Methodology

How Mosaic works

Mosaic turns the world's public information into analyst-grade intelligence. With AI, speed is table stakes — every product can summarize the news fast. What separates intelligence from aggregation is nuance against baselines: knowing what's unusual, what changed, and why it matters.

Live system statusEvery number on the site links back to the signals that produced it.

The engine

Five stages, every item

1CollectWorkers poll 500+ sources on their own cadence — breaking-value Telegram channels every two minutes, ship (AIS) and aircraft (ADS-B) sensors continuously. Every source carries a credibility tier and a state-affiliation label.

2EnrichEach item is translated if needed, then structured by an LLM against a fixed analytic schema — claim status, confidence, significance, geography, entities, market bias. The prompt enforces tradecraft: say what changed, name specific places, separate claims from corroboration, no filler.

3Corroborate & clusterAn engine matches the same event across independent sources, upgrades claim status on confirmation, credits the first reporter, and adjusts source credibility over time. Related signals cluster into Situations — one synthesized headline carrying the full multi-source story.

4Baseline & detectFor every actor, source, and strategic zone, Mosaic keeps rolling 30-day baselines and recomputes deviations every 15 minutes: coverage spikes, sources going silent, ships going dark in chokepoints — and fusion alerts when a physical and a narrative anomaly align in the same place and window.

5ConnectAn entity graph links every actor to its evidence — who appears with whom, which markets react, which situations involve them, what changed this week.

Trust

How every score is computed

Mosaic Score (1–10) — a weighted composite of how credible the source is, how significant the event is, the model's confidence, and how well the core claim is verified. A 9+ means a high-credibility source, a major event, high confidence, and an independently supported claim. It always sits one hover from the reasoning and one click from the sources.

Source credibility is earned, not assigned. Sources start at an editorial tier, then — once they have enough history — are continuously re-blended by their corroboration rate: how often their claims are later confirmed by independent outlets. Reliable sources rise; sources that publish unconfirmed claims sink. Capped at 0.98 — nothing is ever beyond question.

Claim status & confidence are assigned per item, then upgraded automatically when the corroboration engine matches the same event from independent sources. Confidence is always shown, never hidden.

Anomaly severity (0–10) is driven by the size of the deviation from a subject's own 30-day baseline, log-scaled and volume-damped — a 10× spike on 6 mentions is interesting; a 10× spike on 60 is an event. Fusion alerts (aligned physical + narrative indicators) take the highest component severity plus a bonus, because they are the strongest class of early warning.

Design principle: scores compress evidence — they never replace it.

Differentiation

What makes it different

Baseline-aware, not headline-aware

“Iran is in the news” is noise. “Iran coverage is 13× its 30-day norm and the dominant theme shifted from Geopolitics to Conflict” is intelligence.

Multi-INT fusion

650K+ live ship/aircraft observations correlated with narrative anomalies — dark vessels in the Strait of Hormuz while regional coverage spikes produce a single fused alert. No consumer product does this.

Market causality, not quotes

Every notable price move links to the situation driving it; the risk-on/risk-off regime label lists its pressure drivers.

Honesty as a feature

Claim status and confidence are first-class fields, detection methodology is published, and sensor alerts state their own limits (receiver gaps). Nothing is presented as more certain than it is.

See it in action Live system status →