food-review

# Food-Review — Multi-Reviewer Peer Review for Food & Nutrition Manuscripts Give the author the review a good food-science journal would return, from a **panel** rather than a single voice. Original work; architecture informed by open community peer-review skills (see the repo README Acknowledgements). ## Modes - **full** (default) — the whole panel: three domain reviewers + devil's advocate + format check, synthesized by the coordinator into an editorial decision. - **quick** — coordinator + one blended reviewer pass; a fast readiness verdict. - **methodology** — deep dive by `reviewer_methodology` only. - **re-review** — re-assess a revised manuscript against the prior reports and the author's response, verifying each point was addressed. ## Panel (dispatch via the Agent tool; reviewers run in parallel) 1. **`review_coordinator`** (editor-in-chief) — sets the target journal + scope, dispatches `knowledge_builder`, then the reviewers, synthesizes their reports, resolves disagreement, and issues the decision. 2. **`knowledge_builder`** — **runs first**: reads the manuscript's cited sources (Pathway A) and the field's key literature (Pathway B) into a shared **knowledge base** so the panel judges the science from knowledge, not impression. 3. **`reviewer_methodology`** — design, statistics, reproducibility, validation. 4. **`reviewer_domain`** — novelty, significance, scope fit, domain correctness (food/nutrition science). 5. **`reviewer_integrity`** — data & citation integrity, food-safety/ethics, reporting completeness. 6. **`devils_advocate`** — adversarial challenge to the paper's central claim. 7. **`format_checker`** — formatting & reference-style compliance vs the target journal. ## Ground the panel first — the knowledge base Reviewers must **understand the topic and its background before they critique it**. `knowledge_builder` runs **before** the reviewers and builds one knowledge base from: - **A — the manuscript's own citations:** retrieve and **read the full cited articles**, extract what each actually shows, and audit whether it supports the claim it is attached to. - **B — the field's key literature:** extract the manuscript's **keywords and research disciplines**, search the literature for the field's key work (may use the **`food-research` `full review`** branch for discovery/screening — but **knowledge extraction only, no literature-review article**). A + B give the panel the state of the art, standard methods and benchmark values, consensus vs contested points, a novelty map, and gaps — so novelty and correctness are **judged, not guessed**. Never summarize a source that was not retrieved; mark abstract-only and unretrievable items. In **quick** mode, build a light version (Pathway A spot-checks on the load-bearing citations). **Inside `food-pipeline` (Stage 1 already ran):** don't search the field twice — reuse the **Stage-1 evidence base** in place of the Pathway-B search, topped up with `food-research` **quick brief** to find the field's **key review publications** and read them in full; knowledge base = Stage-1 knowledge + key-review knowledge (Pathway A still runs). **Standalone `food-review` is unaffected** and always builds the full A + B. ## Workflow ```mermaid flowchart TD A[Manuscript in] --> B[review_coordinator<br/>resolve target journal + scope] B --> JS[journal-selector<br/>ask once → journal or 'generic'/APA 7.0] B --> KB[knowledge_builder<br/>A: read cited sources<br/>B: key field literature] KB --> KBase[(Knowledge base<br/>state of the art · benchmarks ·<br/>novelty map · cited-source audit)] KBase --> R1[reviewer_methodology] KBase --> R2[reviewer_domain] KBase --> R3[reviewer_integrity] KBase --> R4[devils_advocate] JS --> FC[format_checker] R1 --> C[review_coordinator<br/>synthesize + decision] R2 --> C R3 --> C R4 --> C FC --> C C --> O[Panel report:<br/>per-reviewer reports + format check +<br/>editorial decision + revision checklist +<br/>response-letter skeleton] ``` ## Formatting / target journal The `review_coordinator` first establishes the target journal by calling **`journal-selector/SKILL.md`** (a shared procedure, not an installed skill), which **asks the user which journal the manuscript targets** (they may answer 'generic' → **APA 7.0**). This is asked **once**: the resolved journal's structure, limits, and reference/citation style are recorded and reused, and `format_checker` audits the manuscript against them. Don't re-ask unless the user names a different target journal; reuse the choice if `food-pipeline` already resolved one. ## Output — a Word (`.docx`) report, not Markdown A consolidated **review report delivered as a `.docx`** in the canonical structure of `references/report-format.md`: header (manuscript, target journal, editorial decision, colour legend) → overall assessment → **Part A** editing report by category → summary → **Part B** scientific-quality comments + editorial decision + residual items → **Part C** figure/table consistency audit. Every concern carries a **stable issue ID** (`A#/B#/C#/D#`, `SQ#`, `FC#`) and a `Response (<type>)` line — **Recommendation** or, for a fix that needs the author's data/decision, **Editor query** — with a precise location (`P##` / Table / Figure). Critique the work, not the author. **Markdown is a working format, never the deliverable.** Convert with Pandoc (`pandoc report.md -o report.docx`) or the **`docx` skill**; if neither is available, say so and hand over the Markdown with the conversion command — never claim a `.docx` you did not produce. Apply the colour legend as real Word formatting and leave no Markdown syntax in the file. **Inside `food-pipeline` this report is the single document that `food-paper` later fills responses into — no separate response letter is created** (see `references/report-format.md`). **When the manuscript is a Word (`.docx`) file (or equivalent — LibreOffice / Pages / Google Docs), also deliver the manuscript itself with margin comments.** In addition to the report, insert the panel's concerns as **Word review comments** anchored to the exact text they target (one comment per concern, prefixed with the reviewer lens + severity), using the word processor's Review/Comments feature. See `references/word-review-comments.md`. If the manuscript isn't a Word doc or no Word tooling is available, deliver a location→comment table instead and say so. ## References (load as needed) - `references/report-format.md` — **canonical review report + revision-log format** (Parts A/B/C, issue IDs, `Response (<type>)` taxonomy, editor queries, colour legend). - `references/review-criteria.md` — what each reviewer checks (food-tuned). - `references/quality-rubrics.md` — 1–5 scoring per dimension + weights. - `references/editorial-decisions.md` — `review_coordinator`: Accept/Minor/Major/Reject logic + overrides. - `references/word-review-comments.md` — insert margin comments into a Word/`.docx` manuscript (in addition to the report). - `references/ethics-integrity-checklist.md` — `reviewer_integrity` (canonical; shared with `food-deep-research`). - `food-paper/references/statistics-reporting.md` — `reviewer_methodology`: stats red flags. - `food-paper/references/faithfulness-and-citation.md` — `reviewer_integrity`: verify every citation is real (four-gate) and every claim is source-bound; flag any fabricated/unsupported content. - `food-research/references/full-text-access.md` — **`knowledge_builder`**: how to actually read the cited papers (open access → connected tool → user-supplied PDFs → library session), when to ask the author for access, and how to report what a paywall blocked. - `food-paper/references/privacy-and-confidentiality.md` — check the review report has no local paths/secrets before delivery; `scripts/privacy_scan.py`. ## Handoff Feeds `food-paper` (revise mode) for the author to act on; part of the `food-pipeline` review→revise loop.