Opinion-Fact Separation and Credibility Assessment Engine

What Problem It Solves

Financial research notes, news reports, and investment advice often mix verifiable facts, author judgments, and unsupported claims. Quoting them directly can cause problems: treating “expected revenue growth” as fact, treating anonymous forum posts as authoritative evidence, or being unable to explain why a claim is credible. This skill turns the evaluation into auditable steps: classify information type, judge evidence source, assess logic quality, and then link claims to supporting facts.

How It Works

• Information classification: REALAnalyzer labels content as FACT, OPINION, or UNKNOWN using keywords, data patterns, and opinion markers. It first strips phrase openers like “we believe / we expect / this study” to reduce common misclassification.
• Credibility scoring: It separates HUMAN_TESTIMONY from PHYSICAL_EVIDENCE. Regulatory documents, annual reports, audit reports, major financial media, and brokerage research carry baseline scores; roles such as regulators, executives, chief analysts, retail investors, and KOLs are also scored, then adjusted by recency and source type.
• Logic quality: It checks logical connectors, premises, numeric references, and information volume, then produces judgments such as strong, acceptable, weak, or invalid.
• Fact-claim linking: semantic_support_score connects opinions to supporting facts using lexical overlap and exact number matches, keeping the top three matches by score.

Boundaries

This skill fits Chinese financial text, research-note interpretation, news validation, and investment-advice review. It is rule-based and not guaranteed to be 100% accurate; semantic_support_score is a simplified metric, so a high score does not imply strong causality; batch_analyze does not preserve the fact store across calls, so repeated analysis is better handled with a stateful analyzer.