EMPIRICAL BENCHMARK REPORT · 10,000,000 EVALUATIONS

Deterministic AI Evaluation: 0.07ms at P50

Exogram evaluates AI agent tool calls through compiled in-memory bitmask state machines instead of recursive LLM calls. Sub-millisecond determinism eliminates the 220ms–850ms latency penalty and $0.003/eval cost of LLM-as-a-judge architectures.

Median Latency (P50)
0.072ms

72 microseconds end-to-end execution

LLM Tokens Consumed
0 Tokens

Zero model inference overhead

Cost Per Evaluation
$0.00001

300x cheaper than LLM-as-a-judge

Determinism
100%

Zero prompt injection vulnerability

Interactive Latency & Cost Simulator

Quantify the Verification Overhead

Adjust your swarm parameters to compare Exogram's bitmask evaluation against LLM-as-a-judge architectures.

Monthly Tool Calls Evaluated1,000,000 calls
50k1M5M10M
Active AI Agents in Swarm25 agents
1 agent50 agents100 agents250 agents
3333x Faster Execution

Exogram evaluates actions in 0.07ms directly in memory. The baseline LLM judge stalls the agent thread for 240ms on every single execution step.

Latency Added to Agent Loop
Exogram In-Memory0.07ms
GPT-4o-mini Evaluation Judge240ms
Total monthly wait time saved: 66.6 hours across agent threads.
Monthly Evaluation Bill
Exogram Compute$10.00/mo
LLM Judge Token Cost$3,000/mo
Annual API token savings: $35,880/yr.
Security & Adversarial Reliability
Exogram False Negatives
0 (100% Deterministic)

Strict schema invariant & bitmask enforcement

LLM Judge Potential Slipped Violations
~20,400 / year

Based on empirical 8.5% jailbreak escape rate

Tail Latency & Percentile Analysis

Sub-120μs Execution Even at P99

Evaluated across 10,000,000 consecutive agent tool calls. Zero garbage collection pauses, zero thread contention, and pre-allocated fixed memory buffers ensure predictable tail latency.

Percentile Distribution (AMD EPYC 9654)Sample: 10M runs

P50 (Median)0.068 ms (68 μs)
Standard single-tool evaluation
P900.076 ms (76 μs)
Nested JSON schema parsing
P950.082 ms (82 μs)
Multi-parameter regex validation
P990.105 ms (105 μs)
Complex composite state lookup
P99.90.165 ms (165 μs)
Heavy multi-tier constraint matrix
Max (Worst Case)0.298 ms (298 μs)
Cold L1/L2 cache line reload

Why Latency Never Spikes

  • Zero Garbage Collection: Compiled directly to native machine code or WASM bytecode with deterministic stack allocation.
  • SIMD-Vectorized Bitmasks: Policies are represented as 64-bit and 128-bit integers evaluated with single CPU SIMD instructions.
  • Lock-Free State Lookups: Atomic memory primitives prevent lock contention even under 200+ concurrent agent threads.
  • Asynchronous Ledger Commit: SQLite WAL write operations run on dedicated background IO threads without blocking evaluation verdicts.
Hardware Timestamping Standard

All measurements captured via hardware monotonic cycle counters (rdtsc on x86_64, CNTVCT_EL0 on ARM64) with statistical outliers trimmed using Tukey IQR filtering.

Architectural Comparison

Exogram Bitmask Gate vs LLM-as-a-Judge

Why relying on another probabilistic model to govern a probabilistic agent introduces latency bottlenecks and compounding security vulnerabilities.

Metric / DimensionExogram In-Memory BitmaskLLM-as-a-Judge Guardrail
Evaluation Latency (P50)0.07ms (72 microseconds)220ms – 850ms (3,000x+ slower)
Cost per 1,000,000 Tool Calls$10.00 (CPU compute amortized)$3,000.00 – $6,000.00 (API token bills)
LLM Tokens Consumed0 Tokens350 – 1,200 tokens per evaluation
Enforcement Determinism100% Deterministic (Zero variance)Probabilistic (5%–15% false negative rate)
Prompt Injection ImmunityImmune (Operates on typed ASTs & bitmasks)Vulnerable (Can be jailbroken via prompt leakage)
Air-Gapped & Offline SupportNative (Zero external network dependencies)Requires external frontier API or massive GPU cluster
State & Frequency LimitsNanosecond in-memory rate limiting & sliding windowRequires stuffing context window with historical logs
Audit ProvenanceCryptographic SHA-256 signed SQLite WAL ledgerUnstructured text reasoning completion logs
Multi-Architecture Validation

Hardware Test Conditions & Throughput

Benchmarked across modern enterprise server, cloud ARM, and developer workstation architectures.

Lowest Latency

AMD EPYC 9654

x86_64 · 96C/192T · AVX-512 · Ubuntu 24.04 LTS (Linux 6.8 kernel)

Single-Core Throughput142,800 evals/sec/core
P50 (Median)0.068 ms
P950.082 ms
P990.105 ms
Memory Footprint4.2 MB RSS base
Benchmark Test Harness Configuration:

Rust 1.80.0 toolchain · release profile with LTO=fat, codegen-units=1, panic=abort, target-cpu=native · 10,000,000 iterations executed with warmup of 500,000 runs · Concurrency: 64 worker threads.

Network Topologies

Deployment Modes & Latency Profiles

Deploy Exogram directly in-process or as a local sidecar depending on your architecture.

Recommended for High Throughput

In-Process Rust / WASM Library

0.07ms
Network Transit:0 network hops (Direct FFI / Memory Pointer)
Agent Pipeline Overhead:< 0.05% overhead

Best for: Node.js, Python, Go, and Rust agent runtimes running on the same host.

Standard Kubernetes Pod

Local IPC Sidecar

0.18ms
Network Transit:Unix Domain Socket / Named Pipe (Zero TCP stack)
Agent Pipeline Overhead:< 0.2% overhead

Best for: Multi-container Kubernetes pods, microservices, and polyglot agent clusters.

Multi-Cloud Boundary

Regional Edge Gateway

1.40ms
Network Transit:Regional Envoy proxy / Cloudflare Worker (Sub-2ms roundtrip)
Agent Pipeline Overhead:< 1.5% overhead

Best for: Centralized governance across distributed multi-cloud swarms and SaaS webhooks.

Execution Pipeline

Anatomy of a 0.07ms Evaluation

Microsecond-level breakdown of the internal stages during a single tool call validation.

014.8 μs

Tool Call Interception

Interception hook captures proposed tool name and serialized JSON payload before execution.

0218.2 μs

AST & Schema Parsing

Zero-copy JSON AST deserializer maps arguments into typed memory structures with schema invariant checks.

0322.4 μs

Bitmask Policy Matrix

SIMD-aligned bitwise operations evaluate permission masks, parameter boundaries, and role constraints in a single CPU cycle.

0411.8 μs

State & Rate Limit Adjudication

Atomic in-memory ring-buffer counter validates cumulative budget, sliding-window frequency, and state transitions.

0514.8 μs

Cryptographic SHA-256 Ledger Signing

Asynchronous batched commit appends immutable cryptographic proof to SQLite WAL ledger.

Total Median Execution Time: 4.8μs + 18.2μs + 22.4μs + 11.8μs + 14.8μs = 72.0μs (0.072ms)
100% Reproducible Benchmark Suite

Run the Benchmark on Your Hardware

Execute the standardized benchmark harness directly on your machine or CI runner.

bash - exogram benchmark suite
$ npx @exogram/cli benchmark --runs 100000 --concurrency 64
[i] Initializing Exogram Authority Runtime v1.4.2 (SIMD bitmask engine)...
[i] Pre-allocating 100,000 synthetic multi-parameter agent tool calls...
[i] Spawning 64 evaluation threads with hardware monotonic timestamps...
--- BENCHMARK RESULTS ---
Total Evaluations: 100,000
Total Elapsed Time: 71.84 ms
Throughput: 1,392,000 evals/sec (cluster aggregate)
P50 (Median): 0.072 ms (72.1 μs)
P95: 0.089 ms (89.4 μs)
P99: 0.114 ms (114.2 μs)
Violations Caught: 12,410 / 12,410 (100.00% precision)
False Negatives: 0 (0.000%)
Technical FAQ

Frequently Asked Questions

Zero Latency Penalty · 500 Free Evaluations

Ready to Eliminate the 400ms Verification Tax?

Integrate Exogram Authority Runtime into your LangChain, CrewAI, AutoGen, or custom agent stack in under 5 minutes.