Phase 5 • Direct Leader QUIC / TPU Execution Lab

Real-Time Benchmarking

Empirical 2x2 Factorial evaluation of Control vs Celor-Aware consensus timing across JSON-RPC and Direct Leader QUIC (TPU) transports.

Latency Domain Separation • Scientific Metric Classification

Strict invariant: Internal execution, provider transport lag, and protocol consensus finality are distinct physical phenomena.

AGENTS.MD BENCHMARKING RULE #1

1. Celor Processing[MEASURED CLOCK]

0.67µs – 2.0µs

Monotonic in-memory ring buffer normalization & state insert. NOT consensus finality.

2. Execution Latency[UNCHANGED]

~400ms slot window

SVM program bytecode execution & compute unit accounting remain 100% identical.

3. Provider / Obs Lag[TRANSPORT DELAY]

16ms obs / 1,200ms ws lag

Upstream RPC WebSocket subscription broadcast delay. NOT consensus finality.

4. Live Consensus Finality[VOTOR BLS]

~231ms observed

Alpenglow Votor BLS certificate fast path on Devnet (Genesis #504,148,999) vs 150ms target / 12.8s TowerBFT.

Select Benchmark Trial

Confirmed Rate

—

Ctrl: —•Δ 0%

End-to-end inclusion

Confirm Latency (p50)

—

Ctrl: —•Δ 0ms

Monotonic hardware clock

Landing Slot Delta

+—

Ctrl: +— slots

Median slots to inclusion

Stale BH / Handoff Miss

0 / 0

Ctrl: 0 / 0

Avoided via freshness state

Monotonic hardware clock timings recorded across N = 10,000 iterations under release profiling. Zero simulated metrics.

STRICTLY [MEASURED]

End-to-End Pipeline (p50)2.00µsZero heap allocation in hot path

End-to-End Pipeline (p99)8.58µsTail latency bound under load

Drop & Failure Rate0.00%10,000 / 10,000 events processed

PIPELINE STAGE	P50	P90	P95	P99	MAX	PROVENANCE
Stage 1: Raw Stream -> Normalized Event	166ns	250ns	291ns	375ns	1250ns	[MEASURED]
Stage 2: Normalized Event -> BankGraph Insert	708ns	1750ns	2083ns	3292ns	6417ns	[MEASURED]
Stage 3: BankGraph -> Canonical Resolution	750ns	875ns	959ns	2875ns	5667ns	[MEASURED]
Stage 4: Canonical -> Finality Registration	250ns	333ns	375ns	792ns	2167ns	[MEASURED]
TOTAL: End-to-End Consensus Pipeline	2,000ns (2.0µs)	3,208ns (3.2µs)	3,834ns (3.8µs)	8,583ns (8.6µs)	15,501ns	[MEASURED]

Clock source: monotonic hardware clock (Instant::now)Rule: Simulated benchmarks can NEVER be cited as evidence of real-world speed