Live Timeline — Ukraine War Strategic Analysis

How This System Works

This live timeline is powered by a dedicated Cloudflare Worker (isolated from the Iran War worker — separate D1 database, KV namespace, and Discord webhooks) running on a cron trigger every 2 hours. The complete pipeline:

RSS Ingestion: The Worker fetches RSS feeds from 12 major news outlets in parallel. Only articles published within the last 48 hours are considered.
Relevance Filtering: Articles are filtered against 100+ Ukraine-war-specific keywords (Ukraine, Russia, Kyiv, FPV, Shahed, HIMARS, NATO, Kursk, etc.) using pure string matching — no AI cost.
Priority Scoring: Relevant articles receive a composite score that determines processing order. The default threshold is 8 — articles below it are skipped. Scoring factors:
- +10 — Action verb + Ukraine/Russia party in title (e.g., "Ukraine strikes Russian depot", "Russia advances near Pokrovsk")
- +8 — Critical entity in title (FPV, Shahed, HIMARS, ATACMS, Kursk, Crimea, Putin, Zelensky, Bakhmut, Pokrovsk, NATO, nuclear, ceasefire)
- +5 — Wire service origin (Reuters, AP, BBC)
- +5 — Published within last 2 hours (+2 if within 6 hours)
- +3 — "Ukraine" / "Ukrainian" in title
- +2 — "Russia" / "Russian" in title
- +2 — Substantive description (over 200 characters)
- -5 — Opinion/editorial signals in title
- ±N — Per-source adjustment (configurable, e.g., paywalled sources get a negative base score)
Example: A Reuters article titled "Ukraine HIMARS strike destroys Russian depot near Bakhmut" published 1 hour ago scores: +10 (action+party) +8 (HIMARS+Bakhmut) +5 (wire) +5 (recent) +3 (Ukraine) +0 (Reuters adj.) = 31. An NYT opinion piece about a potential peace plan from 8 hours ago scores: +3 (Ukraine) -5 (opinion) -3 (NYT adj.) = -5 (skipped).
Deduplication: New articles are compared against the last 48 hours of stored events using Jaccard similarity on word sets (threshold: 0.65). This prevents duplicate entries without consuming neurons.
AI Summarization: Unique, high-priority articles are summarized by the configured primary AI model, with automatic fallback if it fails. The primary and fallback models are configurable from the admin dashboard. Available models include local Ollama (zero cost), Cloudflare Workers AI BART-Large-CNN (150 neurons), and Claude Haiku/Sonnet/Opus (API credits).
Strategic Impact: As part of the same AI summarization call, the model generates a one-sentence strategic impact assessment for each article — explaining what the event means in the broader geopolitical, military, economic, or humanitarian context. If the article has no meaningful strategic significance (e.g., human interest or off-topic), the model marks it as "None" and no impact is displayed. This runs at zero additional cost since it is part of the summarization prompt.
Category Classification: Events are classified into 6 categories (Military, Diplomatic, Economic, Humanitarian, Cyber, Political) using keyword scoring rules — no AI cost.
D1 Storage: Events are written to a Cloudflare D1 (SQLite) database as immutable records. Corrections create new records that reference originals. In rare cases, entries produced by coding errors (e.g., raw text posted without AI summarization) may be removed to maintain timeline quality.
Cache Layer: API responses are cached in Cloudflare KV for 2 minutes to reduce D1 reads on high-traffic pages.

AI Models

Summarization + Strategic Impact: Configurable primary/fallback model chain. Available options include Qwen 3.5 9B (local Ollama, 0 neurons), BART-Large-CNN (Cloudflare Workers AI, 150 neurons), and Claude Haiku/Sonnet/Opus (Anthropic API). The primary and fallback models can be changed from the admin dashboard without redeployment.

Fact-Checking: On-demand article reviews via Qwen 3.5 27B (local Ollama) or Claude Sonnet/Opus (Anthropic API). Selectable per-review from the admin dashboard.

Date Validation: Summaries are automatically scanned for dates outside a 48-hour window around ingestion time. Articles with suspicious dates are auto-flagged for manual review.

Review & Correction System

Every event can be independently fact-checked by Qwen 3.5 27B via the admin dashboard. Two review modes are available:

Summary vs Source — compares the AI summary against the original RSS article text for hallucinations, omissions, and misattributions
External Verification — assesses factual plausibility based on the reviewer model's knowledge

If issues are found, the reviewer proposes a corrected summary. An admin reviews the diff and either approves (publishes the correction) or dismisses (marks as verified). No correction is published without explicit human approval. Review status is shown on each event:

✓ Verified — fact-checked and confirmed accurate
⚠ Flagged — potential issues identified by AI reviewer
C Corrected — reviewed, corrected, and republished

RSS News Sources

The Worker monitors these feeds every 2 hours:

Adj.	Source
0	Al Jazeera
0	BBC World Europe
0	BBC World
0	NPR World
-3*	NYT World
0	Reuters World
0	Guardian Ukraine
0	Guardian Russia
0	France 24 Europe
0	Deutsche Welle
-3*	Washington Post
-3*	WSJ / Dow Jones

Adj. = base priority score adjustment applied to all articles from this source.
* Paywalled — most articles require paid subscription to read full text, deprioritized by default.

Filter:

Connecting to live feed...

Live Timeline: Ongoing Developments