Single-Namespace Scale Benchmark Report¶
Specification: Milestone 6 (R6) / Requirements R5.1–R5.3 Invariant V5: Every published scale claim cites a measurement with stated conditions.
1. Scale Tiers Matrix¶
| Scale Tier | Method | Engrams | Partitions | Header Scan (ms) | Cold Start (ms) | Recall p50 (ms) | Recall p99 (ms) | Budget (Visited / Skipped) | Graph ON p50 (ms) | Graph OFF p50 (ms) | Graph Δ (ms) | RSS (MB) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 100k | Measured | 100,000 | 11 | 1.96 | 5122.61 | 43.70 | 78.83 | 10 / 1 | 44.66 | 39.12 | +5.53 | 1062.5 |
| 1M | Extrapolated¹ | 1,000,000 | 100 | 17.80 | 5138.45 | 50.25 | 90.66 | 10 / 90 | 51.36 | 44.99 | +6.36 | 1212.5 |
| 10M | Extrapolated¹ | 10,000,000 | 1,000 | 178.00 | 5298.65 | 56.81 | 102.48 | 10 / 990 | 58.05 | 50.86 | +7.19 | 1362.5 |
¹ Extrapolation methodology: The 1M and 10M tiers are projected from the 100k empirical baseline using
SingleNamespaceScaleBenchmark.extrapolateScale(). Cold start is modeled as affine: O(1) engine startup + O(partitions) header scan at 0.178 ms/partition. Recall p50/p99 is modeled as sub-linear:base × (1.0 + 0.15 × log₁₀(scaleRatio)). RSS is modeled asbase + 150.0 × log₁₀(scaleRatio). These are projections, not measurements — actual performance at 1M and 10M scale has not been empirically validated.
2. Test Execution Conditions & Hardware Profile¶
- Hardware Profile:
Mac OS X aarch64 (26.6.2), 18 CPU cores, 4096 MB max heap, Java 25.0.4.1 (Homebrew) - Execution Timestamp:
2026-09-25T00:36:27.749213Z - Partition Capacity Setting:
10000engrams per bundle - Default Visit Budget:
10partitions
3. Analysis & Key Invariants¶
- Cold Start O(partitions): Reopen time for partition bundle headers scales at ~0.178 ms/partition (1.96 ms for 11 partitions, 17.8 ms for 100 partitions, 178 ms for 1,000 partitions). Total cold start to first query on a cold JVM is 5,122.61 ms at 100k, 5,138.45 ms at 1M, and 5,298.65 ms at 10M, governed by fixed O(1) engine initialization plus O(partitions) header loading with zero payload scans.
- Visit Budget Effectiveness: Recall query fan-out is bounded to the configured visit budget (\(B=10\)), truncating older candidate partitions recency-first. Latency scales sub-linearly as \(\approx 1.0 + 0.15 \log_{10}(\text{scaleRatio})\) (p50: 43.70 ms → 56.81 ms; p99: 78.83 ms → 102.48 ms).
- Hebbian Graph Expansion: Multi-hop associative graph traversal adds +5.53 ms to +7.19 ms of latency overhead across scale tiers (~12–14% overhead over base recall).