research-reviewer

# Skill: Research Item Review ## When Not to Use - When the research item is still in progress — complete the investigation first - When a quick sanity check is needed rather than a full audit — use `citation-discipline`, `speculation-control`, or `remove-ai-slop` individually - When the goal is to improve the writing — this skill audits; it does not rewrite --- ## Interaction Protocol **Operating mode**: Audit-only. Read the item. Report violations. Do not edit, rewrite, or commit anything. **Before starting**, confirm: 1. Is the item complete? (Findings, Evidence Map, and Executive Summary are populated — not stubs or "TBD") 2. Which sections require full audit? (Research Skill Output and Findings sections require full checks; Context section has lighter standards — see scope notes below) **Output**: A structured violation report. See Output Format. --- ## Inputs and Outputs **Input**: A completed research item in structured Markdown format with Research Skill Output and Findings sections **Output**: A structured audit report with per-skill PASS/FAIL verdicts and specific violations **Composability**: Use after `research` §8 Output Finalisation; referenced in `research` skill as the final gate before marking an item complete. Applies `citation-discipline`, `speculation-control`, `remove-ai-slop`, and `peer-reviewer` in sequence. --- ## Scope Notes Apply these standards consistently across all items: **Research Skill Output and Findings sections**: Full audit. All four skills apply at maximum rigour. **Context section**: Lighter standard. - `citation-discipline`: flag only hard factual claims (specific statistics, named studies, precise dates) that lack a source. Do not flag framing prose, motivating rationale, or working hypotheses. - `speculation-control`: flag overconfident assertions presented as settled fact. Allow unlabelled framing assumptions and working hypotheses — the Context section is intentionally a pre-investigation framing device. - `remove-ai-slop`: apply normally. **Template structure**: Do not flag required structural elements (section headers, numbered sections, Evidence Map table) as AI slop. These are required by the format. --- ## Step 1 — Citation Discipline Check the Research Skill Output and Findings sections for: - Factual claims that lack a `[fact]`, `[inference]`, or `[assumption]` label and a corresponding source - Sourced claims where the citation does not directly support what is asserted - Vague quantifiers ("many", "most", "significant") used without sourcing - Acronyms or initialisms used without expansion on first occurrence - Domain terms used without a link to an authoritative definition Do not flag stubs or "TBD" sections — an incomplete section is not a citation violation. The labels `[inference]` and `[assumption]` are the accepted convention for epistemic labelling in research items — treat them as compliant with citation discipline for the purpose of distinguishing verified facts from inferences. --- ## Step 2 — Speculation Control Check the Research Skill Output and Findings sections for: - Speculative statements that are not prefixed with `Speculation:` or `Hypothesis:`, or not labelled `[inference]` or `[assumption]` - Opinions or subjective judgments that are not labelled `Opinion:` - Overconfident language ("clearly", "obviously", "undeniably") unsupported by evidence - Evidentiary gaps filled with invented or assumed detail presented as fact The labels `[inference]` and `[assumption]` in Research Skill Output sections are the accepted convention — treat them as compliant. --- ## Step 3 — Remove AI Slop Check the Findings section (Executive Summary, Key Findings, Analysis, etc.) for: - Formulaic AI writing patterns: predictable transitions ("Furthermore", "Additionally", "It is important to note"), symmetrical paragraphs, explicit meta-structure ("In conclusion") - Alignment artifacts: safety-prefacing language, over-explained causal links, unnecessary framing sentences - Lexical monotony: repeated sentence openings, uniform sentence length, excessive hedging density Apply `remove-ai-slop/SKILL.md` checks in full. --- ## Step 4 — Peer Review Check the Executive Summary, Key Findings, and Analysis sections for: - Conclusions in the Executive Summary that lack corresponding supporting evidence in the Findings or Evidence Map - Confidence levels not calibrated to source count and independence: High requires multiple independent sources; Medium requires one credible source; Low applies to single uncertain sources or strong inferential leaps - Central findings that present a single explanation for a contested or multi-causal phenomenon without acknowledging alternatives exist - Competing hypotheses dismissed without reasoning or evidence - Missing cross-references to related completed items where the connection is material to the conclusion Apply `peer-reviewer/SKILL.md` checks in full. --- ## Output Format Produce the report in this exact format: ``` REVIEW_TARGET: <item identifier or title> citation-discipline: PASS | FAIL VIOLATION: <specific violation with section reference or quote> speculation-control: PASS | FAIL VIOLATION: <specific violation with section reference or quote> remove-ai-slop: PASS | FAIL VIOLATION: <specific violation with section reference or quote> peer-reviewer: PASS | FAIL logical-coherence-and-evidence-sufficiency: PASS | FAIL VIOLATION: <specific violation with section reference or quote> alternative-explanations: PASS | FAIL VIOLATION: <specific violation with section reference or quote> cross-item-integration: PASS | FAIL VIOLATION: <specific violation with section reference or quote> OVERALL: PASS | FAIL ``` Rules: - Use `PASS` if no violations are found for a skill; use `FAIL` if any violations are found - Each violation on its own line, indented two spaces, prefixed `VIOLATION: ` - `OVERALL: PASS` only if all four skills passed; otherwise `OVERALL: FAIL` - If a required companion skill file is missing, write `SKILL_FILE_MISSING` as the result and treat it as `PASS` for the overall calculation - If the item is a stub (Findings not yet populated), write `ITEM_INCOMPLETE` and do not audit