# CLAUDE.md — WP-AI-0002: Human Text Reconstruction Transformer (HTR)

Can a Transformer reconstruct characteristics of pre-2018 human-authored scientific prose from LLM-rewritten passages? Research home: UQO Research (Recherche-UQO/ia/wp-ai-0002). Language: English for code, docs, paper. Identifier: AI-0002 (series "AI", after AI-0001). Status: see docs/STATUS.md.

The author's full mandate (mission, principles, RQ1–RQ6, phases 0–8, gates A–F, definition of done) is reproduced verbatim in docs/MANDATE.md and binds every session. Author additions (2026-11-09): the corpus must be large (target the "Medium" stage, ≥ 10 000 documents, ≥ 200 000 eligible passages), must include economics and finance articles and English reference books published before 2018, and the released model must be genuinely usable afterwards (model release is a first-class deliverable, not a by-product). PyTorch may be installed on other MacLustr nodes if training needs them.

Operational rules that follow from the mandate:

1. Never fabricate data, metrics, sources, licences or checkpoints. Every number in docs/, artifacts/ and paper/ traces to a run id in artifacts/runs/<run>/manifest.json (hashes, seeds, versions, costs).
2. Rights first. Open access ≠ open licence. htr.licenses.policy is the single source of truth; a document is trained on only if its canonical licence code is in the configured training set, and released only if in the release set (CC0, PUBLIC-DOMAIN, CC-BY by default). Dates are verified from the document or an official metadata record, never inferred from upload timestamps.
3. Honest claims. Outputs are synthetic reconstructions. No claim of "restoring human authorship" or of AI detection. Separate verbatim reconstruction, stylistic distribution matching and factual fidelity. Test gains against identity (B0) and generic-paraphrase (B1) baselines; report null/negative findings.
4. No leakage. Splits are by document, near-duplicate cluster and author group (htr.data.splits); held-out discipline, period and rewriting prompt; originals of held-out items are never used at inference.
5. Spend is opt-in. The Anthropic adapter runs in mock/dry-run by default; real calls need --live and a budget ceiling; the pilot ceiling is in configs/rewriting/pilot.json and docs/reproducibility/COMPUTE_BUDGET.md. Spend so far is in docs/STATUS.md (from the job ledgers). Ask before exceeding it, before mass acquisition beyond the planned stages, before any public upload.
6. Gates in order (A repository + governance + 100-paper smoke → B validated pairs → C splits/baselines → D M1/M2/M3 pilot → E scale → F release + paper). Each gate writes docs/gates/GATE_<X>.md with GO/NO-GO.
7. Remote origin = gitsrv:srv/git/wpai0002_uqo.git (spbgit on M1M32). Commit after each deliverable. Raw data, pairs and checkpoints are never committed; paper/main.pdf is.

Quick start: .venv/bin/python -m pytest tests -q -m "not network" · scripts/vertical_slice.sh · .venv/bin/python scripts/discover_pmc.py --max-docs 100 · docs/REPRODUCIBILITY.md.