# Quality verification
Open-model providers can silently quantize, misconfigure, or swap what runs behind a model name.
The endpoint says `deepseek-v4-flash`, but the node serving it may not be running that model at
full precision. Keln checks every node against the released weights, so output quality does not
vary by provider.
## Every provider is verified, continuously
Keln verifies that every provider delivers the model and quality level it claims, using
synthetic probes. Verification runs continuously:
- Each provider's outputs are compared against trusted references, so serving a lower-precision
variant than claimed is detected.
- Providers are scored on test tasks against the other providers serving the same model, which
catches degraded quality and silent context truncation.
- A provider that fails verification is removed from routing automatically and returns once it
passes again.
## Capability checks before listing
Before any provider serves a model, its capabilities are checked end to end: streaming and
non-streaming, tool calling, structured output, image input, the reasoning switch. The same
checks re-run for as long as the provider serves.
Speed targets and what you see when a request fails: [Reliability & SLOs ↗](reliability.md)
Quality verification
Open-model providers can silently quantize, misconfigure, or swap what runs behind a model name.
The endpoint says deepseek-v4-flash, but the node serving it may not be running that model at
full precision. Keln checks every node against the released weights, so output quality does not
vary by provider.
Every provider is verified, continuously
Keln verifies that every provider delivers the model and quality level it claims, using
synthetic probes. Verification runs continuously:
Each provider's outputs are compared against trusted references, so serving a lower-precision
variant than claimed is detected.
Providers are scored on test tasks against the other providers serving the same model, which
catches degraded quality and silent context truncation.
A provider that fails verification is removed from routing automatically and returns once it
passes again.
Capability checks before listing
Before any provider serves a model, its capabilities are checked end to end: streaming and
non-streaming, tool calling, structured output, image input, the reasoning switch. The same
checks re-run for as long as the provider serves.