How we turn noise into signal.
An eight-step pipeline, a four-color severity system, and one rule that governs every published line: say what you know, then say what you think, then say what might happen - and never blend them.
1 · Ingest
We pull from Tier-1 open sources (ECMWF, ENTSO-E, WHO, MarineTraffic AIS, NIST, CERT-EU, IMF), Tier-2 vetted aggregators, and partner analyst tagging. Every signal carries a source URL and a first-seen timestamp. No anonymous submissions, no paid placements.
2 · Deduplicate
Near-duplicates are merged across sources using a combination of semantic similarity and shared entity fingerprints. A dedupe ratio below 0.82 blocks auto-merging.
3 · Classify
Each signal is tagged across six lenses: cyber, geopolitical, infrastructure, climate, health, finance. Signals can carry multiple tags; escalation uses the highest-severity lens.
4 · Score
Four scalars, each 0–100, combined into a composite:
- Severity - potential operational impact
- Confidence - source triangulation and verifiability
- Velocity - how fast the situation is evolving
- Exposure - how concentrated the affected population is
5 · Filter
Signals with confidence below 0.55 and severity below 40 are archived to the low-priority view. They remain queryable via the public API but are hidden from the live feed to prevent alert fatigue.
6 · Publish
Every publication uses three labelled blocks:
FACT - a verifiable, sourced, timestamped claim.
ASSESSMENT - our interpretation, bounded by uncertainty language.
FORECAST - a falsifiable prediction with a confidence score and time horizon.
The boundary between these blocks is the most important editorial rule at Vector//Drift. Crossing it is a publishing incident.
7 · Escalate
Composite score ≥ 75 triggers an analyst briefing; ≥ 90 triggers a same-hour alert to subscribers and pushes an amendment to the live dashboard state strip. Escalations are logged with a visible rationale.
8 · Close the loop
Every forecast gets a post-mortem within 30 days of its horizon. Hit rate, calibration error, and brier score are published per analyst, per lens, per quarter. Corrections follow a 60-minute SLA and are never silently overwritten - we publish the delta.
Severity color system
The palette is not decorative.
- Cyan - verified signal, active monitoring, nominal operating state.
- Amber - elevated risk, ambiguity, or deteriorating condition.
- Red - critical: confirmed, fast-moving, high exposure.
- Static - baseline / archival.
What we will not do
- We do not publish unverified claims dressed as facts.
- We do not hide corrections. Every correction has a public ledger entry.
- We do not sell signal amplification. Advertisers cannot influence severity.
- We do not close-source operational code. Running version is always disclosed under AGPL-3.0.
License and disclosure
All code is distributed under AGPL-3.0. The running production version and its commit hash are displayed in the footer of every page, and at /source. If you operate a modified version of Vector//Drift that users can interact with over a network, you must make your corresponding source available to those users under the same license.
◆ Vector//Drift methodology is itself versioned. See METHODOLOGY.md for the canonical, diff-viewable spec.