How We Gave an AI 18,116 UFO Cases and Built a Self-Learning Observatory

Most AI photo analysis tools work in isolation. You upload an image, the AI looks at it with zero context, and gives you a generic answer. It does not know that 68% of reported aerial phenomena are disk-shaped. It does not know that France has the second-largest observation database in the world. It does not know that 46% of historical reports come from trained observers like pilots and military officers.

We decided to fix that.

The Problem: Blind Analysis

Vigi-Sky is an astronomy and aerial observation platform. Users submit photos of things they see in the sky — unusual lights, formations, objects. Claude AI analyzes each photo.

But every analysis was independent. The AI had no memory. No context. No knowledge of the thousands of cases that came before.

Sound familiar? We had the exact same problem at Strategy Arena, where 14 AI modules were making trading decisions with static prompts and zero memory. We fixed it with a system we call the Living Wiki — a self-evolving knowledge base that feeds context into every AI call.

We exported that technology to Vigi-Sky.

The Data: 18,116 Historical Cases

The Hatch UFO Database is one of the most comprehensive historical observation databases ever compiled. 18,116 documented cases spanning from the year 840 to 2003, each with:

• Location and country
• Date and time
• Detailed description
• Observer type (military, civilian, pilot, scientist)
• Shape classification (disk, cigar, triangle, nightlight, etc.)
• Quality rating

We ran an ETL (Extract-Transform-Load) process to parse every case and build the initial knowledge base.

What the AI Learned

After processing 18,116 observations, the Living Wiki extracted these patterns:

Shapes: - 68.1% are saucers, disks, or spheres — the dominant shape across all decades - 17.1% are nightlights (points of light) — the hardest to identify - 14.8% are cigars or cylinders - 10.3% are triangles, deltas, or boomerangs - 2.8% involve figures or entities

Geography: - USA leads with 7,958 cases (43.9%) - France is second with 2,247 cases (12.4%) - UK and Ireland: 1,185 cases - 20 countries represented in the top tier

Observer Quality: - 46.3% of reports come from high-quality observers (pilots, scientists, military officers) - 26.1% involve military personnel - 92.3% include civilian witnesses

Before vs After

Before (blind analysis):

User: "I saw 3 lights in a triangle formation over Toulouse last night"
AI: "This could be a drone, aircraft, or unidentified object. 
     I cannot determine from the description alone."

After (wiki-enriched analysis):

User: "I saw 3 lights in a triangle formation over Toulouse last night"
AI: "Based on 2,247 historical observations in France and 1,866 
     documented triangle/delta cases worldwide:
     - Triangle formations are the 4th most common shape (10.3%)
     - In urban areas like Toulouse, triangle formations are 
       frequently recreational drones (especially Friday-Sunday evenings)
     - However, 46% of triangle reports come from high-quality observers
     - Recommendation: check Flightradar24 for military activity 
       and ADSB Exchange for drone registrations in the area
     - Credibility score: 62/100 (no photo, single witness, urban area)"

The AI is no longer guessing. It is reasoning from 18,116 data points.

The Architecture

This is the same closed-loop intelligence system we built for Strategy Arena, adapted for observations:

1. Observation Wiki — 18,116 cases parsed into patterns, regional stats, shape distributions, observer credibility data
2. PromptForge (context engine) — before every photo analysis, injects relevant wiki knowledge: regional patterns, similar historical cases, satellite pass times, seasonal trends
3. Component Memory — each analysis is saved and feeds back into the wiki
4. Nightly Evolution — crons mine new patterns, update regional profiles, adjust confidence scores

Every new observation makes the system smarter. The wiki grows. The context gets richer. The analyses get more precise.

What Comes Next

The wiki currently contains knowledge extracted from historical text descriptions. The next steps:

• Real-time satellite correlation — check Celestrak/TLE data to see if the ISS or Starlink was passing at the time of observation
• Aircraft identification — cross-reference with ADSB Exchange for known flight paths
• Weather integration — factor in cloud cover, visibility, atmospheric conditions
• Photo pattern matching — cluster similar photos across the database to detect recurring visual patterns
• User credibility scoring — track each observer's accuracy over time

Open Source

The framework behind this system is ActiveWiki, our open-source Python library for building self-evolving knowledge bases. The same architecture powers both Strategy Arena's 14 AI trading modules and Vigi-Sky's observation analysis.

If you are building an AI application that makes repeated decisions in a specific domain, the pattern is always the same: accumulate data, extract patterns, inject context, learn from results, repeat.

Try It

• Vigi-Sky — submit your own sky observation for AI analysis
• Strategy Arena Living Wiki — see the trading AI brain evolve in real-time
• Evolution Lab — watch 11 nightly engines optimize strategies
• How We Built the Self-Evolving Brain — technical deep dive

Vigi-Sky is a citizen science platform for sky observation. Strategy Arena is an educational trading simulation. Both use the same self-evolving AI architecture to turn raw data into contextual intelligence.