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
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
Comparing Anchor in-process Rust FFI against network sidecars and secondary LLM judges.
Federated Network Topology
Raw AI data never leaves the client's sovereign spoke node. Only cryptographically hashed Ed25519 state packets synchronize with the Hub.
AI agent models, pipelines, and tools running in client VPC / on-premise infrastructure.
@anchor.guard hook
Zero-copy PyO3 Rust kernel evaluates AST parameters in < 0.35ms.
Deterministic verdict
Appends cryptographic SHA-256 chain-hash to local tamper-evident disk ledger.
Immutable proof
Async queue transmits zero-knowledge Ed25519 signed state hashes to the Hub.
Non-blocking sync
Centralized dashboard, dual-key auditor pulls, and automated statutory dialect translation.
Dual-key portal
< 25 MB
Resident Set Size (RSS) per spoke node process during sustained 10K req/sec load.
< 1.2% Core
Zero-copy byte serialization avoids Python GIL bottlenecks and GC pauses.
100K+ ops/s
Single node throughput capacity without scaling sidecar proxies.