Does the model match the label?

Check the evidence an OpenAI-compatible endpoint exposes. No single prompt can prove which weights are behind a proxy, so this tool keeps claims, signals, and uncertainty separate.

01 / metadata 02 / probe 03 / compare

Target endpoint

Send a small audit

A01

API base URL Use the API root, not the full /chat/completions path. The provider must allow browser CORS. API key (optional for local APIs) Use local CORS proxy Local-only. Run python3 proxy_server.py --port 3000; public hosting keeps this off to avoid an open proxy. Run safe behavior fingerprint Runs a small multi-probe battery. It improves evidence quality but still cannot prove the underlying weights. Share verification metadata Opt-in only. Sends provider URL without query data, model ID, status, and aggregate signals. Never sends API keys or raw responses. Model ID to audit Use the exact ID you were given by the provider or router.

Your key stays in this tab and is sent only to the target URL. Collection never receives the key or raw response.

Audit output

Awaiting signal

B01

Enter a target and run the smallest useful check.

Consistency is evidence. It is not provenance proof.

Method

How we test it

C01

We compare independent signals instead of trusting a single self-report.

01
Check transport Read /models and a completion envelope.
02
Compare labels Match claimed, registry, and runtime labels, including safe aliases.
03
Probe behavior Run eight small checks for format, safety, cutoff, and consistency.
04
Make a verdict Score valid evidence; errors and timeouts get no score.

Why this helps It can expose label swaps and wrapper inconsistencies better than asking "what model are you?" alone. It cannot prove the underlying weights; an official endpoint is still the strongest reference.

Community signal

Provider / model observations

These are opt-in, community-reported consistency observations—not a guarantee that a provider is official or that the model weights are authentic. Current A/B/C tiers use valid v2 audits only; older records remain visible as a historical baseline but do not affect the ranking. Open Details to see why.

Loading community observations...