How KVB News Works
KVB News is an AI-native newsroom built around a single invariant: a language model may organize and translate facts from a primary source, but it may never invent a fact, number, quote, or link.
1. Permissioned Source Ingestion
Our Go ingestion worker polls vetted primary institutions, wire agencies, and broadcasters. Every source carries explicit permission metadata (ai_processing_allowed, republishing_mode, and storage_allowed). Sources that only permit link attribution are never summarized beyond their headline.
2. Five-Criteria Triage
Each ingested item is evaluated inside an untrusted <source_document> boundary by the configured triage model (`classify.v3`), scoring 0–20 across five criteria: Affected Population, Institutional Action, Novelty, Consequence, and Source Authority, while extracting ISO country codes, named entities, and primary-statement signals.
3. Closed-Evidence Synthesis
High-scoring items are synthesized by the configured drafting model (`compose.v3`). The model is bound by a strict closed-evidence contract: every name, figure, date, and quote must come strictly from the source document. Even true world knowledge outside the source document is forbidden.
4. Deterministic 6-Check Grounding Gate
Before any draft can reach the editorial queue, deterministic Go code verifies: (1) zero URLs or markdown links in model output, (2) verbatim quote substring matching, (3) exact numeric figure matching, (4) numeric scale matching (million/billion/trillion), (5) number-word consistency, and (6) calendar month verification. Any mismatch rejects the draft automatically.
5. Human Editorial Sign-Off
Grounded dispatches enter the KVB Editor Desk in pending_review state. Editors inspect the side-by-side primary source and AI synthesis before approving publication to the web and Telegram channels.