Verification Methodology
How Sentinel Review assesses, labels, and presents conflict events. Every claim below describes the system as it currently runs in production; figures are pulled from the live pipeline configuration.
Confidence Levels
Corroborated by at least two independent sources from different platforms or organisations, plus at least one of: geolocated footage, official acknowledgment, or matching wire-service report.
Multiple sources exist but all are from the same platform, or a single tier-1 source is paired with one corroborating circumstantial signal. The event likely occurred but cannot be fully confirmed.
A single source, or multiple sources tracing back to a common origin (e.g. all citing the same Telegram channel). Published for awareness; treat all unconfirmed events with significant caution.
Source Trust Tiers
Every source in the system is assigned a trust tier. Tiers affect how much weight a source's posts carry in confidence scoring — a tier-1 source posting alone can reach partial confidence; a tier-3 source posting alone stays unconfirmed.
Government and intergovernmental wire feeds, peer-reviewed conflict research projects, and verified investigative accounts with established geolocation track records. A tier-1 source posting alone can reach partial confidence. Live examples: ISW (Institute for the Study of War), EUCOM, INDOPACOM, OFAC, The Irrawaddy.
Established mapping or conflict accounts with a good verification track record but less transparent methodology, regional press with known editorial stances, and community geolocation projects. Require corroboration before reaching verified.
Anonymous accounts, state-affiliated media, single-contributor Telegram milblogs, and sources with known track records of amplifying unverified or false information. High volume but low signal. Never sufficient alone for any confidence level above unconfirmed.
Event Processing Pipeline
Posts are collected from curated source feeds (X, Telegram public channels, RSS, wire services, GDELT, Bluesky) on a 30-minute schedule. Every post is stored verbatim in an append-only log and never mutated.
An LLM (Anthropic Claude) analyzes each post and extracts: event type, location, approximate time, actor (if identifiable), and a brief description. Posts without a discrete conflict event are discarded. Extraction output is structured JSON — the model cannot free-text outside the schema.
Candidate events within 5 km and a 6-hour window of each other are clustered as potential duplicates. A second LLM pass decides whether two candidates describe the same event or distinct events. Merged candidates inherit all source links.
Each event is scored using the rules above. Source trust tiers are applied as weights. The confidence label shown on the map is deterministic from the rule set — not a probabilistic output of a model.
Events that pass scoring are published to the live map automatically — they are not held for human approval. They appear within the pipeline's processing latency (typically 30–60 minutes after the source post), each carrying its confidence label and full source list so readers can weigh it directly. Events are not individually human-verified.
Content and Graphic Material
Sentinel Review does not embed media directly. Source links open in a new tab and may contain graphic content including footage of combat, casualties, and infrastructure destruction. Source posts are retained verbatim in the underlying data store for archival continuity, but are surfaced on the event detail page only behind explicit user action.
Errors and Corrections
If you believe an event is incorrectly located, mislabeled, or based on fabricated sources, email corrections directly to the editorial address below. Confirmed corrections are applied and recorded openly in the event's change history.
Confidence labels can only move down, not up, once a credible dispute has been filed against an event.
Disclaimer
This platform is a situational awareness tool only. It does not support military targeting or operational planning. Events are algorithmically extracted from open-source reporting and assigned a confidence level based on cross-source corroboration; they are not human-verified and may contain inaccuracies.