Every story carries its receipts.
The Provenance Tag is Narvo Intelligence’s contract with readers, regulators, and partners. This page documents the architecture, the corroboration levels, the trusted sources we verify against, and the compliance posture that backs the guarantee.
What the Provenance Tag is
Every narrative Narvo synthesises ships with a visible provenance band showing the source outlets that supported the story, an explicit AI-generation disclosure, and a corroboration label. Nothing is hidden behind an “AI summary” fog — the user can tap through to each source and to the raw Africa Check / Dubawa verification path.
The four corroboration levels
The ConsensusMeter computes a consensus score from the set of independent sources that cover a clustered event. The score is then bucketed into one of four levels; both the score and the level are exposed in the UI and in the GET /v1/story/:id API response.
- 01 · Single source — One outlet. Not yet corroborated — flagged to the user in the Provenance Tag. (score 0.32)
- 02 · Developing — Two or more outlets. Shared cluster — story still reorganising in real time. (score 0.58)
- 03 · Widely reported — Three or more independent outlets agree. Narrative is stable, safe to broadcast. (score 0.78)
- 04 · Strong consensus — Five or more diverse independent sources. Cross-verified against Africa Check & Dubawa archives. (score 0.91)
The RAG architecture
Source articles are ingested, embedded via gemini-embedding, and stored in pgvector. A clustering job runs at 5 AM WAT (and on demand for breaking events) grouping near-duplicate coverage across outlets into a single canonical event. When a synthesis is requested, the Python/FastAPI intelligence core queries the vector store for the top-K sources in the cluster, re-ranks them for diversity, and issues the consensus score based on cross-source agreement against the Africa Check and Dubawa fact-check indices.
Tier-1 source list
Africa Check, Dubawa, and peer fact-checkers form Tier 0 — they are used exclusively to verify claims, never to originate narratives. The national outlets below are Tier 1:
- Punch Nigeria, Vanguard, Channels TV, BusinessDay, The Guardian Nigeria
- Premium Times, Daily Trust, ThisDay, The Cable
- Reuters / AP / AFP wires — used only as cross-verification signal
Publishing additions to this list pass through a monthly editorial review. New outlets are on probation for six weeks — their contributions count toward the display logo list but not toward the corroboration score until the review closes.
What Narvo generates vs. what Narvo sources
Narvo generates: paraphrased broadcast narrative; translation; TTS audio; summaries. Narvo does not generate: original reporting; quotes; named sources; claims that are not present in the underlying Tier-1 articles. Every factual assertion in a Narvo audio clip is traceable to an outlet in the displayed source list — that is not marketing copy, it is an architectural invariant enforced at synthesis time.
NDPA & GDPR posture
Narvo Intelligence is registered with the Nigeria Data Protection Commission and operates an NDPA-aligned data-handling policy: explicit consent for listening history, local storage of personal queue data via IndexedDB, and a user-initiated export and erase path available at Settings → Privacy in the consumer app. GDPR equivalence is maintained automatically via Termly-managed compliance copy for our diaspora traffic. A full transparency report is published annually at /trust/report.