NEWZ / SOURCE NOTES
NewZ starts with a concrete prediction-market question, then narrows the information field around it. The system is designed to show less when evidence is weak—not to manufacture a full feed.
EIGHT OPERATING PRINCIPLES
Country, sport, league and topic groups use normalized Polymarket tags, series and controlled aliases. Soccer and American football remain separate; a broad Politics or Sports label alone never establishes a close relationship.
Goal: stable filters without guessing from a headline.Open markets that still accept orders rank ahead of near-certain, ended and resolved markets. A live market is shown as >99% or <1% instead of a misleading rounded 100% or 0%; exact endpoints are reserved for confirmed resolutions.
Goal: distinguish conviction from settlement.News retrieval is a union across the selected events. Every article passes event-specific entity, predicate, qualifier and constraint checks. If one event has no credible coverage, NewZ says so instead of inserting lower-relevance stories.
Goal: relevance before volume.Related events share a fixture, series, entity or proposition. Similar events share a controlled theme, such as a US election cycle, or at least the same normalized sport—with league and competition adding weight.
Goal: useful adjacency with visible reasons.Probabilities, volume, outcomes and history come from Polymarket market data. NewZ presents official time-price points as a responsive line rather than inventing OHLC candles from sparse observations.
Goal: interpretation without fabricating precision.Event and reporting search is built only after the same v6 relevance gate used by the News flow. MiniSearch BM25+ weights titles, event entities and summaries; Intl.Segmenter plus character bigrams keep Chinese and Japanese queries useful. A query can narrow an approved set, never restore a rejected item.
Goal: useful retrieval without weakening admission rules.Heat V2 logs each source value, compares it with that source's seven-day median and interquartile range, and shows an empirical percentile only after at least 336 valid buckets across five days. Composite Heat is the median of available source percentiles and requires at least two sources including GDELT or Google. Market volume remains separate.
Goal: relative momentum without mixing incompatible units.Narrative V2 runs only for an explicit event and market target. When enabled after shadow evaluation, multilingual local models rank phrases, tone and claim stance, while user-facing sentences come from fixed templates with source links. Tone, literal resolution alignment and Yes/No claim stance are separate; weak entailment is reported as abstain or unclear.
Goal: traceable summaries rather than free-generated facts.