REJECTED LEADERS
microsoft/BitNet
URL: https://github.com/microsoft/BitNet
License: MIT — VALID
Reason rejected: The 1.58-bit kernel dispatch lives in C++/CUDA infrastructure (src/ggml-bitnet.cpp, src/ggml-bitnet-lut-kernels.h). The Python-facing layer is a thin subprocess wrapper around a compiled binary (run_inference.py). There is no clean Python dispatch loop that can be honestly absorbed to ≤10 lines without either (a) hallucinating function signatures or (b) copying C++ semantics verbatim. The doctrine's "No hallucinated function signatures" rule prohibits reconstruction from documentation. REJECTED.
vllm-project/vllm (MoE expert routing)
URL: https://github.com/vllm-project/vllm
License: Apache-2.0 — VALID
Reason rejected: The MoE expert routing dispatch spans at minimum three files: vllm/model_executor/layers/fused_moe/fused_moe.py (topk softmax + expert selection), csrc/moe/moe_align_block_size.cu (CUDA kernel), and vllm/model_executor/layers/fused_moe/layer.py (forward dispatch). The minimal honest absorption requires importing vllm's fused_moe module — which is not a ≤10 line standalone kernel. The doctrine's "if can't be honestly minimized, write WHY in REJECTED.md and stop" rule applies. REJECTED.
Summary
Both non-chosen candidates have valid open-source licenses (MIT and Apache-2.0). Rejection is solely on minimization grounds, not license grounds.