# name: jof250 description: JoF-250 Writing Engine (UQO WP45) — draft, outline, upgrade, review and audit scholarly manuscripts (LaTeX/Markdown) using measured writing conventions from 276 pre-ChatGPT Journal of Finance articles. Use when the user asks to write or improve a paper section, get a referee report, audit tables/equations/citations, locate evidence for a claim, or inspect JoF writing profiles. Commands: /jof-status /jof-corpus-audit /jof-profile /jof-outline /jof-write /jof-review /jof-upgrade /jof-audit /jof-evidence /jof-benchmark.

# JoF-250 Writing Engine

Backing CLI: /Users/simon-pierreboucher/Desktop/UQO/Recherche/Recherche-UQO/methodes/wp45-jof250/.venv/bin/jof250 (Python 3.12 venv at /Users/simon-pierreboucher/Desktop/UQO/Recherche/Recherche-UQO/methodes/wp45-jof250/.venv). Project root: /Users/simon-pierreboucher/Desktop/UQO/Recherche/Recherche-UQO/methodes/wp45-jof250. Docs: /Users/simon-pierreboucher/Desktop/UQO/Recherche/Recherche-UQO/methodes/wp45-jof250/docs/USER_GUIDE.md.

# What it does

• Profiles (/Users/simon-pierreboucher/Desktop/UQO/Recherche/Recherche-UQO/methodes/wp45-jof250/profiles/*.yaml): measured distributions (medians, IQRs, n) of section lengths, abstract/introduction rhetorical moves, contribution placement, identification language, robustness and table conventions, built from the train split of the corpus. Descriptive, not prescriptive.
• Retrieval (jof250 retrieve --topic "..."): hybrid BM25 + local embeddings over structural annotations and ≤400-char excerpts; every hit carries provenance (doc_id, DOI, year, page, section, confidence). Full texts are never exposed.
• Engine (jof250 write --paper P --section S --mode outline|write|upgrade --condition E_full_engine --output F): inspects the manuscript (LaTeX with \input resolution, or Markdown), detects genre/subfield, selects a profile, retrieves rhetorical patterns, drafts, runs deterministic checks (numeric preservation, unknown citation keys, unsupported claims, corpus overlap) and a specialist-referee revision. Output = new file + .report.md (change log, flags, provenance, audit, diff). The manuscript is never modified.
• Reviews (jof250 review --paper P --role referee|editor|table_editor|integrity|replication), audits (jof250 audit --paper P), evidence lookup (jof250 evidence --paper P --claim "...").

# Rules for the assistant

1. Always pass an explicit --output path under the user's paper folder (jof250_proposals/, jof250_reviews/) or /tmp; present proposals and diffs; apply changes only on explicit request.
2. Read and relay the FLAGS and the deterministic audit; never present a generated number or citation as verified.
3. For theory, methods or non-finance manuscripts, say which JoF conventions do not apply (the engine prints a genre note).
4. API calls cost money (ledger: /Users/simon-pierreboucher/Desktop/UQO/Recherche/Recherche-UQO/methodes/wp45-jof250/data/annotations/api_ledger.jsonl; jof250 status shows spend). The benchmark command spends more: dry-run first.
5. If the venv or index is missing, run cd /Users/simon-pierreboucher/Desktop/UQO/Recherche/Recherche-UQO/methodes/wp45-jof250 && make install && make index (see docs/INSTALL.md).