LI-ESKG integrates uncertain entity resolution with an authoritative event-state knowledge graph. The implementation separates three planes:
- an active, allocation-reusing probabilistic workspace;
- an append-only resolution ledger containing immutable evidence, inference, decisions, revisions, dependencies, outbox entries, and receipts;
- a host-only graph containing accepted native event-state relations.
The typed licorne extension exposes V2 observations, collective inference,
dynamic client-defined candidate/factor providers, the hot workspace and
storage contracts, host-graph materialization, standards projection, and a
bounded asynchronous resolver. Its pooled single-tenant scheduler keeps one
ledger and logical runtime per tenant while sharing only scrubbed workspace
allocations and a fixed execution-thread set. H3 is available only as an
optional provider index; it has no role in default resolution semantics.
See the Python guide for installation and an asynchronous create-then-match example.
Benchmarked on a M2 Apple chip.
| Benchmark | Scale | Execution Time | Throughput |
|---|---|---|---|
| Observation Partition | 500,000 obs | 69.8 ms | 7.16 Melem/s |
| Identity Uniqueness | 400,000 nodes | 2.30 ms | 173.77 Melem/s |
| Causal Acyclicity | 300,000 rels | 3.95 ms | 76.05 Melem/s |
| Batch Insertion | 1,000,000 items | 76.4 µs | 13.09 Gelem/s |
| Active Belief Query | 1,000,000 items | 4.37 ms | 228.78 Melem/s |
| Runtime Pipeline | 10,000x10 | 167 ms | 598.8 Kelem/s |