Live evidence

Live changing data. Controlled AI state reuse.

Persistara was exercised against real, changing Transport for London arrival predictions with source state and decision evidence stored in Databricks Lakebase.

60live scored revisions
43PRESERVE decisions
17natural REGROUND decisions
71.7%control-path model calls avoided in this benchmark

The problem being tested

Persistent AI workflows can keep state after the source data that supported it has changed. Reusing everything risks stale state; recomputing everything can waste model calls, tokens and latency.

What Persistara did

For each scored source revision, Persistara evaluated whether the change remained inside a verified state-preserving transition or crossed a material boundary. It selected PRESERVE for 43 events and REGROUND for 17 events. The run recorded zero false-safe reuse and zero false invalidations.

Why the result matters

Persistara did not simply maximise reuse. It preserved state when the benchmark validity boundary remained intact and forced re-grounding when the real dependency crossed that boundary.

Scope and limitations

This was a designed benchmark using real live external data, not an independent customer production deployment. The ARRIVING/WAITING boundary was synthetic benchmark logic chosen so decisions could be audited event-by-event. The integration demonstrated Databricks Lakebase; it did not claim Databricks certification, endorsement or Managed Agent Memory integration.

Open the one-page PDF case study →