No revision or outcome ledger; treat as an assessment or scenario.
Weighting Model
Regional Brief / 6H
NO CURRENT BRIEFING
Regional Threat Map
Stable
Elevated
Critical
Intelligence Stream
Multi-domain precursor monitor
Early warning desk
Unusual combinations requiring analyst attention. This is a triage signal, not an event probability.
--/ 100
Watch level
--
Data confidence
--
Horizon
0–7 days
Issued
--
Active signals
Coverage record
BNTI Methodology
1. Data Collection
The system monitors RSS/Atom feeds from regional news outlets, wire services, and specialized publications
for each target country. Sources include local newspapers in native languages, international news agencies,
and regional security-focused media. Data is fetched on a 2-hour workflow with bounded retries, caching,
and no-publish safeguards when feed coverage collapses.
2. LLM Attribution
Each headline is sent to OpenRouter Free for final country attribution and final category
selection. The model must decide whether the headline directly concerns one or more of Türkiye's seven
border neighbors or is IRRELEVANT.
Category weights are fixed after the LLM chooses the class, so scoring remains deterministic even though
the model owns the semantic judgment.
military_conflict: +8.0
terrorism: +7.0
political_instability: +4.0
border_security: +5.0
humanitarian_crisis: +3.0
diplomatic_tensions: +2.5
trade_agreement: -2.0
neutral: 0.0
4. Index Calculation
Per-country scores are based on avg(weight) across LLM-attributed events, then normalized to
1-10 with a saturating exponential curve. Composite BNTI = weighted average across all monitored countries.
5. Thresholds
Stable: 1.0-3.9
Elevated: 4.0-6.9
Critical: 7.0-10.0
6. Regional Briefing
A separate LLM-generated regional brief summarizes the most problematic border-neighbor events from the
last completed 6-hour window. The brief is fixed across the interface and refreshes at most four times per
day.
7. Publish Gate
The dashboard publishes only validated feed coverage. Deterministic country and category rules keep the
pipeline keyless-capable; the optional language model improves ambiguous attribution. Failed or thin runs
leave the last valid public snapshot in place, while the header reports its actual age.
8. Early Warning Ensemble
The early-warning desk combines three independently inspectable components: precursor-language pressure
across news domains (40%), unusual five-session moves in WTI crude, VIX, and US high-yield spreads (35%),
and synchronized index acceleration across monitored countries (25%). Missing components are excluded and
the remaining weights are renormalized.
Market anomalies use a median/MAD robust baseline. Narrative alerts require observable terms and retain
evidence links; cross-source confirmation requires at least two independent domains. Data confidence means
coverage quality, not confidence that an event will happen. The 0–7 day label is a triage horizon, and an
elevated reading is never presented as a probability or prediction.
