Qualcomm AI Engine Direct (QNN) SDK

THIS RECORD IS MOSTLY DELIBERATE GAPS, and that is the honest state of the evidence rather than an unfinished record. QNN's compiler mechanics could not be verified from a primary source: the vendor's technical documentation requires an authenticated Qualcomm ID (the docs.qualcomm.com bundle attempted returned a sign-in wall and a page timeout), the public product page carries no compilation detail, and the GitHub API has no official QNN repository - the closest hit, qdsp6sw/qualcomm-ai-engine-direct-sdk, is a 6-star placeholder created and last pushed on the same day in 2024 with no description. CONSEQUENTLY, production_ready is left ABSENT rather than null: the schema types production_ready as a strict boolean, so a null would fail validation and a value would be a guess. Absent is the correct encoding for 'the repo publishes nothing on this point', which is preferred over guessing. production_evidence is an empty list for the same reason - the two integration facts I do have (ONNX Runtime GenAI lists QNN as a backend; ExecuTorch delegates to it on Android) are real and are recorded as strengths, but they are evidence that the SDK ships and integrates, not evidence that the COMPILER is production-ready, and upgrading one to the other is exactly the vendor-claim-as-fact error AGENTS.md forbids. target_backends and output_artifacts are EMPTY because the enum has no term for a Hexagon NPU and none for the artifacts QNN emits (DLC context binaries and Hexagon device binaries); input_languages carries onnx because that is genuinely how these frameworks feed it and it is the one field where the enum has an honest value. first_release_year is null because no reachable source states it. compilation_strategy is aot on the general shape of a device SDK that converts a model offline rather than JIT-compiling at serve time, and it is the weakest claim in this record for that reason - it is the shape every closed inference SDK has, not something Qualcomm documents. To complete this record, get ungated access to the AI Engine Direct SDK documentation and replace w7k-qualcomm-ai-engine-direct-sdk with a real source; that source's own note already recommends this. Confidence is deliberately low (0.5): the integration facts are solid, the compiler facts are absent.