Measured, not guessed
Real inference requests against Nebius-hosted NVIDIA Nemotron. TTFT, P50/P95 latency, throughput and error rate.
Powered by NVIDIA Nemotron on Nebius
Prove an AI infrastructure change is safe before it ships.
Monitoring tells you something broke. Prelight tells you whether it will, by running a controlled experiment on your real workload and returning an evidence-backed SAFE, WARN or BLOCK decision.
Real safe-boundary scan
How far can max_tokens go from 300?
Each value was tested in five paired rounds against the baseline on Nebius Token Factory. Prelight certifies the highest SAFE setting and blocks the one that regresses latency by more than 20%.
| max_tokens | Decision | Mean latency | P95 latency |
|---|---|---|---|
| 300 | SAFE | -10.05% | -8.14% |
| 350 | SAFE | -3.71% | -5.99% |
| 400 | WARN | +6.40% | +10.02% |
| 500 | BLOCK | +23.00% | +22.79% |
From proposal to proof in eight steps
- 01
Analyze
Parse the proposed change and score its risk.
- 02
Research
Pull external evidence on the failure mechanism.
- 03
Plan
Nemotron designs a controlled experiment.
- 04
Baseline
Measure the current config on the real workload.
- 05
Candidate
Run the proposed config, order-balanced.
- 06
Compare
Paired rounds with 95% confidence intervals.
- 07
Decide
Deterministic SAFE / WARN / BLOCK policy.
- 08
Roll out
Get a canary-ready rollout plan.
Built to be trusted
Fail-closed by design
Missing or incomplete evidence never passes. Incomplete runs resolve to BLOCK, not SAFE.
Paired statistics
Baseline and candidate alternate order across rounds, so drift can't masquerade as regression.
Safe boundary explorer
Scan a range of settings to find the highest value that still certifies SAFE.
Stop shipping on hope.
Open the control plane and see the evidence behind every decision.
Open dashboard