Food Research

Food-Research — Comprehensive Evidence Synthesis for Food & Nutrition Science

Build a broad, defensible understanding of a topic by searching many sources, screening them consistently, and synthesizing across them. Original work; no third-party research text is reused. Architecture informed by open community literature-search skills (see Acknowledgements in the repo README).

Streams — pick one and when to use it

Four streams share the same search/screening machinery but differ in depth. Three of them (quick brief, full review, deep research) prioritize sources by journal ranking via journal_ranker; the systematic stream does not (inclusion is by pre-specified eligibility, not prestige).

Overall flow

flowchart TD
    Q[Research question] --> M{Which stream?}
    M -- quick / full --> S1[search_strategist]
    M -- deep research --> DR[food-deep-research skill<br/>scope, plan, investigate,<br/>verify, critique loop]
    M -- systematic --> SR[systematic_reviewer<br/>PRISMA pipeline]
    S1 --> S2[source_scout<br/>four-layer search + dedup]
    S2 --> S3[screener_appraiser<br/>two-phase screening + quality tags]
    S3 --> JR[journal_ranker<br/>Tier 1 preferred; Tier 2 to fill gaps;<br/>avoid Tier 4]
    DR --> JR
    JR --> SY[synthesis<br/>evidence matrix, grading, gaps, coverage advisory]
    SY --> WR[writer<br/>manuscript, APA7 / target journal]
    WR --> RV[reviewer<br/>editorial + integrity]
    RV -- revise --> SY
    RV -- accept --> DOCX[Final review manuscript .docx]
    SR --> SROUT[Systematic-review manuscript .docx<br/>PRISMA + OHAT bias + per-RQ synthesis]

Both the full review and systematic streams finish by writing a manuscript, passing it through the reviewer loop, and delivering a Word document (writer). The quick brief and deep research streams do not (quick brief returns a short brief; deep research is handled by the food-deep-research skill).

Stream detail — invocation & subagent call sequence

Quick brief

• When it wakes: the user wants fast orientation, not an exhaustive review. Phrases like "give me a quick brief on…", "what's known about…", "quick overview of…", "brief me on…", "orient me on…", "TL;DR of the research on…". Also the default when the user asks a scoped factual research question and signals speed ("quickly", "just the highlights").
• How it runs (lightweight — may be inline, no subagents required):
  1. Frame the question in one line (concepts + scope).
  2. One search pass over 2–3 high-yield sources (PubMed/Consensus/CrossRef via MCP, else web search) — no four-layer expansion.
  3. Apply journal_ranker Tier 1 only — keep Q1/Q2 food-science & nutrition, Nature/Science/Cell, and Q1/Q2 other-discipline hits; ignore the rest unless nothing Tier 1 exists.
  4. Skim-appraise (relevance + obvious rigor red flags) — no full rubric.
  5. Write a short brief: 3–6 key findings with citations, the consensus vs open questions, and 2–3 sources to read next.
• Subagents: optional. Run inline for speed; only spin up source_scout if the topic is broad. journal_ranker is applied as a filter step, not necessarily a separate dispatch.

Full review (default)

• When it wakes: the user wants a thorough, citable review/evidence brief — "do a literature review on…", "comprehensive review of…", "survey the field of…", "build an evidence brief on…", "review the evidence for…" — or asks to research a topic without signalling that speed matters.
• How it runs (full subagent pipeline): dispatch subagents in this order (independent retrieval runs in parallel):
  1. search_strategist → search plan (concepts, controlled vocabulary, Boolean strings, source list).
  2. source_scout → four-layer search + dedup → candidate set (parallel per source).
  3. screener_appraiser → two-phase screening + quality rubric → included set with High/Medium/Low tags.
  4. journal_ranker → prioritize by tier (Tier 1 preferred; Tier 2 only to fill gaps; avoid Tier 4).
  5. synthesis → evidence matrix, grading, contradiction resolution, coverage advisory, gaps.
  6. writer → write the review manuscript (APA 7.0 default, or target journal via journal-selector).
  7. reviewer → editorial + integrity review; if not Accept, loop back to synthesis/writer to revise, then re-review (cap ~2–3).
  8. writer → export the accepted manuscript to Word (.docx).
  • Output: a finished review manuscript (.docx) + annotated bibliography + .bib/.ris.

Deep research

• When it wakes: "deep research on…", "investigate … thoroughly", "I need a deep dive / full briefing on…", or a question extending beyond the literature (regulatory, market, technology landscape). Calls the food-deep-research skill; its literature portion still passes through journal_ranker.

Systematic

• When it wakes — use the systematic stream when the user needs a defensible, reproducible, publishable systematic review, signalled by any of:
  • Explicit terms: "systematic review", "systematic literature review", "PRISMA", "meta-analysis".
  • A methodological requirement: "follow a protocol / PROSPERO", "two independent reviewers / dual screening", "with risk of bias", "OHAT", "PRISMA flow diagram".
  • A rigor/audit intent: the user wants the review to be reproducible and auditable (every search string, screening decision, and exclusion reason recorded), not just a narrative overview.
  • If the user only wants a broad narrative overview, use full review instead; if unsure which they want, ask one question ("narrative review or a full PRISMA systematic review with risk-of-bias?").
• How it runs: the systematic_reviewer orchestrator drives protocol → sr_search (≥3 databases) → dual independent three-step screening (sr_screener ×2 + sr_moderator) → PRISMA flow → data_extractor results table → risk_of_bias (OHAT) → sr_synthesis → reviewer loop → writer Word .docx. Journal ranking is not applied (eligibility-based inclusion).

Subagents (dispatch, don't inline)

Run these as subagents (via the Agent tool). Layers that are independent — e.g. per-source retrieval — run in parallel.

1. search_strategist — turns the question into a search plan: concepts, synonyms/controlled vocabulary (MeSH, FSTA/CAB thesaurus terms), Boolean strings per database, filters, and the source list.
2. source_scout — executes the four-layer retrieval across sources, records hit counts, and deduplicates into one candidate set.
3. screener_appraiser — two-phase screening + the food-science quality rubric; outputs the included set with quality tags.
4. journal_ranker — prioritizes the screened sources by journal ranking (Q1/Q2 food-science & nutrition, plus Nature/Science/Cell families and Q1/Q2 in any other discipline = highest; Q3 second; Q4 avoided). Used by quick brief, full review, and deep research only — never inside a systematic review.
5. synthesis — evidence matrix, contradiction resolution, evidence grading, gap analysis, and the coverage advisory.
6. writer — writes the review manuscript and exports Word (.docx) (APA 7.0 default, or target journal via journal-selector). Full review + systematic.
7. reviewer — combined editorial + integrity review with a revision loop. Full review + systematic.

Systematic-review subagents (systematic stream only): 8. systematic_reviewer — PRISMA orchestrator (protocol → search → dual screening → PRISMA → extraction → risk of bias → synthesis → review → Word). 9. sr_search — ≥3 databases (Web of Science, Scopus, PubMed preferred); combine + deduplicate; log all strings/counts. 10. sr_screener — run as two independent instances; three steps (title → abstract → full text) with per-record include/exclude + reasons. 11. sr_moderator — after each step, compares the two screeners, resolves conflicts, keeps PRISMA counts. 12. data_extractor — pulls the results table (by research question) from the final shortlist. 13. risk_of_bias — OHAT risk-of-bias assessment (in vitro / human / animal) by default. 14. sr_synthesis — PRISMA description → risk-of-bias results → per-RQ synthesis; formats APA 7.0 or target journal.

For a quick brief you may run the workflow inline without subagents.

Step 1 — Frame the question (search_strategist)

• Interventions/nutrition: PICO (Population, Intervention/Exposure, Comparator, Outcome).
• Composition/process/safety: define the food matrix, factor/treatment, and measured response.
• State scope, timeframe, languages, and exclusions. Break the question into concepts and list synonyms + controlled-vocabulary terms per concept.

Step 2 — Plan the sources

Cover several source classes so the picture isn't skewed by one index:

• Bibliographic: FSTA (Food Science & Technology Abstracts — the core food index), PubMed/MEDLINE, Web of Science, Scopus, CAB Abstracts, AGRICOLA, AGRIS (FAO).
• Preprints: bioRxiv, ChemRxiv, agriRxiv.
• Semantic / aggregators: CrossRef, Semantic Scholar, Consensus, Dimensions, Lens.org.
• Safety & regulatory / grey: EFSA, US FDA, USDA (incl. FoodData Central), Codex Alimentarius, WHO, EU/national food-standards bodies.
• Chemistry / bioactives: PubChem, ChEMBL, FooDB, Phenol-Explorer.
• Methods / standards: AOAC, ISO.

Tooling: use whatever literature MCP tools are connected (e.g. PubMed, Consensus, bioRxiv, CrossRef, Scopus/ScienceDirect) for live retrieval; fall back to web search for any source without a tool. Record which tool/source produced each result so the search is reproducible.

Step 3 — Four-layer search (source_scout)

1. Layer 1 — structured search: Boolean/keyword + controlled vocabulary across the bibliographic databases (target 100–500 raw hits). Apply date/language filters.
2. Layer 2 — backward chaining: mine the reference lists of the key reviews and seminal papers for older frequently-cited work.
3. Layer 3 — forward chaining: "cited by" from seminal works to catch the latest research.
4. Layer 4 — semantic / cross-disciplinary: related-article and semantic tools to catch methodologically or disciplinarily adjacent work (chemistry, engineering, nutrition, microbiology) that keyword search misses.

• Deduplicate by DOI/title across sources. Record the hit count at each layer.
• Stop when the search saturates — e.g. ≥3 of: no new themes appearing, citation loops closing, timeframe covered, key authors/venues all seen, new hits <10% novel.

Step 4 — Two-phase screening & appraisal (screener_appraiser)

• Phase A — title/abstract: apply inclusion/exclusion; narrow to ~30–50 candidates.
• Phase B — full text: read the semantically strong and borderline items; land ~15–30 (more for systematic).
• Quality rubric (score each source): study design & rigor; replication and whether n is biological (not pseudo-replicated); method validation (LOD/LOQ, recovery, controls, appropriate standards); journal quality and predatory/fabrication check; relevance to the question; recency/currency. Tag each source High / Medium / Low.
• Universal gates (relevance, methodological soundness, predatory/fabrication) are never waived; only publication-type/recency expectations flex by subfield.

Step 4.5 — Prioritize by journal ranking (journal_ranker) — quick / full / deep only

• Tier every screened source: Tier 1 = Q1/Q2 in Food Science & Technology or Nutrition & Dietetics, any Nature/Science/Cell-family journal, or Q1/Q2 in any other WoS discipline/multidisciplinary category; Tier 2 = Q3; Tier 3 = Q4 (avoid).
• Prefer the highest tier that covers each point — if Tier 1 sources suffice, don't include Tier 2/3 for it; drop to Tier 2 only when Tier 1 is insufficient; use Tier 3 only when nothing better exists, and flag it.
• Uses references/journal-priority.csv for food/nutrition quartiles; JCR knowledge for other fields.
• Skip this step entirely in the systematic stream — inclusion there is by eligibility, not journal ranking.

Step 5 — Synthesis (synthesis)

• Evidence matrix: source × theme grid showing coverage density and method spread.
• Integrate & resolve conflicts: weigh by design and rigor; separate consistent findings from contested ones; explain disagreements (matrix, method, dose, population).
• Grade the evidence: prefer systematic reviews/RCTs for health/nutrition claims; require standardized measurement (AOAC/ISO) for compositional/process claims. State confidence and why.
• Coverage advisory: flag when >70% of sources share one publication year, region, food matrix, method, or venue family — a bias risk.
• Gaps: under-powered areas, missing methods, population/geographic voids; propose the next study.

Deliverables

An evidence brief containing: question & scope; reproducible search strategy (sources, Boolean strings, filters, dates); screening funnel with counts; annotated bibliography (per source: design, findings, relevance, quality tag, intended paper section); literature/evidence matrix; graded conclusions; coverage advisory; and a gap list. Export references as .bib/.ris (deduplicated) for reuse.

Deep dives

For a subtopic that needs open-ended investigation beyond the literature (e.g. regulatory landscape, market/technology state), call the food-deep-research skill and fold its sourced synthesis back into the evidence brief.

References (load as needed)

• references/literature-sources.md — databases + APIs (FSTA/PubMed/WoS/Scopus/CrossRef/OpenAlex + EFSA/FDA/USDA) for search_strategist/source_scout/sr_search.
• references/source-quality-hierarchy.md — evidence grading for screener_appraiser/synthesis.
• references/reporting-guidelines.md — EQUATOR/PRISMA/CONSORT/STROBE for the systematic stream and appraisal.
• references/ohat-risk-of-bias.md — full OHAT tool (11 questions, 4-point scale, design applicability incl. corrected in-vitro Q3/Q4 = NA, and in-vitro criteria) for risk_of_bias.
• food-paper/references/faithfulness-and-citation.md — grounding + four-gate citation check. Every finding, number, and citation traces to a real source; never fabricate. scripts/verify_citations.py audits the reference set.
• food-paper/references/privacy-and-confidentiality.md — privacy scan before delivering the brief/report (no local paths/secrets); scripts/privacy_scan.py.

Handoff

Sources tagged and assigned by section feed food-paper (Introduction and Discussion evidence, reference list) and are orchestrated by food-pipeline.

Food & nutrition rigor notes

Watch for pseudo-replication (analytical replicates as biological n); matrix effects and single-cultivar/single-batch over-generalization; unvalidated assays; and undisclosed funding/conflicts, which are common and material here.