spb/wpai0002_uqo

1% ============================================================================2% Ethics and licensing3% ============================================================================4\providecommand{\evid}[1]{\texttt{\detokenize{#1}}}56\section{Ethics and Licensing}7\label{sec:ethics}89\paragraph{Rights in the source texts.} Open access is not open licensing. Every document carries its own licence10code and evidence, the document's statement prevails over provider metadata, and documents whose licence is11unknown, text-mining-only, no-derivatives, or the default arXiv licence are excluded from training12(Table~\ref{tab:licence-tiers}). Rewrites and reconstructions are derivative works and inherit the terms of their13source. Non-commercial and share-alike sources may be used for research training under the default policy but are14never redistributed in the main benchmark. Public-domain status of Project Gutenberg books is verified for the15United States only, and book dates inferred from contributor death years are flagged as upper bounds and can be16excluded. Retraction status is checked at acquisition for PMC articles and will be re-checked before any release.17Acquisition uses official bulk channels, honours \texttt{robots.txt} and rate limits, and identifies itself with a18research user agent; publisher text-mining links are used only for Creative Commons content. Where a publisher's19\texttt{robots.txt} names AI crawlers specifically, the policy to follow is an open decision for the author and is20recorded in the project's risk register.2122\paragraph{Provider terms.} Rewrites are generated with Anthropic models under the provider's commercial terms. The23provider's usage policy restricts the use of outputs to develop competing models. Whether training a small,24task-specific reconstruction model on Claude rewrites and publishing it falls under this restriction is treated as25an open question, not a conclusion: it will be checked against the terms in force and with UQO before any public26weight release. The author took that decision on 2026-11-11 (decision log D23): the six served models are published27on the Hugging Face Hub under the \texttt{spboucher04} namespace with the model cards described below, the28fine-tuned weights are licensed CC~BY-NC~4.0 (non-commercial research use), and the rewrites themselves are kept29internal and are not redistributed; a review of the provider-terms question with UQO remains advisable. Cross-provider generalisation tests (RQ4) use only outputs whose terms permit this use; in this30version all held-out models are from the same provider. The pretrained backbones of Section~\ref{sec:htr2} are31used under their own licences as stated on their model cards at the time of download (Apache-2.0 for flan-T5 and32Qwen2.5-1.5B; the Qwen Research Licence for Qwen2.5-3B); their fine-tuned derivatives inherit both those terms and33the open question on the rewrites, and are marked for non-commercial research use because the training pairs34include CC~BY-NC sources.3536\paragraph{Release.} Six models are served by the public research platform (\url{https://htr.spboucher.ai}):37HTR-Tiny~0.3.0 ($M^\ast$), HTR-Base~0.1.0, flan-T5-large-HTR, flan-T5-xl-HTR, Qwen2.5-1.5B-HTR (the current default)38and Qwen2.5-3B-HTR, each with a model card, the preservation warnings and the detector-sensitivity notice; the39choice of the default model is an author decision that this paper's results inform but do not settle. The40release bundles (SafeTensors weights, tokenizer, configuration, manifest with SHA-256 digests, licence note) load41in a clean interpreter and are published on the Hugging Face Hub (\texttt{spboucher04/htr-tiny}, \texttt{htr-base},42\texttt{flan-t5-large-htr}, \texttt{flan-t5-xl-htr}, \texttt{qwen2.5-1.5b-htr}, \texttt{qwen2.5-3b-htr}; the two HTR43repositories ship the \texttt{htr} inference package). These are \emph{research weights}: they are trained on the full44training set, including non-commercial and share-alike sources, and are therefore released under CC~BY-NC~4.045(Qwen2.5-3B-HTR additionally inherits the Qwen Research Licence), with the synthetic-output notice, the46no-detection and no-integrity-use clause, the provenance of the training pairs and the detector-sensitivity caveat47on every card. A \emph{release-weight} set trained only on CC0, public-domain and CC-BY data remains planned. The48benchmark redistributes only source and derivative text whose licence is in the release set, with an attribution49file listing title, authors, venue, year, identifier, licence and the changes made, plus the NLM/PMC50acknowledgement. No further upload is made without the author's explicit approval.51Takedown requests are handled within five working days by withdrawing the document from all release bundles and52recording the action; if a withdrawn document was in a released model's training data, the next model card lists53it, and retraining is considered if memorisation tests show verbatim recall.\footnote{\evid{docs/data_governance/LICENSE_MATRIX.md}, Sections 5--6.}5455\paragraph{Misuse and dual use.} A model that maps LLM-rewritten text toward the style of pre-2018 human prose can56be misused in two ways: to present generated text as human-authored, and to ``humanise'' generated text so that it57evades AI-text detectors, the class of attack studied by \citet{zheng2025thbench}, \citet{sadasivan2023reliably}58and, with learned rewards, by HUMPA \citep{wang2025humpa}, AuthorMist \citep{david2025authormist} and adversarial59paraphrasing \citep{cheng2025adversarial}. The exploratory tracks of Section~\ref{sec:htr2} sharpen the risk: the60original-versus-rewrite discriminator (AUROC 0.994) is exactly the reward such attacks optimise, which is one61reason the discriminator-reward variant was not run and no reported system uses it; and the flan-T5 backbones62remove every Pangram flag in the sample while the Qwen backbones remove most.63We take three measures. Every output carries the notice that it is a synthetic reconstruction and not evidence of64human authorship (Section~\ref{sec:models}). The model card states the intended research uses and excludes65authorship claims and detector evasion. And we make no detection or authenticity claims ourselves: probe scores are66exploratory, and we do not report detector-evasion rates as a success measure. The residual risk is reduced, not67removed, by these measures; it is weighed against the scientific value of measuring what LLM rewriting destroys.6869\paragraph{Privacy and memorisation.} The corpus consists of published scientific and scholarly text; we do not70collect personal data beyond bibliographic metadata (author names, affiliations as printed). Because a71reconstruction model might regurgitate training originals, the memorisation audit of Section~\ref{sec:design} is a72release condition.7374\paragraph{Human participants.} No human participants have been involved. The qualitative audit of rewrite pairs75reported in Section~\ref{sec:rewrite} was not a study with human raters. No ethics application has been filed;76the planned human evaluation (Section~\ref{sec:human}) will begin only after recruitment and ethics approval at77UQO, and is future work.7879\paragraph{Cost and environment.} Total spend for this version is about 258~USD: 43.20~USD of rewriting API calls80(26.85 Opus pilot, 1.43 and 1.40 for the held-out sets, 13.52 for the 120\,000 Medium Haiku requests), recorded81per job in the ledgers; about 57.17~USD of Pangram detector calls (38.93 for the Gate~E runs and 18.24 for the82HTR~2.0 systems, list-price estimates); and about 158~USD of cloud GPU time (three Lambda H100 instances,832026-11-10, author-approved under D22; to be confirmed against the invoice). Training compute and84accelerator-hours are reported per run (Section~\ref{sec:robustness}); energy was not measured. All large spends85and long training runs required the author's explicit approval.8687\paragraph{Use of AI assistance.} The software, documentation and drafts of this working paper were produced with88the assistance of Claude Code under the author's supervision. All numbers are traced to run manifests and ledgers,89and all references were checked against bibliographic records.90