Bring LLM-rewritten scientific prose back towards its pre-2018 human register.
Paste a paragraph that a language model has rewritten or polished. HTR proposes a reconstruction in the style of scientific writing before 2018, keeps numbers, citations and equations protected, and flags anything it could not preserve.
Research preview. The output is a synthetic reconstruction, not the original text and not evidence of human authorship. This instance runs the pilot model (HTR‑Tiny, 27 M parameters); larger models are in training.
Input — LLM-rewritten passage
Reconstruction — synthetic
API
One endpoint, JSON in and out. Rate limit 30 requests per minute per client; inputs up to 8 000 characters.
curl -s https://htr.spboucher.ai/api/reconstruct \
-H 'Content-Type: application/json' \
-d '{"text": "…rewritten passage…", "discipline": "economics", "pub_year": 2012, "section_role": "results", "beam": 1}'Response fields: reconstruction, warnings (preservation issues), preservation
(number recall/precision, placeholder exactness, negation and hedging deltas), metadata_used, latency_s,
notice. Model metadata: GET /api/info. Health: GET /health.
About
HTR is the model family of UQO working paper AI‑0002 — Human Text Reconstruction: Learning the Inverse of LLM‑Induced Transformations in Scientific Writing. Training pairs are built from rights‑verified scientific passages published before 2018 (PMC Open Access, CC‑BY economics and finance journals, arXiv, public‑domain reference books) and their rewrites by a large language model; the model learns the inverse map. Reconstruction from a paraphrase is non‑identifiable: many originals share one rewrite, so the output is one plausible estimate, and every number, citation and negation must be checked.
Weights and code: Apache‑2.0 (non‑commercial research use where CC‑BY‑NC sources were included). Not for AI‑text detection, authorship attribution or academic‑integrity decisions. © Simon‑Pierre Boucher, Université du Québec en Outaouais.