War Monitor

The problem with conflict data

Every open conflict feed carries the same defect: a single event reported by five outlets that all sourced it from one wire becomes five confirmations. Count naively and a rumour outranks a verified strike.

War Monitor treats that as the central problem rather than a cleanup step. Every event carries an independence-weighted trust score computed before it is eligible to render. Correlated sources collapse toward the weight of one; genuinely independent corroboration compounds. The globe only draws what clears the gate, so the map is a claim about confidence, not a claim about volume.

A map that plots everything reported is a map of how loud the reporting was. The gate is the difference between a feed and an assessment.

Architecture

A TypeScript monorepo of two apps and eight packages, orchestrated by Turborepo. apps/web is Next.js 15 on React 19, apps/api is Fastify 5, packages/db is Prisma 6 over PostgreSQL 16 across 46 models — plus a raw-managed PostGIS geom column and a pgvector event_embeddings table that sit outside the schema on purpose, behind a capability guard. packages/shared holds the types both ends compile against, and the typed API client is regenerated from the OpenAPI hash — CI runs the generator in check mode, so a hand-edited or stale client fails the build.

Migrations do not run through prisma migrate. A checksummed engine in packages/db/scripts journals a two-layer manifest of 47 entries, declares the database capabilities it needs (pgcrypto, PostGIS, pgvector), and refuses out-of-order application.

Ingest runs against six first-class sources — GDELT, UCDP, GDACS, ReliefWeb, Telegram and ACLED — plus a generic RSS adapter contract for the long tail. Every source name resolves to a family (a wire, a primary outlet, a dataset), and the trust spine uses distinct families, not report counts, to decide how much a second report is actually worth.

The four systems of the 2.0 rebuild

The trust spine (apps/api/src/services/trust/score.ts) is pure and deterministic: one source family caps the score at 25, two score 50, three 70, four or more 85, a primary outlet adds 12 — and three or more reports that collapse to a single family are flagged echo-chamber and capped at 20. The discrete ladder behind it reserves CONFIRMED for high-severity events with a primary-family corroboration, and golden tests pin the bands so a tuning change is a reviewed diff, not a vibe.

The ingest layer normalises six inconsistent source schemas into one event shape, then deduplicates on Unicode-aware token similarity inside geo-and-time blocks — a fix that exists because Arabic and Cyrillic titles once tokenised to nothing and could never corroborate. Corroboration is monotonic: late-arriving sources merge into the stored event, trust is recomputed over the union, and an event’s status only ever rises.

The globe renders scored events on CesiumJS, code-split into a lazy chunk that the 300 KB first-load bundle budget explicitly excludes, with a time scrubber for replay and horizon culling at altitude. The console is the authenticated analyst surface — review queues, source management, and adjudications that are themselves stored as append-only records freezing the evidence as the reviewer saw it, with a compare-and-set on status so two analysts cannot silently overwrite each other. Analyst-facing AI — briefings, entity extraction, natural-language query — runs through a local model gateway behind a CI-enforced local-only policy, with golden release evals and quality budgets as gates.

Verification

Structural counts in the documentation are generated, not typed: a script walks the codebase and writes docs/generated/ARCHITECTURE-STATS.md — 57 API route modules, 264 service modules, 46 Prisma models, 731 test files holding 5,983 test declarations, 23 architecture decision records — and CI fails if the committed file drifts from what the code actually contains. The same pipeline enforces gates on documentation, configuration, roadmap, migrations, policy, performance contracts, types, tests, build and security.

That habit — making the docs a build artifact rather than a promise — is the part of this project I have reused most everywhere else.