Relay Guard — is the AI model you're using real?
API relays routinely water down their upstream — quietly swapping the Claude / GPT you paid for with a cheaper model (Qwen, DeepSeek, a quantized variant), often prompted to claim it is still Claude. Relay Guard catches that by behavioral fingerprint, end to end.
Upstreams leak
A relay buys real Claude today, then silently routes to a cheaper model tomorrow to widen margin. You never see the switch.
Disguised identity
A swapped upstream is told to answer "I am Claude." Asking the model who it is proves nothing — fingerprinting sees through it.
Silent quality loss
Your product degrades, your users churn, and you blame your own prompts — when the real cause is the model under you changing.
Three ways in
From a free spot-check anyone can run, to continuous monitoring for relay operators, to running your own relay on our stack.
Relay Guard Check
A free behavioral-fingerprint checker. Paste your relay's answers — or run a one-file local CLI — and find out in seconds whether your Claude / GPT is genuine or substituted.
- ✓ No signup, key never leaves your machine
- ✓ Sees through disguised upstreams
- ✓ Genuine / Suspected-substitution / Unknown + evidence
Relay Guard Monitor
Continuous-monitoring SaaS for everyone running on relays. Watch all your upstream channels around the clock and get alerted the moment a model is swapped or degraded.
- ✓ Scheduled probes across every upstream
- ✓ Drift & substitution alerts
- ✓ History & audit trail per channel
Relay Guard System
Want to run your own relay? License our full relay system with the detection module built in — sell upstream you can prove is genuine, and stand out from relays that water theirs down.
- ✓ Production relay stack + billing
- ✓ Built-in fingerprint detection module
- ✓ "Verified-genuine" as a selling point
Why you can trust the verdict
Relay Guard is built on a behavioral-fingerprinting method grounded in published academic work on LLM identification — not vibes.
Fingerprint identification accuracy measured against our reference model set.
Still identifies the real model even when the upstream is prompted to disguise itself as Claude.
Based on peer-reviewed academic fingerprinting methods, not self-reported model identity.
⚠️ Results are probabilistic signals, not legal proof. Quantized / distilled variants and models outside our reference set cannot be determined.