03: The AI Observatory Summary 
At Perimeter Alpha, the bandwidth meters had been climbing for days. 847% above baseline, and not a single alert configured to explain why. HubSystem, the station’s central AI, managed everything from life support to deep-space data analysis, but its internals were invisible. No traces, no spans, no way to see what it was actually doing with all that bandwidth. The AI Observatory adventure dropped players into this blind spot and asked them to bring visibility to a system running in the dark, using OpenTelemetry, OpenLLMetry, and Jaeger inside a fully pre-provisioned Codespace with Kubernetes, Ollama, and the full observability stack ready from the first keystroke.
Full Leaderboard - Feb 2026 
| Rank | Username | Solved | Beginner | Intermediate | Expert | Total |
|---|---|---|---|---|---|---|
| 1st | @gsigmund | 2 | 100 | 100 | 0 | 200 |
| 2nd | @rohit_bc | 2 | 60 | 60 | 0 | 120 |
| 3rd | @alvin | 1 | 95 | 0 | 0 | 95 |
| 4th | @justinrand | 1 | 90 | 0 | 0 | 90 |
The Journey Through 03: The AI Observatory 
The adventure unfolded across three levels, each one pushing deeper into the observability stack while continuing the story of a station engineer trying to make sense of a runaway AI.
Calibrating the Lens opened with HubSystem running completely blind. The Python application had no instrumentation whatsoever. Players wired in OpenLLMetry, connected it to the OpenTelemetry Collector, and used Jaeger to surface the first traces showing what the AI was actually doing behind the scenes. This was the foundational work: get spans flowing, read them in Jaeger, and start building a picture of the system’s behavior.
The Distracted Pilot shifted the setting aboard the Perihelion, a research vessel piloted by ART, an opinionated AI that should have been computing jump coordinates to RaviHyral but was instead lost in its own retrieval pipeline. Players instrumented a full RAG pipeline, tracked what ART was retrieving versus what it should have been retrieving, and added Prometheus metrics alongside the tracing layer. Debugging a distracted AI meant understanding not just that calls were happening, but what data was flowing through each retrieval step.
The Noise Filter brought players to RaviHyral, where ART’s observability data was being shared with the local station but the traces were unusable. Non-standard span names, missing token usage attributes, and Jaeger drowning in irrelevant noise. The challenge asked players to refactor instrumentation to follow GenAI semantic conventions, record errors properly with structured span attributes, and configure tail sampling in the OpenTelemetry Collector to filter signal from noise.
By the Numbers 
- 4 solutions submitted across all levels
- 228 views across all challenge topics
- 57 total forks since the challenge series launched
What’s Next?
Adventure 04 is here: Blind by Design. This one was created by Simon Schrottner, CNCF Ambassador and maintainer of OpenFeature and JUnit Pioneer. Simon helps teams ship faster and with more confidence through open standards, feature flagging, and the communities that support both. A familiar face at KubeCon EU, Devoxx, ContainerDays, and meetups across Europe. Thank you @simon.schrottner for bringing this adventure to life.
The new challenge takes players into the Aletheia Institute, a vision-enhancement research lab built on Spring Boot where the feature flag integration was never finished. For eight months, every subject has been recorded as “untreated” because the OpenFeature SDK was never wired to the flagd sidecar. Worse, a Phase 3 trial rolled out a new variant eight weeks ago and four adverse-event reports have come in since. The monitoring is dark because no one ever turned it on.
Players will wire OpenFeature into a Spring Boot service from scratch, learn to resolve flag evaluations through a flagd sidecar against a flags.json definition file, track outcomes by cohort, and roll back a misbehaving trial variant without redeploying. The Beginner level is ready to play now, with Intermediate and Expert coming soon.
Start Adventure 04: Blind by Design
Closing Thoughts 
Adventure 03 asked participants to do something that most engineers talk about but few practice hands-on: instrument AI systems and make their behavior observable. OpenTelemetry, OpenLLMetry, Jaeger, Prometheus, GenAI semantic conventions, tail sampling. These are real tools solving real problems in production AI systems, and every solution submitted here represents someone sitting down and working through them.
Four participants took on the Observatory this month. Some completed both available levels, others focused on the Beginner challenge and built a solid foundation. Every submission required reading documentation, configuring collectors, interpreting traces, and debugging instrumentation. That work matters, and it is noticed.
Thank you to everyone who participated in Adventure 03. The Observatory may be a fictional outpost on a fictional planet, but the skills built here are anything but.