Local models. Explicit contracts. Measurable trade-offs. Explore the engineering evidence behind AI systems that have to work beyond the demo.
Inspect every input, output, and decision in a fixed local-model evaluation.
The decision
A useful evaluation separates supported configuration facts from claims about actual runtime incidents. The next study should predeclare its revised interface and use independently authored cases.
Read the full evaluation →Inspect the method
The baseline implements the same rules used to construct these labels. No result here establishes production performance, broad NiFi compatibility, or the ability of frontier models.
What survives when two sampling policies receive exactly the same byte budget?
Retention, not model accuracy
The experiment boundary
| Policy | Bytes | USD |
|---|
Illustrative $2 / decimal GB, $0 fixed fee. Not a vendor price or a customer savings measurement.
A transparent monthly sensitivity calculation. This changes cost projections only; it does not rerun sampling or predict evidence retention on your workload.
Operational question
Rare-event recall and complete trace retention measure different things. Retaining isolated anchors does not guarantee the context needed to understand what happened. The published uniform sampler retained no complete rare traces in these 20 trials.
Reproduce the experiment →Next experiment
Test incorrect tags, unknown failure families, variable event sizes, priority overload, and the cost of attaching tags to earlier context. No result here establishes measured customer ROI.