Deep Research

Deep-Research — Source-Validated Literature Review Engine

Answer a hard question properly and hand back a written, formatted, integrity- checked literature review — not just notes. Scope → design → discover → screen by journal ranking → validate sources → extract & verify evidence → synthesize → stress-test → write → review-loop → final report. Original work; architecture informed by open community food-deep-research skills (see the repo README Acknowledgements). Usable standalone, or as the deep-dive engine called by food-research.

Modes

• quick brief — scope → discover → screen (Tier 1) → light synthesis → short sourced answer. Skips the full validation/compile/review loop.
• full — the default: the complete 12-subagent pipeline below with the iterate-to-saturation and compile↔review loops, ending in a finished review.

Subagent team (dispatch via the Agent tool)

Workflow

flowchart TD
    A[research_scope<br/>scope brief] --> B[research_architect<br/>methodology blueprint]
    B --> C[investigator Pass 1<br/>discover candidate sources]
    C --> D[source_screener<br/>journal-ranking tiers]
    D --> E[source_verifier<br/>validate → Source Quality Matrix]
    E --> F[bibliography<br/>dedupe + format + citation map]
    E --> G[investigator Pass 2<br/>extract evidence from validated sources]
    G --> H[claim_verifier<br/>verify claims vs validated sources]
    H --> I[synthesizer<br/>matrix, themes, conflicts, grading, gaps]
    I --> J[critic<br/>stress-test synthesis]
    J -- gaps --> C
    J -- sound --> K[compiler<br/>write + format review<br/>APA 7.0 / target journal]
    F --> K
    K --> L[editor + ethics_reviewer<br/>editorial + integrity review]
    L -- minor/major revision --> K
    L -- accept --> M[Final literature review]

Two loops: (1) evidence loop — critic sends gaps back to investigator (cap ~2–3); (2) writing loop — editor/ethics_reviewer send revisions back to compiler until Accept (cap ~2–3), then deliver.

Source discipline (non-negotiable)

• Investigation and claim-checking operate only on validated sources — those that passed source_screener (ranking) and source_verifier (validity). Retracted/unresolvable sources are excluded and logged.
• Journal ranking favors Tier 1 (Q1/Q2 food-science & nutrition, Nature/Science/Cell families, Q1/Q2 in any other discipline); Tier 2 (Q3) only to fill gaps; Tier 4 avoided.
• Every claim carries a source and locator; inference is labelled as inference.

Formatting — resolve the target journal once

At the start, call journal-selector, which asks which journal the review targets (they may answer 'generic' → APA 7.0). Ask once: record the choice and reuse it in bibliography and compiler to format the review to that journal's structure, limits, and reference style — don't ask again. Reuse a journal already resolved by food-research/food-pipeline; re-resolve only if the user asks to switch journals.

Principles

Every claim sourced; fact separated from interpretation; disagreement shown, not averaged; uncertainty surfaced. Upstream evidence beats parametric knowledge — mark missing evidence [EVIDENCE GAP], never fabricate.

References (load as needed)

• references/reasoning-and-fallacies.md — synthesizer/critic: sound argument + fallacies.
• food-research/references/literature-sources.md — databases/APIs for investigator.
• food-research/references/source-quality-hierarchy.md — evidence grading for source_verifier/synthesizer.
• food-research/references/reporting-guidelines.md — EQUATOR/PRISMA/CONSORT/STROBE.
• food-paper/references/apa7-quickref.md — APA 7.0 for bibliography/compiler (default style).
• food-review/references/ethics-integrity-checklist.md — for ethics_reviewer.
• food-paper/references/faithfulness-and-citation.md — grounding + four-gate citation check for bibliography, source_verifier, claim_verifier, compiler. Never invent sources/data; run scripts/verify_citations.py.
• food-paper/references/privacy-and-confidentiality.md — privacy scan before delivering the report (no local paths/secrets); scripts/privacy_scan.py.

Handoff

Standalone → deliver the final review. Called by food-research → return the validated synthesis (or the finished review) to fold into the evidence brief.