SENTINEL
Defense Mobilization Intelligence Platform — Sources & Methodology
Sentinel is an autonomous intelligence system that continuously monitors public defense, geopolitical, and economic signals — then synthesizes them into assessed intelligence products using a knowledge graph and 15 AI agents. It runs 24/7 without human intervention, producing briefings, risk assessments, alerts, and forecasts that follow Pentagon IIR (Intelligence Information Report) conventions.
The Core Idea
The value of Sentinel is not in any single data source. It is in the ontology + agent mesh — a general-purpose engine for cross-domain intelligence fusion.
The ontology defines what exists (26 entity types, 21 relationship types) and how things connect. Companies manufacture weapon systems. States deploy forces in theaters. Contracts fund supply chains. News signals mention entities. Threats indicate risks. This is domain knowledge encoded as a traversable graph.
The agents define what to do about it — 15 specialized AI agents that continuously enrich, correlate, assess, and synthesize the graph into intelligence products. Each agent has a specific mandate and output discipline. Together, they form an autonomous intelligence production line.
Plug in any data source and the engine does the rest. A new RSS feed, a vessel tracking API, a satellite imagery pipeline — each becomes nodes and edges in the same graph, processed by the same agents, assessed against the same threat/risk framework. The more sources, the denser the graph, the richer the correlations, the sharper the assessments.
What Makes It Different
Cross-Domain Fusion
Most tools show data from one domain. Sentinel connects procurement to military flights to news to arms trade to prediction markets in one knowledge graph. A SAM.gov contract automatically links to the company, weapon system, deployment theater, and active conflict.
Assessed Intelligence, Not Data
Every briefing answers “so what?” and “what would change this assessment?” with stated assumptions, denial indicators, and named action indicators. Unfalsifiable assessments are explicitly prohibited.
Threat/Risk Framework
8 threat types (Force Buildup, Procurement Surge, Arms Transfer, Rhetoric Escalation, Industrial Mobilization, Tech Proliferation, Alliance Fracture, Anomalous Pattern) feed into 7 risk categories. Each has posture, trend, and trigger conditions tracked continuously.
Prediction Market Calibration
Polymarket crowd probabilities are compared against Sentinel's signal-based assessments. Divergences — where the system disagrees with the crowd — are flagged as the most valuable intelligence signals.
Below: every data source, analysis method, and its known limitations — full transparency on what Sentinel can and cannot see.
Important Disclaimer
Sentinel uses only publicly available, open-source data. It does not access classified information, restricted government systems, or proprietary databases. Signals reflect what is visible in public records — not the full picture.
Absence of signal does not mean absence of activity.
Intelligence briefings are AI-generated analytical summaries, not verified intelligence products. Classification markings are simulated. Do not use as the sole basis for operational or investment decisions.
Data Sources
ADSB.fi
Military Flight Tracking
Known Limitations
- —Transponder-dependent — aircraft must actively broadcast
- —Stealth aircraft invisible by design
- —Dark flights (transponder off) undetectable
- —Coverage gaps over oceans and remote regions
OpenSky Network
Aircraft Near NATO Ports
Known Limitations
- —Tracks aircraft positions, not vessels
- —Port proximity used as vessel activity proxy — not AIS data
- —Only 5 port areas monitored (Bremerhaven, Gdynia, Constanta, Piraeus, Alexandroupolis)
- —No actual ship identification or classification
SAM.gov
US Federal Contracts
Known Limitations
- —US government contracts only — no allied nation procurement
- —Classified contracts excluded by definition
- —API rate-limited (~100 records/day for backfill)
- —PVI baselines require years of history to be meaningful
TED Europa
EU Procurement
Known Limitations
- —Article 346 TFEU exemptions exclude most sensitive defence contracts
- —Publication delays of days to weeks
- —Less granular than SAM.gov — vendor attribution harder
- —Inherently understates European defence procurement activity
UN Comtrade
Arms Trade Flows
Known Limitations
- —2–3 month data lag — not a real-time signal
- —Arms trade frequently under-reported or mis-classified
- —Aggregated by country, not by company
- —Does not capture government-to-government transfers or aid
GDELT Project
Global Events
Known Limitations
- —Machine translation errors — especially from non-English sources
- —No human curation or editorial validation
- —High noise-to-signal ratio; sentiment scoring is approximate
- —Source quality varies widely across regions
Defense RSS Feeds
Curated News
Known Limitations
- —Coverage limited to English-language Western outlets
- —Editorial framing affects sentiment scores
- —Not comprehensive — significant events may be missed
AI-Defense RSS Feeds
AI-Military News
Known Limitations
- —AI relevance filtering is keyword-based — may miss novel terminology
- —Sentiment scoring is approximate, not contextual
- —Private company coverage depends on public reporting
Civil Aviation RSS Feeds
Airline Disruption Intelligence
Known Limitations
- —Route cancellation news lags actual airline decision by hours to days
- —English-language sources only — regional carrier cancellations may be missed
- —No structured schedule data — relies on news reports of cancellations
- —Cannot distinguish commercial decisions (unprofitable route) from security-driven suspensions without AI classification
SEC EDGAR
AI Company Filings
Known Limitations
- —US-listed companies only — no coverage of private AI companies (Anthropic, OpenAI, Anduril)
- —Defense keyword matching may miss indirect military references
- —Filing text excerpts are truncated — full analysis requires manual review
- —10-K/10-Q are backward-looking; 8-K are timelier but less detailed
UN Comtrade (Semiconductors)
Chip Supply Chain
Known Limitations
- —Same 2-3 month data lag as arms trade Comtrade data
- —HS code classification may not capture all AI-relevant chips
- —Does not distinguish military vs. commercial semiconductor use
- —Country-level aggregation hides company-specific flows
Cloud Status Dashboards
Infrastructure Events
Known Limitations
- —Only covers AWS, Azure, GCP — no coverage of Oracle Cloud, Alibaba Cloud
- —Status pages may underreport or delay disclosure of incidents
- —Region-to-country mapping is approximate for multi-AZ regions
- —Physical attacks (e.g., Bahrain bombing) appear as outages, not as security events
Polymarket
Prediction Markets
Known Limitations
- —Crowd sentiment, not verified intelligence — markets can be wrong
- —Liquidity varies — low-volume markets have unreliable probabilities
- —Geo-restricted in some jurisdictions (fetched via Supabase Edge Function)
- —Only geopolitics and politics tags monitored — other categories excluded
- —Entity matching is keyword-based — some false positives possible
Yahoo Finance
Market Data
Known Limitations
- —v8 chart API — no official support, may change without notice
- —Market cap data unavailable (v6/v7/v10 APIs require paid auth)
- —Non-US exchanges may have delayed data
- —Does not capture pre/post market or dark pool activity
Analysis Methods
Procurement Velocity Index (PVI)
Measures rate-of-change in defence contract awards vs. 12-week rolling baseline. Sole-source ratio applied as urgency multiplier.
Only as good as public contract data. Classified spend invisible. Baseline requires years of history.
Flight Track Computation
ADS-B positions interpolated into tracks. Military aircraft identified by ICAO hex code ranges and callsign patterns.
Transponder-off flights produce gaps. Military hex ranges are incomplete. Misidentification possible.
Anomaly Detection
Z-score and rolling mean deviation applied to port aircraft counts, PVI values, and trade volumes to surface statistical outliers.
Statistical anomaly ≠ operational significance. Low historical baseline inflates anomaly scores.
AI Briefing Generation
Claude (claude-sonnet-4-6) synthesises anomalies and news into structured intelligence assessments every 6 hours.
LLM output is not intelligence. No classified context. Briefings are analytical summaries, not verified reports.
AI Mobilization Index
Weekly composite score per AI company from 4 dimensions: procurement velocity (30%), semiconductor supply chain stress (25%), cloud infrastructure risk (25%), and news sentiment (20%). Normalized to 0–4 scale.
New metric with limited historical baseline. Weights are heuristic, not empirically optimized. Private company coverage limited to news and procurement — no financial filings.
Knowledge Graph
Links companies, contracts, trade flows, SEC filings, infrastructure events, and geopolitical entities into a traversable graph. Enables multi-hop queries like 'show everything connected to Nvidia's military exposure'.
Graph completeness depends on pipeline coverage. Proposed edges from AI analysis require human validation. No causal inference — only correlational links.
Prediction Market Sentiment
Polymarket geopolitics and politics markets are fetched, entity-matched against Sentinel's taxonomy, and displayed as a sentiment layer. Crowd probabilities are compared against Sentinel's proprietary signals to identify divergences.
Prediction markets reflect crowd consensus, not ground truth. Thin markets can be manipulated. Sentinel only monitors markets that match its entity taxonomy — coverage is not exhaustive.
Forecast Agent
Every 6 hours, Claude surveys the prediction landscape, cross-references Polymarket crowd sentiment against Sentinel's proprietary signals (PVI, trade flows, movements, news), and produces independent probability estimates. Highlights divergences where Sentinel disagrees with the crowd and generates Sentinel-originated predictions for scenarios the market hasn't priced.
LLM-generated probability estimates, not actuarial models. No formal calibration or backtesting yet. Confidence tiers (high/medium/low) are self-assessed by the model. Sentinel-originated predictions may lack the crowd validation that market-backed scenarios have.
Situation Memory
Every 3 hours, a dedicated agent reads recent news and updates persistent situation nodes in the knowledge graph. Each situation has a summary (current state) and a chronological timeline (how events unfolded). All intelligence agents consume situation awareness, giving them institutional memory spanning weeks instead of a 7-day rolling window.
Situation identification and merging is AI-judged — may miss subtle distinctions between related events or over-merge separate situations. Timeline entries are compressed summaries, not full article text. Auto-dormancy after 14 days may be too aggressive for slow-burning situations.
Divergence Detection
Every 2 hours, a computational check compares Polymarket crowd probabilities against Sentinel's composite signals for linked entities. When divergence exceeds 20 percentage points, a PREDICTION_DIVERGENCE alert fires and triggers the full forecast agent.
Divergence calculation maps composite scores to implied probabilities heuristically. A high composite score for a company doesn't directly translate to a specific scenario probability. False divergence alerts possible when entity matching is imprecise.
Evolution Brief
Daily strategic synthesis. Reads all briefings, alerts, correlations, news, and situation awareness from the past 7 days. Produces a 2x2 scenario matrix, forward scenarios with probabilities, drift analysis, zombie assumption identification, and priority adjustments.
Weekly cadence means rapidly evolving situations may be assessed with stale framing. Drift scores are heuristic. Zombie assumption identification depends on the corpus containing enough prior assessments to detect stale beliefs.
Signal Thresholds
Procurement Velocity Index (PVI)
Measures the rate-of-change in defence contract awards relative to a rolling 12-week baseline. PVI 1.0 = normal pace.
AI Mobilization Index
Weekly composite of 4 dimensions: procurement (30%), supply chain (25%), infrastructure risk (25%), news sentiment (20%). Scale 0–4.
Live Data Pipeline Status
Real-time counts from the Supabase database. Updated on every page load.
Total rows: 294,422 across 15 tables
Architecture
15 AI Agents — Situation, News, Graph Enrich, Graph Correlate, Graph Patterns, Graph Watchlist, Graph Briefing (Defense/Market/Strategic), Briefing, Mobility, Forecast, Forecast Divergence, Evolution, Anomaly Digest, Risk Assessment
Knowledge Graph — 26 node types, 21 edge types. Companies, states, weapon systems, contracts, conflicts, threats, risks, predictions, situations — all connected and continuously enriched
Threat/Risk Framework — 8 threat types feeding 7 risk categories, each with posture (CRITICAL/ELEVATED/BASELINE), trend, trigger conditions, and denial indicators
Database — Supabase Postgres with RLS, pgvector for semantic search
Frontend — Next.js App Router on Vercel, server components + Supabase Realtime
AI Layer — Claude (claude-sonnet-4-6) for all intelligence production, Intel Chat, and Code Red investigations
Intel Chat — 13 tools: query signals, search web, traverse graph, create/connect/update entities, fire alerts, launch investigations
Code Red — 7-step autonomous investigation workflow: scope → graph walk → temporal → fusion → supply chain → hypotheses → IIR briefing. Auto-triggered on CRITICAL threats
Prediction Markets — Polymarket crowd sentiment, entity-matched, tracked with hourly snapshots, divergences flagged
Maps — Leaflet with live conflict zones, ADS-B flight tracks, and asset layers
Open source — github.com/simseve/sentinel
Built by @simseve
Open-source intelligence for an informed public