Benchmarks

Real numbers from real code. The test scripts are in the repo. Run them yourself.

Destruction Test — 1,000 Agents

1,000 agents (85% normal, 10% bad actors, 5% manipulators) running full governance lifecycles through 100 stress rounds. In-process, single node.

6.6ms

Enforcement p99

Error rate

100%

Bad actor denial

102.7MB

Peak heap

Metric	Value	Target	Status
Registration throughput	104,068 agents/sec	> 100	PASS
Enforcement p50	0.1ms	< 5ms	PASS
Enforcement p95	1.1ms	< 50ms	PASS
Enforcement p99	6.6ms	< 100ms	PASS
Proof emission	19,884/sec	> 50/sec	PASS
Bad actor denial rate	100.0%	> 50%	PASS
Peak heap usage	102.7MB	< 512MB	PASS
Error rate	0.00%	< 1%	PASS

10/10 thresholds passed. Nothing broke. 323,200 actions processed.

Full results + reproduction commands →

k6 HTTP Benchmark — Cognigate API

50 VUs (Normal Load)

Global p9519.79ms

Throughput286 req/sec

Error rate0%

500 VUs (Stress)

Global p953.49s

Throughput231 req/sec

Error rate0%

500 VU test saturates a single Node.js process. System stayed correct (0% errors) but slow. Production architecture uses 3+ replicas behind Traefik with KEDA autoscaling.

Enterprise Soak v3 — 500 Agents, 500K Signals

8,740

signals/sec

5,873

proof writes/sec

28,061

concurrent writes/sec

No leaks

memory stable

6/6 tests passed. 21-minute sustained run. 18 MB growth over 30 cycles (stable).

Scaling Path to 50k+ RPS

Configuration	Expected RPS	Basis
Single node (local)	230	Measured
Single node (4 vCPU VM)	1,000-2,000	CPU headroom
3 replicas + Traefik	3,000-6,000	Linear scaling
K8s + KEDA (10 pods)	10,000-20,000	Horizontal
K8s + KEDA (50 pods) + Redis	50,000+	Target

Reproduce These Results

# Destruction test (1,000 agents)
npx tsx scripts/destruction-test.ts

# k6 HTTP benchmark (requires cognigate-api running)
k6 run k6/steady-state.js

# Enterprise soak v3
npx vitest run tests/battle/enterprise-soak-v3.test.ts

All test scripts are in the open-source repository. Apache 2.0 licensed.