Measured. Not estimated.
Every number below comes from running deliverables/benchmark.py with 3 warm-up passes and 7 timed repetitions on Apple M-series hardware. No simulations. No projections.
Measured. Not estimated.
Every number below comes from running deliverables/benchmark.py with 3 warm-up passes and 7 timed repetitions on Apple M-series hardware. No simulations. No projections.
BloomFilter.add() — Time vs. Input Size
Benchmark methodology
All benchmarks were run using deliverables/benchmark.py on an Apple M-series chip (exact model redacted to avoid hardware anchoring). Each benchmark uses 3 warm-up passes followed by 7 timed repetitions. The median of the 7 timed repetitions is reported.
The _save() disk write is patched out of the ConfidenceCalibrator benchmark to isolate mathematical computation cost from I/O variance. All other benchmarks reflect end-to-end wall time.
The BFS graph benchmark uses a degree-4 ring graph as a representative topology. Real-world context memory stores may have different topologies; O(V+E) complexity holds regardless.
LLM inference latency is not benchmarked here — it is network- and provider-dependent. See the LLM Routing section on the home page for observed figures.