Transparent scoring framework, source coverage, normalization logic, confidence construction, limitations, compliance notes, and auditability.
Sentiment, risk, volume, novelty, confidence, and narrative-shift scores are normalized indicators. They are designed to expose language change, not to produce investment recommendations.
Sentiment scores are normalized within company, sector, and index contexts. A score of 80 does not mean 80% positive; it indicates the company's current language environment is strongly positive relative to its historical and peer baselines.
Confidence blends source reliability, source dispersion, evidence count, entity mapping quality, recency, and agreement across document types.
Earnings transcripts, filings, regulatory documents, and high-reliability news receive higher evidence weights than low-frequency or lower-verifiability sources.
Percentiles compare the current score to company, sector, and SPY baselines. Z-scores show distance from a trailing baseline and are used for alert thresholds.
This prototype uses local mock data. Production usage would require entitlement-aware ingestion, source validation, audit trails, and documented model governance.
The system avoids buy/sell recommendations and presents source-backed signal context for research workflows. Output should be reviewed by qualified investment professionals.
Every summary should retain links to contributing sources, timestamps, topic tags, and signal contribution metadata so analysts can inspect the evidence chain.