fact-checking

name	fact-checking
description	Use when working on claim verification in graph-protokoll: detecting checkable statements, assigning verdicts (bestätigt/teilweise/irreführend/falsch/unbelegt), attaching sources, the evidence corpus, or the "every fact-check has a source" invariant. Trigger on pipeline/factcheck.py, data/evidence/evidenz.json, Aussage/Verdikt/Quelle/Evidenz tasks.

Fact-checking (Aussage → Verdikt + Quelle + Audio-Beleg)

The project's distinctive feature: quantitative claims in speeches are checked against an official evidence corpus, and each verdict always carries a source plus the audio timestamp of the original statement.

How it works

extract (Aussage) → factcheck → Faktencheck{verdikt, begründung, quelle} → graph: GEPRUEFT_ALS → BELEGT_MIT

• Demopfad factcheck_rule_based(aussagen, evidenz_path): deterministic keyword-overlap match against data/evidence/evidenz.json (≥0.6 → that verdict; else unbelegt).
• Produktivpfad factcheck_with_retrieval (contract only, raises NotImplementedError): dense retrieval over an official corpus (Destatis, Bundesregierung answers, Drucksachen via DIP-API) + an NLI-style LLM verifier (entailment/contradiction) with citation duty.
• Claims come from extract._is_checkable (a quantifiable figure present, and not a promise like "wird … vorlegen" — those become Aufgabes, not facts).

The invariant — do not break it

assert all(fc.quelle for fc in results), "Faktencheck ohne Quelle — verletzt Grundsatz."

• Every FactCheck has a quelle. Even unbelegt references the corpus it was checked against, with stand: "not provable against source X up to date Y" — not "not checked". Never produce a verdict without a source; never let the assert weaken.
• unbelegt ≠ falsch. Missing evidence is reported as missing, never as refutation.
• Verdict scale is fixed (VERDIKTE): bestätigt · teilweise · irreführend · falsch · unbelegt. Don't add verdicts without updating VERDIKTE, the accessible._VERDIKT_WORT verbalization, the web frontend .fc styles, and the Neo4j few-shot examples.

Boundaries (keep these honest)

Opinions are not checked (only quantifiable factual claims); a verdict is a sourced suggestion, not a ruling — always Mensch-im-Loop with a Widerspruchspfad; evidence carries a stand so stale belege are flagged. The demo evidence (data/evidence/evidenz.json) is fictional — never add real statistics or real sources to it.

Hard policy — no verdicts on real people. The rule-based checker against the fictional corpus is not a reliable fact-checker (it produced a spurious falsch on a real MdB when run on a real protocol). Real sessions are ingested with --no-factcheck — never auto-publish Faktencheck/Quelle/GEPRUEFT_ALS verdicts on real, named persons (Persönlichkeitsrecht). Real fact-checking needs a real evidence corpus + LLM-NLI verifier + human review. The fact-check mechanism is demonstrated only on the fictional bundestag scenario. See [[bundestag-ingestion]] and docs/spark-und-echtdaten.md Teil D.

Verify

python3 run_demo.py --scenario bundestag (prints the fact-check block); the assert runs on every build. Background: docs/bundestag-protokollierung-analyse.md §4.