| name | claim-appraise |
| description | Appraise a CLAIM (e.g. a popular diet or health product) — decompose into PICO sub-questions, run a Q1/post-2016/CrossRef-verified systematic review per question, GRADE each, cross-check with OpenEvidence, and ship a verdict-style bilingual site to Cloudflare Pages. |
| user_invocable | true |
Claim Appraisal Workflow (/claim-appraise)
You are turning ONE real-world claim into a defensible, peer-checkable evidence
appraisal. Unlike /lit-review (which reviews a topic into a manuscript), this
adjudicates a claim and renders a verdict per sub-question.
Worked example shipped with this repo: the 4+2R 代謝飲食法 appraisal →
https://verdict-4plus2r-sr.pages.dev/.
When to use
"Is legit?", "appraise the diet/supplement/protocol", "fact-check this health claim".
Hard quality gates (non-negotiable)
- Literature ≥ 2016 and Q1 journals only (
assess_journal_quality(strict=True) + filter_by_year).
- Every DOI resolved via doi.org and CrossRef-verified to exist (
utils/crossref.py) — the anti-hallucination gate. Never cite a work CrossRef can't confirm.
- Per sub-question GRADE certainty (recomputed deterministically, LLM rating is advisory).
- OpenEvidence cross-check where the relay is available (non-fatal if not).
- Overall verdict logic audited with
/argdown before wording ships.
- Attach the claim's primary source URL so peers can confirm no 斷章取義 (cherry-picking).
Steps
-
Identify the real claims. Read the claim's PRIMARY source (official site / book), not secondary blogs. Enumerate every distinct sub-claim. Save the source URL(s).
-
Decompose into PICO sub-questions — one per distinct outcome domain (efficacy AND safety AND any over-extended claims). See scripts/run_4plus2r.py for the PICOQuestion shape. Get the list approved before searching (wrong PICOs waste the run).
-
Run the per-PICO SR (real API search; needs .env keys). For each PICO:
ClaimAppraisalPipeline.run_pico_search → Scopus+PubMed+Embase → dedup → year≥2016 → Q1-strict → DOI-validate → CrossRef-verify → PicoPrismaFlow + included studies. Drive it like scripts/run_4plus2r.py (saves output/<claim>/pico_NN.json).
-
Write one chapter per PICO following output/4plus2r/CHAPTER_SPEC.md: dispatch a subagent per chapter that reads only its pico_NN.json, selects the most relevant HUMAN studies (flag animal/in-vitro as 機轉假說), assigns a verdict chip (vc-strong/moderate/weak/contra/risk/unproven) and GRADE pips, and cites via <cite class="ref" data-doi="DOI">. NEVER invent a citation.
-
GRADE + OpenEvidence cross-check per PICO (grade_judge.py, crosscheck.py). On OE disagree, cap confidence and surface dissent. If the OE browser-relay is down, mark cross-check pending — do not fabricate.
-
Synthesize the executive summary (verdict table), methods (PRISMA totals), and overall verdict. Audit the overall verdict argument with /argdown — confirm the conclusion follows and excludes the obvious overreaches.
-
Render + deploy. Assemble references (DOI→number, AMA, CrossRef/PubMed/OE badges) and render with blueprint_renderer.py (reuses the verdict-site style.css/app.js). Build like scripts/build_4plus2r_site.py, then:
wrangler pages deploy output/<site> --project-name=<name> --branch=main.
Key files (reuse, don't rewrite)
src/litreview/pipeline/claim_orchestrator.py — per-PICO pipeline + assembly
src/litreview/pipeline/{claim_decomposer,grade_judge,crosscheck,verdict_builder}.py
src/litreview/utils/crossref.py — existence gate
src/litreview/pipeline/blueprint_renderer.py + templates/blueprint/ — verdict-style site
scripts/run_4plus2r.py, scripts/build_4plus2r_site.py — runnable templates
output/4plus2r/CHAPTER_SPEC.md — chapter authoring contract
Scale it — dynamic workflow (optional)
For many claims or many PICOs, the parallel fan-out (Steps 3–5) can run as the
bundled dynamic workflow .claude/workflows/claim-appraise.js (background,
resumable, with adversarial verify per verdict). Two-phase hybrid, because a
workflow can't prompt mid-run:
- Phase A (this skill): research the primary source → decompose → approve PICOs → write
output/<slug>/picos.json.
- Phase B (workflow):
ultracode: run the claim-appraise workflow for <slug> (needs Claude Code ≥ v2.1.154 + the ultracode opt-in). Agents call scripts/run_one_pico.py, GRADE, verify, and write chapters.
- Phase C (this skill): synthesis chapters →
/argdown audit → render → deploy.
Allowlist wrangler/uv/scripts/* (see .claude/settings.local.json) so the workflow agents don't stall on prompts.
Output
A bilingual, dual-audience (專業/民眾) verdict site on Cloudflare Pages, plus the
per-PICO pico_NN.json evidence (every study Q1 · ≥2016 · CrossRef-verified).
The renderer emits an OG share-card (assets/og-image.png, 1200×630) for link previews.