the data files with their own dates; nothing is retro-edited.