Transparency

Methodology

IntelCap transforms source articles into event intelligence. Every output remains linked to its evidence. The pipeline does not claim that a model has verified truth.

Collection and provenance

Feeds come from public and official endpoints in the source registry. IntelCap preserves publisher, original URL, title, language, timestamps and a limited snippet. It detects likely attribution and syndication so copies are not counted as independent confirmations.

Deduplication and event clustering

Canonical URLs and fingerprints are combined with title and SimHash similarity, multilingual terms, entities, actions, event type, amounts and time proximity. Contradictory actions or amounts block automatic merging. Optional multilingual embeddings can add a signal but never decide a merge alone.

Impact Score

Impact estimates potential importance, not truth. Event magnitude, market relevance, entity significance, breadth, novelty and urgency are weighted by event profile. Unavailable capital or market-reaction inputs are omitted and remaining weights are normalized.

Event profile × (magnitude + market relevance + entity significance + breadth + novelty + urgency + available market inputs)

Confidence Score

Confidence is independent from Impact. It combines category-specific source trust, diminishing independent corroboration, primary evidence, consistency, extraction and attribution. Rumors, anonymous attribution, circular reporting, contradictions and retractions apply explicit penalties.

30% Source trust + 25% Independent corroboration + 20% Primary evidence + 10% Claim consistency + 10% Extraction + 5% Attribution − explicit penalties

Limitations

Extraction, lineage detection and clustering can make mistakes. Embeddings and DeepL translation are optional and disabled without configured providers. Market movement is shown only from real cached provider observations and describes timing, not causality.