Benchmark evidence

One clock for each claim.

Historical internal measurements are reported with their workload and boundaries. Customer acceptance still requires matched, independently reproducible testing.

The available material contains promising internal measurements and later notes that some reruns failed. The figures below preserve their original scope. They do not establish current production performance or a matched comparison with all competing systems.

Reported resultMeasurement scopeInterpretation
13.834 µs p50; 24.875 µs p99Isolated FHRR recall; 1,800 trials; 12 June 2026Historical internal lookup timing, excluding a complete generated answer
100 / 100 successful trialsSeven distinct role bindings; 16 June 2026Bounded test; Wilson 95% lower confidence bound approximately 96.3%
292,705 queries / secondCPU E8-LSM benchmark; batch 32; 10 August 2026Historical peak batched throughput; not network-wide or GPU throughput
47.8 µs p50CPU E8-LSM benchmarkSeparate path and timing scope from FHRR recall
0.876 ms tick; 0 / 826 over budgetInternal loop with 33.333 ms budget; June 2026Loop timing, not transaction finality or model response time
130:1 raw-data compressionCompany-reported Phoenix workloadCorpus, full encoded-byte accounting and raw execution logs needed

What can be said today

The historical evidence supports a development thesis around microsecond-scale local lookup and potentially substantial compression on suitable raw data. It does not support an unrestricted claim that the platform is faster or more accurate than every competitor across every benchmark.

The supplied business-plan material references later rerun failures and does not include the complete raw artifacts needed to resolve them. A release-specific replication is therefore part of the seed program. Each published result should carry a code revision, hardware configuration, corpus identifier and evaluation script.

One clock for each claim

Local lookup, batched query throughput, network round-trip time, retrieval-plus-verification and complete answer generation are separate metrics. The plan does not compare a microsecond memory operation to another product’s full API response and call the ratio a system speedup. The commercial benchmark is the customer’s complete task under a matched operating envelope.


$20 million is an ask. No customer logos yet. Forecasts are a plan.