From traces to annotations

Select a trace sample from the RL run. Summarize traces for clustering.

Sample traces, discover a hierarchy, then annotate the full corpus 200 of 10,000 illustrative traces are sampled into five uneven clusters, then named. The taxonomy includes a No failures category. All 10,000 traces are classified and aggregated, then their labels are linked back to the original traces in AC2 in one batch. Icons are schematic; counts show illustrative scale. Full corpus DISCOVERY SAMPLE TracesFailuresAgentEnvironment 10,000 traces · 200 sampled Discovery sample
0:00 / 0:16

Illustrative scale: 10,000 traces · 200 sampled. Tiles and rows show representative traces.