Performance Overhead
Sub-millisecond control-loop latency. Constant CPU and memory footprint. Sealed benchmark receipts from V26 twin-pair runs.
++20.0%%
Whole-system gain
++22.2%%
Kubernetes gain
< 1 ms
Control-loop latency
Sealed Performance Receipts
All results from sealed twin-pair runs. SHA-256 signed and published on GitHub. Numbers shown are live-fetched from the sealed file.
| Metric | Result | Runs | Note |
|---|---|---|---|
| Whole-system gain | ++20.0%% | omni-v26, 97 paired reps | Work, speed, and machines scored |
| Kubernetes scheduler gain | ++22.2%% | omni-v26, 9 tests | One LOSS: fairness −2.6% |
| 99th-pct latency | −11.0%% | omni-v18, Azure AKS, 97 pairs | Tail latency reduction beyond noise |
| GPU energy (A10) | −1.3%% | omni-v23, A10 card, own meter | GAIN all 3 runs |
| Mean latency | −2.7% | omni-v18, Azure AKS, 97 pairs | Inside noise floor |
| Cloud bill | +1.2% | omni-v18, Azure, power given | Inside noise floor — not a cost claim |
Results sealed at omni-v26 (2026-11-12). Numbers are live-fetched from GitHub on every page load. Fallback values shown if GitHub is unreachable. Full methodology at Benchmark Results.
Engine Footprint
Constant overhead regardless of governed node count. Statically linked, zero GC, no heap allocation in the hot path.
| Metric | Value | Note |
|---|---|---|
| Control-loop latency | < 1 ms | Measured at 1 kHz on bare metal |
| CPU overhead | < 2% | Single core, constant regardless of governed nodes |
| Memory footprint | < 250 MB | Static allocation, no GC, no heap in hot path |
| Network overhead | < 0.1% | Metrics polling only; no data plane traffic |
| Startup time | < 3 s | Includes admissibility envelope computation |
| Kill switch latency | < 1 s | From signal to full handback to governed system |
Scaling Behavior
The engine overhead is O(1) with respect to governed node count. This is a consequence of the closed-circle convergence law, not an engineering optimization.
1 node
< 2% CPU
Same as 1,000 nodes
100 nodes
< 2% CPU
Constant — not additive
1,107 nodes
< 2% CPU
Full catalogue. Same budget.
Next in Platform
Kill Switch