HTA Quality Check (scoring + consistency)
Manually evaluate a drafted HTA document against the official Swiss rubrics. You — the agent — are the evaluator: read the document, apply the rubric criteria yourself, and justify every score with evidence quoted or referenced from the document. Do not delegate the judgment to any external scoring service.
Step 1 — Identify the document and its type
Locate the document to review. Default locations in a dossier: applications/ch/<form>/<lang>.md or chapters/*.md (for report-level reviews). If the user did not name a file, list the candidates and ask.
Determine the document type — it selects the rubric:
| Document type | Quality rubric | Also run consistency? |
|---|---|---|
| MiGeL application | references/rubrics/ch-migel.md |
yes |
| Analysenliste form | references/rubrics/ch-analysenliste.md |
yes |
| Antrag neue Leistung | references/rubrics/ch-antrag-neue-leistung.md |
yes |
| Meldung neue Leistung | references/rubrics/ch-meldung-neue-leistung.md |
yes |
| Umstrittenheitsabklärung | references/rubrics/ch-umstrittenheit.md |
yes |
| EurHTA / chapter set | (no form rubric) | consistency only |
The consistency rubric for all types is references/rubrics/consistency.md.
Review language follows the document: German for BAG forms, English for EurHTA reports, unless the user asks otherwise.
Step 2 — Gather context
Read alongside the document, where present in the dossier:
dossier.yaml— intervention description and notes, to judge whether the form reflects the actual intervention.prisma/prisma-diagram.mdandprisma/included-studies.json— to verify PRISMA counts and citations in the document against the actual literature data.models/runs/*.json— to verify that economic figures in the document match the engine outputs. Never recompute model results yourself; compare against the run files only (regeneration is done with thepackages/heor-enginescripts,npx tsx <script>, if the user asks).
Context files may be absent; then evaluate the document on its own terms and note the missing cross-checks as limitations.
Step 3 — Apply the quality rubric
Read the applicable rubric file completely and score every dimension, module/section, and (where defined) WZW or dispute category it lists. For each score:
- Cite concrete evidence from the document (section number, quoted fragment) — no unexplained numbers.
- Check the rubric's "critical issues" list explicitly; a critical issue must appear in the report even if the affected section otherwise scores well.
- Systematically hunt for unresolved placeholders (
[Bitte hier eintragen],{{...}},[Veuillez compléter],[Da compilare], empty table cells) — these are always at least a missing-field finding. - Verify data against the dossier: PRISMA counts in the form vs.
prisma/, economic figures vs.models/runs/, intervention facts vs.dossier.yaml. Any mismatch is a critical issue.
Step 4 — Apply the consistency rubric
Read references/rubrics/consistency.md, pick the document-type-specific check block, and work through it plus the general principles. Focus on cross-section agreement: comparator and population definitions, PICOT (for Antrag), repeated numerical values, terminology, CE/regulatory status, and citation-reference integrity. Score consistency 1–10 on the applicable scale (German rubric for BAG forms).
Step 5 — Write the review report
Write a structured Markdown report next to the reviewed document as <name>.review.md (e.g., applications/ch/migel/de.review.md); do not modify the reviewed document itself. If a review already exists, write a new dated one or diff — never blind-overwrite. Structure:
- Header: document reviewed, date, rubrics applied.
- Scores at a glance: overall quality score, consistency score, readiness level — as a table.
- Dimension / module / section scores: table per the rubric's output structure, each with a one-line justification.
- WZW / dispute-category assessment (where the rubric defines one).
- Critical issues — must-fix before submission.
- Missing fields / open placeholders — exhaustive list.
- Consistency findings — per-section issues, recommendations, strengths.
- Recommendations — prioritized, specific, actionable.
Summarize the top findings to the user in chat (scores, readiness level, the 3–5 most important fixes).
Step 6 — Mandatory expert review disclaimer
End the report with, and repeat to the user:
Hinweis / Disclaimer: Diese Bewertung wurde automatisiert erstellt und ist eine Entscheidungshilfe, keine behördliche oder fachliche Prüfung. Scores und Bereitschaftseinschätzung ersetzen nicht die Beurteilung durch qualifizierte Fachpersonen; vor einer Einreichung beim BAG/ELGK/EAMGK ist eine Expertenprüfung zwingend.
This step is not optional.