StarCodex uses AI Codex to automate Starlink risk detection

Starlink and other LEO satellite broadband systems generate massive, diverse measurement data across regions and time. However, converting this data into reusable experimental evidence still relies on manual code development and expert-guided inspection—a bottleneck for network research. StarCodex, proposed by Luo et al., tackles this by acting as a dynamic coding harness. It first detects analysis gaps in the current measurement state, then converts those gaps into structured coding tasks. Using OpenAI's Codex, it generates or repairs executable analysis artifacts (code, scripts) and validates them through four layers: code correctness, data-interface compatibility, measurement-semantic consistency, and output plausibility. This creates a closed-loop automation pipeline that adapts as new data streams in.

In experiments on real Starlink measurements, StarCodex discovered 49 out of 56 system-risk cases that were previously uncovered by manual analysis. It also attained higher average precision than the strongest predefined analysis baseline. Beyond risk detection, StarCodex constructed a benchmark with denser and broader system-risk evidence, enabling more thorough evaluation. The generated prediction and replay artifacts further revealed prediction risks and significant quality-of-experience (QoE) differences among adaptive bitrate (ABR) controllers. These results demonstrate that a dynamic coding harness can convert evolving satellite measurements into validated, reusable artifacts—automating experiment workflows that previously required extensive human effort.