# WP-AI-0002 — Human Text Reconstruction Transformer (HTR)

Question. Can a Transformer reconstruct characteristics of pre-2018 human-authored scientific prose from LLM-rewritten versions of the same passages?

Deliverables. The project, part of UQO Research (series AI, id AI-0002), has two deliverables:

• a peer-review-ready paper, Human Text Reconstruction: Learning the Inverse of LLM-Induced Transformations in Scientific Writing;
• a released model family (HTR-Tiny/Small/Base, plus Large if feasible) that can actually be used.

Prior work. Paraphrase inversion already exists (Rivera Soto et al., 2024, 2026). This project studies something different:

• Reconstruction fidelity against identity and paraphrase baselines.
• Rights-verified, pre-2018 scientific originals with controlled Claude rewrite variants.
• Factual integrity.
• Non-identifiability.
• Memorisation.

See docs/RESEARCH_GAP.md.

# Status (2026-11-09)

• Gate A: GO. 100-paper PMC smoke test with 2 013 eligible passages; licence governance in place (docs/gates/GATE_A.md).
• Pilot corpus. 1 180 PMC documents with 23 658 eligible passages (data/parsed/pilot_pmc). Book trials: 11 Gutenberg books and 17 NCBI Bookshelf chapters. arXiv and Crossref economics/finance adapters are in progress.
• Vertical slice. Runs end to end with mock rewrites at 0 USD spend. M1-nano passes the one-batch overfit test.
• Not started: real Claude rewrites (Gate B, ceiling 30 USD), trained models at release scale, and the paper draft.
• Full status: docs/IMPLEMENTATION_CHECKLIST.md, docs/architecture/RESEARCH_PLAN.md.

# Quick start

python3 -m venv .venv && .venv/bin/pip install -e ".[api,hf,dev]"
.venv/bin/python -m pytest tests -q -m "not network"      # unit tests, no network
scripts/vertical_slice.sh                                  # corpus → mock rewrites → pairs → M1-nano → metrics (0 USD)
.venv/bin/python scripts/discover_pmc.py --candidates 400 --max-docs 100 --seed 1 --out data/raw/pmc_smoke

Using a trained checkpoint:

from htr.inference.pipeline import HTRReconstructor
r = HTRReconstructor.from_pretrained("artifacts/runs/<run>/best").reconstruct(text, {"discipline": "economics"})
print(r.reconstruction, r.warnings)

Spending money. Real API calls require --live and respect the budget ceilings (docs/reproducibility/COMPUTE_BUDGET.md). All commands are listed in docs/reproducibility/REPRODUCIBILITY.md.

# Layout

configs/      rewriting (pilot.json, pricing.json); data/training/evaluation/release (to fill)
src/htr/      ingest · licenses · parse · segment · dedup · data · rewrite · validate · models · training ·
              inference · evaluation · release · utils
scripts/      discover_{pmc,arxiv,books} · build_corpus · make_splits · rewrite · validate_pairs ·
              train_tokenizer · train · predict · evaluate · vertical_slice.sh
tests/        unit · integration (network-marked) · regression
data/         manifests (committed) · raw · parsed · pairs · splits (not committed)
artifacts/    runs/checkpoints (not committed) · metrics · figures · tables · reports
docs/         MANDATE · RESEARCH_GAP · PREREGISTRATION · DECISION_LOG · RISK_REGISTER · IMPLEMENTATION_CHECKLIST ·
              architecture/ · data_governance/ · experiments/ · reproducibility/ · gates/ · literature/
paper/  model_cards/  demo/

# Honest-claims notice

HTR outputs are synthetic reconstructions. They are a model's estimate of pre-2018-style scientific prose given an LLM rewrite. They are not the original text, they do not restore authorship, and they are not evidence that a text was written by a human. HTR is not an AI-text detector. Every result is reported against identity and paraphrase baselines with confidence intervals, and null or negative results are reported.

# Licence

• Code: Apache-2.0 (pyproject.toml).
• Data and derived text: per-document licences. Only CC0, public-domain and CC BY items are redistributable (docs/data_governance/LICENSE_MATRIX.md).
• Model weights: licence to be decided at release. It depends on the training licence set and on Anthropic's terms for model outputs (open question, LICENSE_MATRIX.md §4).
• Attribution: NLM/PMC and all source authors are acknowledged.

# Citation

@techreport{boucher2026htr,
  author      = {Boucher, Simon-Pierre},
  title       = {Human Text Reconstruction: Learning the Inverse of {LLM}-Induced Transformations in Scientific Writing},
  institution = {Universit{\'e} du Qu{\'e}bec en Outaouais},
  type        = {Working Paper},
  number      = {AI-0002},
  year        = {2026},
  note        = {In preparation}
}