Measured evidence, separated by workload and build lineage.
Persistara publishes measured results with explicit test boundaries. Newly completed v1.6.10 blind, live-data, concurrency and recovery tests appear first; earlier Databricks benchmarks remain separately attributed.
Frozen Enterprise v1.6.10: evidence on unseen inputs and failure recovery.
The following results were measured on the same frozen v1.6.10 product build in self-run engineering evaluations. They are not independent third-party certification, Databricks production benchmarks or promised customer-wide savings.
Exact REUSE / REGROUND / HOLD action matches across 13 predefined scenario classes. Zero observed false-safe reuse, false invalidation or engine errors.
Stage 01; precommitted oracle; newly generated corpus at runtime. Internally designed and executed test, not independent lab certification.
Exact actions on 3,500 distinct package transitions, including 177 material support changes. 3,323 REUSE and 177 REGROUND decisions; zero observed safety errors.
EBV-02 v1.1; precommitted future public inputs and external randomness. 94.943% counterfactual recomputations avoided on this observed mix; no live lifecycle-hook change occurred.
DS-01 completed 310,000 / 310,000 resolves with zero bad statuses, cross-tenant leakage or engine errors, alongside 744 / 744 mixed-action checks.
One Windows host, 12 logical CPUs; isolated partitions across processes. Demonstrates local process scaling, not distributed multi-node scale-out.
Three live-data workloads, each with its own recomputation rate.
981 correct REUSE and 19 correct REGROUND decisions from 1,000 scored repeat updates; 0 observed false-safe reuse, 0 false invalidations, and 357 / 357 scheduled revalidations completed.
Frozen v1.6.10; synthetic derived state evaluated against real public change events. Not a finance, legal or compliance decision.
538 correct REUSE and 462 correct REGROUND decisions across 1,000 future public readings; 1,000 / 1,000 exact oracle matches, with zero observed false-safe reuse or false invalidations.
Future external input and a public randomness seed; self-run live-blind test, not flood-safety certification or independent third-party testing.
3,323 safe REUSE decisions and 177 necessary REGROUND decisions across 3,500 future package transitions. Exact decision agreement: 3,500 / 3,500.
Support predicates were predeclared. This does not certify software package safety or imply unchanged package source code.
Safety held under local load and deliberate database faults.
50,000 / 50,000 calls completed
All 50,000 local threaded calls completed, with zero bad statuses, deadlocks, timeouts or engine errors. Across all measured phases, p99 was 0.4911 ms and p99.9 was 1.0995 ms.
Engineering load characterization on a local machine; not a production application SLA or Databricks throughput result.
20 / 20 injected fault cycles recovered
400 / 400 logical events durably reconciled, including 20 / 20 duplicate replays without duplicate downstream effects. HOLD outcomes produced zero downstream effects; no unrecovered persistence errors.
Controlled fault injection against one Docker PostgreSQL instance; not multi-region disaster recovery or production RTO evidence.
41,773 resolves/s at eight workers
Aggregate throughput rose from 11,051 resolves/s at one worker to 41,773 at eight, with p99 of 0.2811 ms. At 16 workers, throughput plateaued near 43,793 resolves/s and p99.9 rose to 34.94 ms.
One host with 12 logical CPUs. The 16-worker configuration was deliberately oversubscribed; the tail-latency degradation is retained in the evidence.
Recorded outcomes, not universal guarantees
These tests report the observed behaviour of their exact workloads and environments. Detailed protocol, hashes and raw evidence are reserved for controlled technical diligence.
No outcome from the separate ongoing 72-hour open-system evaluation is claimed here. No external laboratory, customer production certification or independent penetration test is implied.
Earlier live Databricks and mixed-benchmark findings
The historical results below were obtained on their disclosed earlier validated product builds and workloads. They are not relabelled as v1.6.10 runs.
978 of 1,000 control-path recomputations avoided in a live Companies House benchmark, with 0 observed false-safe reuse and 0 false invalidations.
Historical live benchmark on an earlier validated v1.4.2-line build. Workload-specific.
43 of 60 control-path model calls avoided in a live TfL + Databricks benchmark. Measured tokens fell 71.5%; estimated model cost fell ~72.0% in that benchmark.
Test15 matched every expected decision with 0 observed false-safe reuse and safely avoided full recomputation on 70.0% of cases.
Companies House
The control path recomputed on every scored repeat update. This percentage must not be extrapolated to all AI traffic, tokens or customer spend.
TfL + Databricks
Real TfL source data and a real Databricks Lakebase trial workspace; the benchmark validity boundary was synthetic.
Test15
Test20
The 30-call baseline intentionally contained unchanged-source REUSE cases. It demonstrates avoidable work when reuse is valid, not a 100% workload-wide savings forecast.
Test21 completed 1,500 live events with 1,042 safety-approved REUSE skips, all 182 required REGROUND calls completed, and 0 final App/model/safety errors.
A separate standard-library process reproduced 20,000/20,000 decisions with 0 mismatches and loaded 0 Persistara runtime modules. Modification and record reordering were detected.
Two persisted evaluation passes produced deterministic action-correctness and no-false-safe scorers true on every row.
Fail closed, recover deliberately
Enterprise release validation
Controlled engineering measurements, not independent certification or universal production SLOs.