Empirical Benchmarks & System Topography

Sub-Millisecond Guardrails.
Zero Hot-Path Overhead.

Measured under rigorous high-frequency load on dedicated hardware testbeds: 100,000+ operations/sec per node with p99 runtime latency under 0.38 milliseconds.

p50 Runtime Intercept

0.18 ms

180 microseconds (PyO3 Rust FFI)

p90 Runtime Intercept

0.28 ms

280 microseconds

p99 Runtime Intercept

0.38 ms

380 microseconds (tail latency)

Layer 1 AST Full Scan

18.4 ms

50,000 LOC / 120 AST rules

Benchmark Parameters & Testbed Topography

Why Anchor Achieves < 0.4ms Execution Latency

Unlike LLM proxy wrappers that route payloads through external network hops or invoke secondary language models, Anchor executes in-process via zero-copy PyO3 Rust FFI pointer passing. When decorated with @anchor.guard, input arguments are evaluated directly against pre-compiled AST invariants in memory.

Cloud Instance

AWS c6i.4xlarge (Dedicated Compute)

Processor

Intel Xeon Platinum 8375C @ 2.90GHz (16 vCPUs)

Memory

32 GB DDR4-3200MHz ECC

Operating System

Ubuntu 22.04 LTS (Linux Kernel 6.5 x86_64)

Sample Size

1,000,000 synthetic payload iterations

Execution Layer

In-Process Zero-Copy PyO3 FFI pointer passing

Execution Overhead Comparison

Inference Latency Impact (Lower is Better)

Comparing Anchor in-process Rust FFI against network sidecars and secondary LLM judges.

Anchor Layer 2 (In-Process PyO3 Rust FFI)0.35 ms
LangSmith Async Telemetry Sink85.00 ms
Guardrails AI (Python Regex & Wrapper)180.00 ms
NeMo Guardrails (Secondary LLM Judge)240.00 ms
Full Secondary LLM Review (e.g. GPT-4o-mini)380.00 ms

Federated Network Topology

How Anchor Connects: Spoke to Hub

Raw AI data never leaves the client's sovereign spoke node. Only cryptographically hashed Ed25519 state packets synchronize with the Hub.

01

Client Execution Node

AI agent models, pipelines, and tools running in client VPC / on-premise infrastructure.

@anchor.guard hook

02

Anchor Core Engine

Zero-copy PyO3 Rust kernel evaluates AST parameters in < 0.35ms.

Deterministic verdict

03

Local DAC Storage

Appends cryptographic SHA-256 chain-hash to local tamper-evident disk ledger.

Immutable proof

04

Spoke Telemetry Relay

Async queue transmits zero-knowledge Ed25519 signed state hashes to the Hub.

Non-blocking sync

05

Enterprise Hub & Oversight

Centralized dashboard, dual-key auditor pulls, and automated statutory dialect translation.

Dual-key portal

Memory Footprint

< 25 MB

Resident Set Size (RSS) per spoke node process during sustained 10K req/sec load.

CPU Utilization

< 1.2% Core

Zero-copy byte serialization avoids Python GIL bottlenecks and GC pauses.

Throughput Scaling

100K+ ops/s

Single node throughput capacity without scaling sidecar proxies.