Docs
Evaluation
The June 15, 2026 T3-lite evaluation measured identity connections and name-based retrieval on a synthetic corpus of 60,610 entities. The results below apply to that workload.
Workload and scope
- Evaluation date: June 15, 2026.
- Corpus: 60,610 synthetic entities.
- Identity recall and identity precision: 124,392 verified identity pairs on that corpus.
- Recall@10 and MRR: 20 name-lookup queries.
- Measurements ran locally through in-process handlers. They are not measurements of deployed-service latency or throughput.
Retrieval and connection quality
Recall@10 is the share of queries whose target identity appeared in the top 10 results. MRR is the mean reciprocal rank of the correct identity. Identity recall is the share of verified identity pairs recovered; identity precision is the share of returned identity connections that matched the verified pair set.
The evaluation found one pair of same-type entities whose similarity was strong enough to be treated as an identity connection even though the verified data did not identify them as the same entity. Some additional correct shared-reference connections counted against identity precision because they were not listed in the verified identity-pair set.
Multi-hop reasoning was measured only as retrieval of supporting evidence without a judge model and remained partial. A five-minute reliability check completed 123 operations with zero errors; that check does not establish throughput or a service guarantee.
Model usage
Match retrieval, relationship computation, graph exploration, and history in this evaluated workload ran without language-model or embedding API calls. Incoming classification and property extraction are separate model-assisted operations that use the configured language-model provider. That measurement does not describe answer-mode search, which uses a model to compose a response and requires explicit mode: "answer". Avoiding model API calls for those kernel operations does not mean they have zero total operating cost.