Arm Kleidi Libraries product page (developer.arm.com)

Vendor product page, used ONLY for the integration list and the CPU-family scope, both of which are vendor-published and carry no performance figure. States Kleidi libraries are 'a lightweight suite of open source routines' optimised for 'Arm Cortex-A, Cortex-X, and Arm Neoverse CPUs' and for ML and CV tasks, with 'low overhead, high efficiency' and minimal resource demands. The integration list is the deployment-relevant claim: 'Integrated with popular ML and AI frameworks like PyTorch, ExecuTorch, MediaPipe, ONNX Runtime, LiteRT, and llama.cpp, enabling rapid acceleration of inference models'. Four of those six are engines/records in this repo, which is why the compiler record's production_evidence names them rather than asserting a speedup. The page also carries an entitlements banner ('Entitlements could not be checked due to an error reaching the service') - noted because it is the reason no gated documentation was reachable for this pass. No throughput, latency or accuracy number appears anywhere on the page, so none is asserted.