{ "generatedAt": "2026-08-26", "corpusSize": 48, "regressionSamples": 11, "humanSamples": 24, "aiSamples": 24, "byRegister": { "esl-nonnative": { "label": "human", "n": 5, "misses": 0, "meanScore": 25 }, "academic-abstract": { "label": "human", "n": 4, "misses": 1, "meanScore": 35.8 }, "government-bureaucratic": { "label": "human", "n": 4, "misses": 0, "meanScore": 19.8 }, "technical-documentation": { "label": "human", "n": 4, "misses": 0, "meanScore": 21.8 }, "casual-first-person": { "label": "human", "n": 4, "misses": 0, "meanScore": 12.8 }, "business-marketing": { "label": "human", "n": 3, "misses": 0, "meanScore": 14.7 }, "default-assistant": { "label": "ai", "n": 7, "misses": 6, "meanScore": 43.3 }, "instructed-casual": { "label": "ai", "n": 5, "misses": 5, "meanScore": 9 }, "instructed-varied-rhythm": { "label": "ai", "n": 5, "misses": 5, "meanScore": 9.4 }, "humanizer-output": { "label": "ai", "n": 3, "misses": 3, "meanScore": 9.7 }, "domain-specific": { "label": "ai", "n": 4, "misses": 4, "meanScore": 19.8 } }, "caveats": { "dateBasis": "Every sample was published between 2010 and 2019, i.e. before any general-purpose LLM was publicly available. Human authorship here is a date fact, not a stylistic judgement. Each sample records how the date was verified, preferring identifiers that cannot be back-dated: DOIs (checked against Crossref), PMC ids and Europe PMC firstPublicationDate, Federal Register document numbers, MediaWiki revision ids, git tags, and Stack Exchange server-assigned creation timestamps.", "eslSourcing": "READ THIS BEFORE PUBLISHING ANY ESL NUMBER. The five \"esl-nonnative\" samples are NOT learner-corpus essays. No learner corpus with a license compatible with commercial use could be sourced: PELIC is CC BY-NC-SA (NC excludes us), the Cambridge Learner Corpus FCE set, ICLE, ICNALE, NUCLE and the BEA-2019 W&I+LOCNESS data are all behind restrictive or registration-gated licenses, and the ETS TOEFL11 corpus is a paid LDC product. WHO and FAO publications were also rejected as CC BY-NC-SA. Rather than substitute something unlicensed, these five are CC BY / CC BY-SA scholarly prose written by authors whose institutional affiliations are all in non-anglophone countries (Indonesia, Turkey, Iran, China), selected because visible L2-transfer features survived peer review uncorrected. That is evidence of non-native authorship, not proof of any individual author's first language, and the register is academic rather than the undergraduate essay that produces our worst false positives. Treat any FPR computed on this slice as a floor, not as the ESL-essay false-positive rate.", "registerVariance": "Registers are unevenly hard to source under an open license. Government and technical-documentation samples come from a small number of institutional voices (three Federal Register rules, two Kubernetes docs pages, three English Wikipedia revisions), so within-register variance understates the real world.", "modelDiversity": "Read this before quoting any FNR computed from this file.\n\nWHICH MODELS ARE ACTUALLY IN HERE (24 samples):\n- claude-opus-5 (Anthropic), generated 2026-08-26 for this benchmark: 18 samples.\n- OpenAI gpt-4o-mini, as the engine of the Codaone production humanizer, rewriting claude-opus-5 text: 3 samples. Mixed lineage — Claude wrote it, GPT rewrote it.\n- OpenAI ChatGPT, December 2022 web release (gpt-3.5 era), via the HC3 dataset: 3 samples.\n\nWHICH ARE NOT: Google Gemini. Meta Llama. Mistral. DeepSeek. Qwen. xAI Grok. Cohere. Every current-generation OpenAI model (GPT-4o, GPT-5 family) except gpt-4o-mini in its humanizer role. Every current-generation Claude except opus-5. Nothing from a consumer \"undetectable AI\" service (StealthGPT, Undetectable.ai, Phrasly) — those are the strongest evasion tools in the wild and we have zero samples of their output.\n\nTHEREFORE: an FNR computed over this corpus is an FNR *for Claude Opus 5 output, for our own humanizer's output, and for three-year-old ChatGPT output*. It is not an estimate of how often we miss AI text in general, and it must not be published as one. If the number is quoted on /ai-detector/accuracy, this sentence has to travel with it.\n\nA SECOND BIAS, LESS OBVIOUS: 18 of the samples were written by the same model that assembled this corpus, while it knew it was building a detector benchmark. That is a demand-characteristic risk in both directions — the model may have written more stereotypically \"AI\" prose for the default-register samples, and may have tried harder than a naive user would on the evasion samples. The prompts are recorded verbatim so a third party can rerun them on any model and check. Doing that on a non-Claude model is the cheapest available fix for both biases.\n\nTHIRD, THE REGISTER MIX IS NOT A USAGE DISTRIBUTION: {\"default-assistant\":7,\"instructed-casual\":5,\"instructed-varied-rhythm\":5,\"humanizer-output\":3,\"domain-specific\":4}. It is deliberately weighted toward evasion (13 of 24 samples are instructed-casual, instructed-varied-rhythm, or humanizer output) because §1.4 of AI-DETECTOR-COMPETITIVE-2026-08-25 measured that all of our false negatives come from disguised AI. A corpus weighted this way will report a WORSE FNR than a corpus of default-register text would. That is intentional and honest; it is not comparable to a competitor's headline accuracy number, which is measured on whatever mix flatters them.\n\nFOURTH, ON THE HUMANIZER SAMPLES: the free anonymous quota is 3/day/IP and all three were used on 2026-08-26, so there are exactly three and no reruns. They are the highest-value rows in the file — a false negative on any of them means our own paid product defeats our own free product, which is the central question the accuracy page exists to answer honestly." }, "falsePositives": 1, "falseNegatives": 23, "threshold": 50, "fpr": 0.041666666666666664, "fnr": 0.9583333333333334, "humanMean": 22.041666666666668, "aiMean": 20.958333333333332, "separation": -1.0833333333333357, "results": [ { "label": "human", "note": "esl-nonnative: Indonesian authors, education research abstract. L1-transfer markers: \"The subject of this research are\", \"the interaction that occurs in students\", \"This interaction is built with heterogeneous student skills.\"", "provenance": "Katarina Tri Utaminingtyas, Rachmadina Eka Herdianti, Inti Hayatul Fitria & Anton Prayitno, \"Small Groups: Student Productive Interactions in Learning Cooperative (Case Study of Mathematics Learning at Junior High School in Pakis, Malang)\", Educational Process: International Journal 6(2), 2017 — https://doi.org/10.22521/edupij.2017.62.3", "register": "esl-nonnative", "sourced": true, "score": 24, "verdict": "human", "words": 93 }, { "label": "human", "note": "esl-nonnative: Turkish nursing academics, structured abstract. L2 markers: \"The scale was carried out to the students\", \"While analyzing the research data;\", \"Fisher' Exact\".", "provenance": "Ebru Özen Bekar, Dilek Konuk Şener, Çetin Yılmaz & Şengül Cangür, \"The Evaluation of Professional Self-esteem of Nurses and Social Workers Before and After Graduation\", Sağlık ve Hemşirelik Yönetimi Dergisi (Journal of Health and Nursing Management) 4(3), 2017 — https://doi.org/10.5222/SHYD.2017.050", "register": "esl-nonnative", "sourced": true, "score": 15, "verdict": "human", "words": 119 }, { "label": "human", "note": "esl-nonnative: Iranian-Persian L1 authors at Turkish/Iranian institutions; discursive, argumentative register — the closest thing in this set to a student essay.", "provenance": "Sahar Pouya & Homa Irani Behbahani, \"Landscape visual assessment: A case of Iran-Iraq war memorial garden\", Turkish Journal of Forestry / Türkiye Ormancılık Dergisi 18(4), 2017 — https://doi.org/10.18182/tjf.294916", "register": "esl-nonnative", "sourced": true, "score": 35, "verdict": "human", "words": 136 }, { "label": "human", "note": "esl-nonnative: Iranian clinical-trial abstract. Article omission throughout (\"in treatment of patients\", \"in case they had\"), a classic L1-Persian marker.", "provenance": "Shakiba M, Moazen-Zadeh E, Noorbala AA, Jafarinia M, Divsalar P, Kashani L, \"Saffron (Crocus sativus) versus duloxetine for treatment of patients with fibromyalgia: A randomized double-blind clinical trial\", Avicenna Journal of Phytomedicine, 2018 — https://europepmc.org/article/PMC/PMC6235666", "register": "esl-nonnative", "sourced": true, "score": 9, "verdict": "human", "words": 109 }, { "label": "human", "note": "esl-nonnative: Chinese-hospital author team, medical case series. Semicolon-splice and \"which may demand an additional approach to the ongoing practice\" are non-native constructions that survived peer review.", "provenance": "Pandey S, Li L, Deng XY, Cui DM, Gao L, \"Outcome Following the Treatment of Ventriculitis Caused by Multi/Extensive Drug Resistance Gram Negative Bacilli; Acinetobacter baumannii and Klebsiella pneumonia\", Frontiers in Neurology 9:1174, 2019 — https://doi.org/10.3389/fneur.2018.01174", "register": "esl-nonnative", "sourced": true, "score": 42, "verdict": "human", "words": 142 }, { "label": "human", "note": "academic-abstract: Review abstract, UK. Uniform long sentences, heavy formal connectors (\"Furthermore\", \"Overall\") — the canonical AI-lookalike register.", "provenance": "Barton AJ, Hill J, Pollard AJ, Blohmke CJ, \"Transcriptomics in Human Challenge Models\", Frontiers in Immunology 8:1839, 2017 (Oxford Vaccine Group, University of Oxford) — https://doi.org/10.3389/fimmu.2017.01839", "register": "academic-abstract", "sourced": true, "score": 42, "verdict": "human", "words": 149 }, { "label": "human", "note": "academic-abstract: Structured systematic-review abstract with labelled sections. Extremely uniform sentence length and near-zero first-person voice.", "provenance": "Cairns AE, Pealing L, Duffy JMN, Roberts N, Tucker KL, Leeson P, \"Postpartum management of hypertensive disorders of pregnancy: a systematic review\", BMJ Open 7(11):e018696, 2017 (University of Oxford) — https://doi.org/10.1136/bmjopen-2017-018696", "register": "academic-abstract", "sourced": true, "score": 9, "verdict": "human", "words": 95 }, { "label": "human", "note": "academic-abstract: Plant-genetics abstract, US. Dense nominalisation and hedged claims; the kind of prose humans write that scores as \"too uniform\".", "provenance": "PLOS ONE 13(12):e0207723, 2018 — MSU-DOE Plant Research Laboratory, Michigan State University — https://doi.org/10.1371/journal.pone.0207723", "register": "academic-abstract", "sourced": true, "score": 42, "verdict": "human", "words": 95 }, { "label": "human", "note": "academic-abstract: Statistical-ecology abstract, US. Long subordinate clauses, \"However\"/\"In these settings\" transitions.", "provenance": "PLOS ONE 13(12):e0204150, 2018 — Department of Forestry, Michigan State University — https://doi.org/10.1371/journal.pone.0204150", "register": "academic-abstract", "sourced": true, "score": 50, "verdict": "ai", "words": 108 }, { "label": "human", "note": "government-bureaucratic: FAA final rule summary. \"Additionally\", \"Finally\", \"These actions are necessary to\" — the exact connector profile our detector punishes.", "provenance": "Federal Aviation Administration, \"Regulatory Relief: Aviation Training Devices; Pilot Certification, Training, and Pilot Schools; and Other Provisions\", final rule, Federal Register document 2018-12800 — https://www.federalregister.gov/documents/2018/06/27/2018-12800", "register": "government-bureaucratic", "sourced": true, "score": 30, "verdict": "human", "words": 159 }, { "label": "human", "note": "government-bureaucratic: FDA final rule summary — one 100-word sentence built out of semicolon-separated clauses. Human, and maximally machine-like.", "provenance": "Food and Drug Administration, \"Food Labeling: Revision of the Nutrition and Supplement Facts Labels\", final rule, Federal Register document 2016-11867 — https://www.federalregister.gov/documents/2016/05/27/2016-11867", "register": "government-bureaucratic", "sourced": true, "score": 21, "verdict": "human", "words": 162 }, { "label": "human", "note": "government-bureaucratic: Department of Education interim final rule. Statutory cross-references and self-referential procedural language.", "provenance": "Department of Education, \"Student Assistance General Provisions, Federal Perkins Loan Program, Federal Family Education Loan Program, William D. Ford Federal Direct Loan Program, and Teacher Education Assistance for College and Higher Education Grant Program\", interim final rule, Federal Register document 2017-22851 — https://www.federalregister.gov/documents/2017/10/24/2017-22851", "register": "government-bureaucratic", "sourced": true, "score": 15, "verdict": "human", "words": 133 }, { "label": "human", "note": "government-bureaucratic: CDC surveillance report opening. Numbered citations, passive voice, agency-as-actor (\"CDC analyzed data from...\").", "provenance": "Cree RA, Bitsko RH, Robinson LR, Holbrook JR, Danielson ML, Smith C, et al., \"Health Care, Family, and Community Factors Associated with Mental, Behavioral, and Developmental Disorders and Poverty Among Children Aged 2-8 Years — United States, 2016\", MMWR Morbidity and Mortality Weekly Report 67(50), 2018 — https://europepmc.org/article/PMC/PMC6342550", "register": "government-bureaucratic", "sourced": true, "score": 13, "verdict": "human", "words": 100 }, { "label": "human", "note": "technical-documentation: Encyclopedic network-protocol description. Flat declaratives, list-like enumeration, zero first person.", "provenance": "Wikipedia contributors, \"Transmission Control Protocol\", English Wikipedia, revision 843439827 (2018-05-29T05:04:10Z) — https://en.wikipedia.org/w/index.php?title=Transmission_Control_Protocol&oldid=843439827", "register": "technical-documentation", "sourced": true, "score": 15, "verdict": "human", "words": 110 }, { "label": "human", "note": "technical-documentation: Cryptography explainer. Long conditional sentences and \"For this to work it must be\" constructions.", "provenance": "Wikipedia contributors, \"Public-key cryptography\", English Wikipedia, revision 736786168 (2016-08-29T20:54:56Z) — https://en.wikipedia.org/w/index.php?title=Public-key_cryptography&oldid=736786168", "register": "technical-documentation", "sourced": true, "score": 42, "verdict": "human", "words": 106 }, { "label": "human", "note": "technical-documentation: Product documentation, procedural voice. Repetitive parallel clauses (\"does not kill... does not create...\") read as templated.", "provenance": "Kubernetes documentation, \"Deployments\" (content/en/docs/concepts/workloads/controllers/deployment.md), kubernetes/website at tag release-1.16 — https://github.com/kubernetes/website/blob/release-1.16/content/en/docs/concepts/workloads/controllers/deployment.md", "register": "technical-documentation", "sourced": true, "score": 18, "verdict": "human", "words": 81 }, { "label": "human", "note": "technical-documentation: Networking internals documentation. Comparative technical claims with hedging, written by contributors for whom English varies.", "provenance": "Kubernetes documentation, \"Service\" (content/en/docs/concepts/services-networking/service.md), kubernetes/website at tag release-1.16 — https://github.com/kubernetes/website/blob/release-1.16/content/en/docs/concepts/services-networking/service.md", "register": "technical-documentation", "sourced": true, "score": 12, "verdict": "human", "words": 72 }, { "label": "human", "note": "casual-first-person: Blunt forum answer. Sentence fragments, an em dash, a typo (\"polystychrene\"), and a joke — high burstiness.", "provenance": "Stack Exchange user \"Criggie\", answer 51225 on Bicycles Stack Exchange, 2017-12-01 — https://bicycles.stackexchange.com/a/51225", "register": "casual-first-person", "sourced": true, "score": 16, "verdict": "human", "words": 78 }, { "label": "human", "note": "casual-first-person: Advice post with imperatives, a parenthetical aside and an editorialising last line. Wildly uneven sentence lengths.", "provenance": "Stack Exchange user \"keshlam\", answer 82857 on The Workplace Stack Exchange, 2017-01-12 — https://workplace.stackexchange.com/a/82857", "register": "casual-first-person", "sourced": true, "score": 9, "verdict": "human", "words": 80 }, { "label": "human", "note": "casual-first-person: Reflective first-person explanation of social convention, with quoted dialogue and a self-deprecating closing parenthesis.", "provenance": "Stack Exchange user \"Max\", answer 38180 on Travel Stack Exchange, 2014-11-03 — https://travel.stackexchange.com/a/38180", "register": "casual-first-person", "sourced": true, "score": 12, "verdict": "human", "words": 153 }, { "label": "human", "note": "casual-first-person: Wikipedia talk-page comment. Enumerated grievance, slang (\"crabon\"), signed and timestamped by the editor.", "provenance": "Wikipedia editor Dennis Bratland, comment dated 22:30, 20 July 2014 (UTC) on Talk:Bicycle; captured in English Wikipedia revision 664757797 (2015-05-30T21:01:42Z) — https://en.wikipedia.org/w/index.php?title=Talk%3ABicycle&oldid=664757797", "register": "casual-first-person", "sourced": true, "score": 14, "verdict": "human", "words": 139 }, { "label": "human", "note": "business-marketing: Enforcement press release. Announcement-lede structure, superlatives (\"record\", \"by far the largest\"), third-person institutional voice.", "provenance": "Federal Trade Commission, press release \"Google and YouTube Will Pay Record $170 Million for Alleged Violations of Children's Privacy Law\", September 4, 2019 — https://www.ftc.gov/news-events/news/press-releases/2019/09/google-youtube-will-pay-record-170-million-alleged-violations-childrens-privacy-law", "register": "business-marketing", "sourced": true, "score": 17, "verdict": "human", "words": 99 }, { "label": "human", "note": "business-marketing: Press release with an executive quote — the register that reads most like generated corporate copy.", "provenance": "Federal Trade Commission, press release \"Uber Agrees to Expanded Settlement with FTC Related to Privacy, Security Claims\", April 12, 2018 (quote from Acting FTC Chairman Maureen K. Ohlhausen) — https://www.ftc.gov/news-events/news/press-releases/2018/04/uber-agrees-expanded-settlement-ftc-related-privacy-security-claims", "register": "business-marketing", "sourced": true, "score": 16, "verdict": "human", "words": 155 }, { "label": "human", "note": "business-marketing: Product release announcement. Feature-benefit sentences and capability claims — vendor marketing prose written by engineers.", "provenance": "Kubernetes 1.16 Release Team, \"Kubernetes 1.16: Custom Resources, Overhauled Metrics, and Volume Extensions\", Kubernetes blog, 2019-09-18 — https://kubernetes.io/blog/2019/09/18/kubernetes-1-16-release-announcement/", "register": "business-marketing", "sourced": true, "score": 11, "verdict": "human", "words": 86 }, { "label": "ai", "note": "default-assistant: Classic essay register: abstract subject, \"Furthermore\", uniform long sentences. The easy case — if we miss this we have nothing. Also the input to humanizer sample #1.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "default-assistant", "sourced": true, "score": 42, "verdict": "human", "words": 120 }, { "label": "ai", "note": "default-assistant: Default register with one em-dash. Included deliberately: commit ba4f0aef demoted em-dashes from conviction evidence to corroboration, and this sample is the regression guard for that decision on the AI side.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "default-assistant", "sourced": true, "score": 23, "verdict": "human", "words": 102 }, { "label": "ai", "note": "default-assistant: SEO-blog intro register — the highest-volume real-world use of an LLM and the text most likely to be pasted into a detector by an editor.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "default-assistant", "sourced": true, "score": 70, "verdict": "ai", "words": 110 }, { "label": "ai", "note": "default-assistant: Default register, informational/advisory. Also the input to humanizer sample #2, so the pair isolates what the humanizer actually changes.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "default-assistant", "sourced": true, "score": 42, "verdict": "human", "words": 114 }, { "label": "ai", "note": "instructed-casual: Texting register: lowercase, no terminal punctuation on the last line, fragments. Direct replication of the §1.4 miss that scored 15.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "instructed-casual", "sourced": true, "score": 9, "verdict": "human", "words": 83 }, { "label": "ai", "note": "instructed-casual: Reddit-comment imitation — the §1.4 miss that scored 9, and the single cheapest evasion a real user can perform (one sentence of prompt).", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "instructed-casual", "sourced": true, "score": 9, "verdict": "human", "words": 121 }, { "label": "ai", "note": "instructed-casual: Fake user review in a chatty voice — the commercial evasion case (review farms), and a register where sentence-capitalization is preserved so the detector cannot key on lowercase alone.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "instructed-casual", "sourced": true, "score": 9, "verdict": "human", "words": 108 }, { "label": "ai", "note": "instructed-casual: Self-deprecating personal confession — tests whether emotional first-person content plus low lexical formality is enough to push the score under threshold.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "instructed-casual", "sourced": true, "score": 9, "verdict": "human", "words": 93 }, { "label": "ai", "note": "instructed-casual: Personal-blog voice with the transition words explicitly banned in the prompt — isolates how much of our AI signal is carried by \"moreover / additionally\" alone.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "instructed-casual", "sourced": true, "score": 9, "verdict": "human", "words": 99 }, { "label": "ai", "note": "instructed-varied-rhythm: Extreme length variance (2-word sentences next to 45-word sentences). This is the direct attack on a CV-of-sentence-length feature.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "instructed-varied-rhythm", "sourced": true, "score": 9, "verdict": "human", "words": 118 }, { "label": "ai", "note": "instructed-varied-rhythm: Varied rhythm carrying an anecdote with reported speech — the register a student actually gets when they ask for \"a personal essay that does not sound like AI\".", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "instructed-varied-rhythm", "sourced": true, "score": 9, "verdict": "human", "words": 118 }, { "label": "ai", "note": "instructed-varied-rhythm: Argumentative op-ed with varied rhythm and a concrete verifiable claim (Buffalo 2017). Tests whether specific facts and dates read as human to us.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "instructed-varied-rhythm", "sourced": true, "score": 9, "verdict": "human", "words": 132 }, { "label": "ai", "note": "instructed-varied-rhythm: Literary reflective memoir with varied rhythm — sits directly on top of the human Woolf/Fitzgerald samples in the human half. If this scores like they do, the score is not separating authorship, it is separating genre.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "instructed-varied-rhythm", "sourced": true, "score": 11, "verdict": "human", "words": 128 }, { "label": "ai", "note": "instructed-varied-rhythm: Technical explainer written with varied rhythm and a second-person analogy — the hardest combination for us, because it is also exactly how a good human technical writer writes.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "instructed-varied-rhythm", "sourced": true, "score": 9, "verdict": "human", "words": 115 }, { "label": "ai", "note": "humanizer-output: Our production humanizer applied to default-assistant sample #1 (remote work). API metrics on the call: aiScoreBefore 24, aiScoreAfterPass1 26, aiScoreAfter 8, iterations 2, 120→141 words. Note the humanizer's own estimator already scored the untouched Claude essay at only 24 — that estimator disagreeing with detect.js is its own finding.", "provenance": "Two-stage pipeline. Stage 1: claude-opus-5 (Anthropic) produced the input text (default-assistant sample #1, remote-work essay) on 2026-08-26. Stage 2: that text was POSTed verbatim to the Codaone production humanizer (POST https://www.codaone.ai/api/tools/humanize, Origin: https://www.codaone.ai, body {text, mode:\"standard\"}, anonymous/free plan) on 2026-08-26. The API reported model \"gpt-4o-mini\" (OpenAI), fallback:false. The value of the response's \"humanized\" field is stored below verbatim, including its paragraph breaks. — https://www.codaone.ai/api/tools/humanize", "register": "humanizer-output", "sourced": true, "score": 9, "verdict": "human", "words": 141 }, { "label": "ai", "note": "humanizer-output: Our production humanizer applied to default-assistant sample #4 (small-business cybersecurity). API metrics: aiScoreBefore 24, aiScoreAfterPass1 18, aiScoreAfter 14, iterations 2, 114→152 words.", "provenance": "Two-stage pipeline. Stage 1: claude-opus-5 (Anthropic) produced the input text (default-assistant sample #4, small-business cybersecurity paragraph) on 2026-08-26. Stage 2: that text was POSTed verbatim to the Codaone production humanizer (POST https://www.codaone.ai/api/tools/humanize, Origin: https://www.codaone.ai, body {text, mode:\"standard\"}, anonymous/free plan) on 2026-08-26. The API reported model \"gpt-4o-mini\" (OpenAI), fallback:false. The value of the response's \"humanized\" field is stored below verbatim, including its paragraph breaks. — https://www.codaone.ai/api/tools/humanize", "register": "humanizer-output", "sourced": true, "score": 9, "verdict": "human", "words": 152 }, { "label": "ai", "note": "humanizer-output: Our production humanizer applied to a formal academic abstract (domain-specific sample #1, lightly shortened to fit the free word budget). API metrics: aiScoreBefore 16, aiScoreAfter 20, iterations 1, 123→141 words. This is the one call where our estimator scored the output HIGHER than the input (16→20) — the humanizer made formal text read more like generic AI, not less.", "provenance": "Two-stage pipeline. Stage 1: claude-opus-5 (Anthropic) produced the input text (domain-specific sample #1, an academic abstract, with the \"(n = 4,812)\" parenthetical and the final clause about student effort removed so the input fit the 300-word free budget cleanly) on 2026-08-26. Stage 2: that text was POSTed verbatim to the Codaone production humanizer (POST https://www.codaone.ai/api/tools/humanize, Origin: https://www.codaone.ai, body {text, mode:\"standard\"}, anonymous/free plan) on 2026-08-26. The API reported model \"gpt-4o-mini\" (OpenAI), fallback:false. The value of the response's \"humanized\" field is stored below verbatim, including its paragraph breaks. — https://www.codaone.ai/api/tools/humanize", "register": "humanizer-output", "sourced": true, "score": 11, "verdict": "human", "words": 141 }, { "label": "ai", "note": "domain-specific: AI-written empirical abstract. Pairs with the false-positive side: if human academic prose scores high AND this scores high, the feature is \"academic\", not \"AI\". Also the stage-1 input for humanizer sample #3.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "domain-specific", "sourced": true, "score": 9, "verdict": "human", "words": 136 }, { "label": "ai", "note": "domain-specific: AI-written contract-law analysis. Direct counterpart to the human Marbury v. Madison sample in the human half — same register, opposite label.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "domain-specific", "sourced": true, "score": 42, "verdict": "human", "words": 134 }, { "label": "ai", "note": "domain-specific: AI-written API reference. Counterpart to the human Wright brothers patent specification in the human half: both are uniform procedural prose with near-zero sentence-length variance, which is precisely where a variance-based score has no information.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "domain-specific", "sourced": true, "score": 14, "verdict": "human", "words": 131 }, { "label": "ai", "note": "domain-specific: AI-written internal corporate memo. The register a real employee most plausibly delegates to an LLM, and the register a manager most plausibly runs through a detector.", "provenance": "claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.", "register": "domain-specific", "sourced": true, "score": 14, "verdict": "human", "words": 124 }, { "label": "ai", "note": "default-assistant: CROSS-FAMILY CONTROL (OpenAI, not Claude). Encyclopedic explainer register. Three years older than our own samples, so it also probes whether we only detect current-generation phrasing.", "provenance": "OpenAI ChatGPT, December 2022 web release (gpt-3.5 era; exact checkpoint not disclosed by OpenAI). Collected by Guo et al., \"How Close is ChatGPT to Human Experts?\" (arXiv:2301.07597) into the HC3 dataset, config \"wiki_csai\", row_idx 11 (record id 11), field chatgpt_answers[0]. Retrieved 2026-08-26 via the Hugging Face datasets-server rows API. — https://huggingface.co/datasets/Hello-SimpleAI/HC3", "register": "default-assistant", "sourced": true, "score": 42, "verdict": "human", "words": 149 }, { "label": "ai", "note": "default-assistant: CROSS-FAMILY CONTROL (OpenAI, not Claude). Financial explainer, domain register. Contains a real generation artifact — a missing space at \"potential investors.Stock splits\" — which is preserved verbatim rather than cleaned up.", "provenance": "OpenAI ChatGPT, December 2022 web release (gpt-3.5 era; exact checkpoint not disclosed by OpenAI). Collected by Guo et al., \"How Close is ChatGPT to Human Experts?\" (arXiv:2301.07597) into the HC3 dataset, config \"finance\", row_idx 15 (record id 15), field chatgpt_answers[0]. Retrieved 2026-08-26 via the Hugging Face datasets-server rows API. — https://huggingface.co/datasets/Hello-SimpleAI/HC3", "register": "default-assistant", "sourced": true, "score": 42, "verdict": "human", "words": 172 }, { "label": "ai", "note": "default-assistant: CROSS-FAMILY CONTROL (OpenAI, not Claude). ELI5 prompt, but note the answer is still in default assistant register — the casual PROMPT did not produce a casual REGISTER. That contrast with our instructed-casual block is the point of including it.", "provenance": "OpenAI ChatGPT, December 2022 web release (gpt-3.5 era; exact checkpoint not disclosed by OpenAI). Collected by Guo et al., \"How Close is ChatGPT to Human Experts?\" (arXiv:2301.07597) into the HC3 dataset, config \"reddit_eli5\", row_idx 10 (record id 10), field chatgpt_answers[0]. Retrieved 2026-08-26 via the Hugging Face datasets-server rows API. — https://huggingface.co/datasets/Hello-SimpleAI/HC3", "register": "default-assistant", "sourced": true, "score": 42, "verdict": "human", "words": 170 } ], "regressionResults": [ { "label": "human", "note": "formal scientific prose, 19th c.", "provenance": "Charles Darwin, On the Origin of Species (1859), Introduction. Public domain. gutenberg.org/ebooks/2009", "register": null, "sourced": false, "score": 12, "verdict": "human", "words": 76 }, { "label": "human", "note": "legal reasoning, formal connectors", "provenance": "Chief Justice John Marshall, Marbury v. Madison, 5 U.S. 137 (1803). US government edict, public domain.", "register": null, "sourced": false, "score": 10, "verdict": "human", "words": 107 }, { "label": "human", "note": "technical specification, uniform procedural register", "provenance": "Orville & Wilbur Wright, US Patent 821,393 \"Flying-Machine\" (1906), specification. Public domain.", "register": null, "sourced": false, "score": 12, "verdict": "human", "words": 82 }, { "label": "human", "note": "casual humorous first person", "provenance": "Mark Twain, The Innocents Abroad (1869), ch. III. Public domain. gutenberg.org/ebooks/3176", "register": null, "sourced": false, "score": 12, "verdict": "human", "words": 69 }, { "label": "human", "note": "literary prose, heavy punctuation (the trap case)", "provenance": "Virginia Woolf, The Voyage Out (1915), ch. I. Public domain in the US. gutenberg.org/ebooks/144", "register": null, "sourced": false, "score": 9, "verdict": "human", "words": 54 }, { "label": "human", "note": "reflective first-person narrative", "provenance": "F. Scott Fitzgerald, The Great Gatsby (1925), ch. I. Public domain in the US since 2021. gutenberg.org/ebooks/64317", "register": null, "sourced": false, "score": 12, "verdict": "human", "words": 75 }, { "label": "ai", "note": "classic LLM essay register", "provenance": "Generated by Claude Opus 5 (Anthropic), 2026-08-25, prompted for default essay register.", "register": null, "sourced": false, "score": 78, "verdict": "ai", "words": 83 }, { "label": "ai", "note": "LLM with heavy em-dash usage", "provenance": "Generated by Claude Opus 5 (Anthropic), 2026-08-25, prompted for em-dash-heavy commentary.", "register": null, "sourced": false, "score": 80, "verdict": "ai", "words": 80 }, { "label": "ai", "note": "LLM listicle body", "provenance": "Generated by Claude Opus 5 (Anthropic), 2026-08-25, prompted for listicle register.", "register": null, "sourced": false, "score": 50, "verdict": "ai", "words": 73 }, { "label": "ai", "note": "LLM instructed to write casually (the evasion case)", "provenance": "Generated by Claude Opus 5 (Anthropic), 2026-08-25, prompted to evade detection with casual register.", "register": null, "sourced": false, "score": 9, "verdict": "human", "words": 77 }, { "label": "ai", "note": "LLM with deliberately varied sentence length", "provenance": "Generated by Claude Opus 5 (Anthropic), 2026-08-25, prompted to vary sentence rhythm.", "register": null, "sourced": false, "score": 9, "verdict": "human", "words": 78 } ] }