# Historical campaign excerpt

Source: M1-CYC34-LDBC-IS1-OPTIMIZED.md

SHA-256 of preserved original: a4f6b1c2befecaec9a8b8df21883a0d20ef80e803a8e282bb578f41c29fe658b

This selected excerpt preserves the recorded table and its historical commentary. It is not an audited LDBC result. The article distinguishes service configurations, client/transport and returned data; words such as matched or ceiling in the original report are not claims of universal engine performance. The traversal log was not recovered.

## The rows (interleaved ×3, same window, all arms value-verified bad=0)

| row | 8DB (median) | Neo4j (best) | ratio |
|---|---|---|---|
| **Matched per-op** (8 workers, 1 in-flight/worker, UDS vs Bolt) | 218K ops/s (p50 23µs) | 3,410 QPS warm (p50 1.9ms) | **64×** |
| **Best-vs-best** (any config on this box) | 2.01M ops/s (batch=128; window-depressed — quiet-box value 3.21M, M1-CYC31) | 7,068 QPS (4×8 multiproc, warm) | **284×** |

Scaled per-op data point: 231K at c=16 in-window (394K quiet). The sweep
window was contaminated (loadavg 24–32, external + the 4-JVM-client arm
itself) — symmetric across interleaved arms; ours suffers more from load
(CPU-elastic) while Neo4j sits at its own CPU ceiling regardless.