I built a small Python prototype that converts runtime logs into four normalized signals:
- self-call ratio
- concurrency
- external admission rate
- recent-state diversity
It then evaluates the direction of change between time steps against a fixed table of symbolic patterns. For example, rising concurrency combined with falling external admission is labeled as a possible bottleneck.
I tested it on several synthetic traces. One early run produced a false positive because the entropy calculation compared a warmed-up window with an incomplete one. I changed the scanner so both windows must be fully initialized before pattern matching.
I am now exploring whether an LLM could explain these symbolic matches while remaining grounded in the measured values.
My questions:
- Has anyone used an OpenAI model to explain structured telemetry rather than raw logs?
- Would tool calling with a fixed schema be preferable to placing the pattern table directly in context?
- Are there useful evaluation methods for checking whether the model’s explanation stays faithful to the underlying measurements?
4.I would love any feedback.
Minimal code and sample traces:
