Vet a provider
Enter an upstream endpoint, key, and claimed model. TrueLLMs runs the existing billing, tokenizer-fingerprint, and capability checks, then reports a high / medium / low / uncertain risk headline with decisive evidence and next actions.
Scope and hard limits
This page is a workflow wrapper around the main audit engine, not a separate certification path. It is intentionally adversarial: it can flag substitution, billing inflation, and downgrade evidence, but it cannot prove a supplier is clean.
A full run sends the usage samples, tokenizer-fingerprint probes, and capability floor prompts. Adding a trusted reference endpoint doubles request count and cost. If an API format or provider does not return usage, billing inflation is unavailable.
Current plan: 50 test cases, 50 upstream requests before retries or provider-side redirects. The reference endpoint runs the same set again.
Billing inflation depends on returned usage fields. Some Anthropic-native or OpenAI-compatible gateways omit or normalize usage, in which case this page must report that signal as unavailable.
Base URL without the final /messages or /chat/completions path.
Sent only to /api/proxy for this run; not stored or logged by this app.
Tokenizer estimator selected for this model: tiktoken-o200k.
Enter the supplier endpoint's per-token price as a multiple of official pricing. This converts token inflation into real cost; leave blank to report token inflation only, without a cost conclusion.
Trusted reference endpoint (official key, optional)Optional
If filled, the exact same cases run against your trusted official endpoint. The reference key is kept in memory only and is not printed, persisted, or logged.
Use your own official key. It is used only for this run.
Reference tokenizer estimator: tiktoken-o200k.
No supplier review has run yet.
Keys are kept in browser memory and sent only to /api/proxy for this run. Results are evidence summaries, not guarantees, certifications, or legal advice.