# Brief template — specialist reviewers (Agents E AI, F Systems, G Theory, H Security), Phase 5

Inputs you receive (generated by Phase 4, never edited by hand):

• literature_synthesis/landmarks/<area>_landmarks.csv — the candidate landmark works of your area: OpenAlex id, DOI/arXiv, title, venue, year, categories, c12/c36 citation windows, NCS36 (field-normalised), FWCI, abstract availability, OA full-text URL.
• literature_synthesis/landmarks/<area>_abstracts.jsonl — title + abstract for each candidate (metadata only).
• results/graph/category_flows_yearly.csv, results/graph/category_counts_monthly.csv, results/econometrics/* for your categories.
• manuscript/references.bib (verified entries) — you may cite only keys that exist there or add entries you verify with Crossref/OpenAlex/arXiv single lookups (record them in references_verification.csv).

Deliverables (structured, one directory per agent: literature_synthesis/<area>/):

1. claims.csv — one row per scientific claim you make in the review: claim_id, claim (one sentence), claim_type ∈ {established, contested, speculative}, evidence_type ∈ {metadata_only, abstract, full_text, corpus_statistic}, supporting_keys (bib keys or OpenAlex ids), contradicting_keys, corpus_artifact (file + column/row), confidence (0–1), verification_status ∈ {unverified, machine-verified, human-verified}, notes.
2. contributions.csv — for each landmark you discuss: openalex_id, contribution_type ∈ {theory, method, empirical, benchmark, engineering, survey}, one-line description, whether the type was judged from abstract or full text.
3. review.md — the disciplinary synthesis (3,000–6,000 words), sectioned by category, written in the WP44 academic style (CLAUDE.md §10): critical comparison, established vs speculative, unresolved problems (feeds Part X), conflicting findings reconciled explicitly, no causal language about ChatGPT, no "largest ever" claims.
4. open_problems.csv — problem_id, statement, categories, why unresolved, testable direction, supporting_keys.

Integrity rules (CLAUDE.md §1): never invent content of a paper you have only metadata for — say "the abstract reports"; never call a preprint peer-reviewed; an abstract is not a full-text review (evidence_type must say which); every number comes from a corpus artefact you cite by file. Your review is checked by Agent K (critic) before any sentence enters the manuscript.