Agri-Deep-Research — Source-Validated Reviews for Agricultural Science

Run the food-deep-research skill exactly — its 12-subagent team (research_scope, research_architect, investigator, source_screener, source_verifier, bibliography, claim_verifier, synthesizer, critic, compiler, editor, ethics_reviewer), both loops (evidence loop and compile↔review loop), and its source discipline — with the agriculture substitutions in agri-research/references/agriculture-domain.md. Read that file first. No new machinery here.

The substitutions

1. Persona — a senior agricultural scientist of the specific discipline; name it and apply its standards (domain §2). research_architect designs the method to that discipline's conventions.
2. Evidence base — source_screener ranks agriculture + multidisciplinary literature: Tier 1 = Q1/Q2 of the seven agriculture categories (journals/_coverage_agriculture.md) + Nature/Science/Cell/PNAS + Q1/Q2 adjacent disciplines; Tier 2 = Q3 for gaps; Q4 avoided. FAO/USDA/CGIAR/EFSA and extension sources are evidence with a source and date (domain §3).
3. Journal routing — bibliography and compiler format via journal-selector using the agriculture coverage map (domain §4); APA 7.0 by default.

Source discipline (inherited, non-negotiable)

Investigation and claim-checking operate only on validated sources — those that passed source_screener (ranking) and source_verifier (existence, venue legitimacy, retraction, predatory check). Every claim carries a source and locator; inference is labelled as inference; [EVIDENCE GAP] rather than filling from memory.

Agricultural rigour

Apply domain §5 — the critic should attack the usual agricultural weak points: single site-year generalised to a recommendation, pseudoreplication (subsamples treated as replicates), pot-to-field extrapolation, missing G×E, and causal language unearned by the design.

Inherited unchanged (not optional)

Four-gate citation verification (scripts/verify_citations.py), privacy scan, academic style + AI-tell removal (food-paper/references/writing-style.md with human-writing.md), and the mandatory AI-use disclosure. Also the full-text-access first move — food-deep-research's highlighted, one-time request for the user's EndNote .Data folder / reference PDFs, and full-text extraction via the ladder before the evidence loop (food-research/references/full-text-access.md).