What happens in those 28-43ms
LCAC Context EvaluationMinimum-context enforcement · what is this agent authorized to access?
Policy EnforcementEvaluate all active governance policies for this tenant and agent.
RIS ScoringCompute chain stability, semantic coherence, drift, variance dimensions.
CII CalculationUnified Cognitive Integrity Index = (RIS + LCAC trust) / 2.
Audit Ledger WriteImmutable LCAC governance ledger entry with hash chain verification.
Usage TrackingPer-tenant, per-model, per-call usage increment for billing and limits.
Benchmark Results
Measured in production on live ATOM Platform infrastructure · Groq llama-3.3-70b backend · Jan 2026
10ms
ATOM overhead (min)
Single request
Cached policy + warm governor
28-43ms
ATOM overhead (typical)
Production average
Full governance stack
326ms
Total latency (p50)
10ms ATOM + 316ms Groq
Single request baseline
10/10
10 concurrent
100% success rate
All governed and audited
25/25
25 concurrent
100% success rate
After concurrency fix
619ms
Console API (20 req)
20 concurrent requests
Enterprise Console API
| Test |
Concurrent |
Success |
Avg Latency |
P95 Latency |
Result |
| Baseline single request |
1 |
1/1 |
326ms |
326ms |
PASS |
| Light concurrency |
10 |
10/10 |
341ms |
398ms |
PASS |
| Medium concurrency |
25 |
25/25 |
378ms |
512ms |
PASS |
| Console API burst |
20 |
20/20 |
619ms |
741ms |
PASS |
| Governance-only overhead |
1 |
1/1 |
28ms |
43ms |
PASS |
Overhead Breakdown
Where the 28-43ms governance overhead is spent (approximate, measured).
Comparison: Governance Approaches
28-43ms is the cost of knowing every AI decision was authorized before it executed.
No governance
Raw LLM calls, no oversight
0ms
Zero protection. No audit trail. No blocking. No enforcement. Damage discovered after execution or not at all.
Insufficient
Post-execution monitoring
Log and alert after the fact
0ms
0ms overhead, but detects damage after it's done. Cannot block. Cannot roll back. Compliance audit after an incident ≠ governance.
Too Late
ATOM pre-execution
Govern before every call executes
28-43ms
Full policy enforcement, RIS scoring, LCAC context evaluation, audit ledger write · all before the model call executes. Blocks on policy violation. Zero post-hoc damage.
Governed
"28-43ms is the cost of knowing every AI decision
in your system was authorized before it executed."
ATOM Labs Production Benchmark · January 2026
Methodology
All benchmarks measured on the live ATOM Platform production infrastructure. Backend: Groq LPU (llama-3.3-70b-versatile). Governance overhead isolated by measuring total_latency - model_latency. Load tests run with Python asyncio concurrent task batches. No synthetic warmup · results reflect cold-path latency including Redis reads and Postgres writes.
Governance overhead varies based on: policy complexity (number of active rules), Redis cache hit rate, Postgres write latency, persona blend vector complexity, and RIS scoring depth. The 28-43ms range covers >90% of production calls. Outliers above 60ms occur during first-call-after-restart and during governance rollup operations.
These results reflect the ATOM Platform as deployed on a DigitalOcean Droplet (4 vCPU, 8GB RAM, Ubuntu 22.04). Enterprise deployments on dedicated infrastructure will show lower overhead.