Paths are relative to src/frontends/gguf/. Read only the route the task needs. First identify the input path: a native .gguf file or a supplied GgufDecoder (llama.cpp cgraph). They share converters; builders and validation differ. Registration or conversion alone never establishes accuracy.

Task	Read
Is a model supported?	supported_models.md or mmproj.md; quote the checkpoint result and limitations, not just the catalog entry
Load a model, make it stateful/GenAI/PA-ready, tokenizer, IR export	runtime.md
Add or port an architecture	architectures.md and the family contract; extensions.md when shipping it as a plugin
Add or replace an mmproj branch	extensions.md, then the graph contracts in mmproj.md and multimodal contracts
Missing ggml operation or `op_case`	how_to_add_op.md
Validate generated-text or quantized-model quality	WWB validation; compare both backends against the same saved reference while preserving numerical and integration gates
Wrong numbers	debugging_accuracy.md
GenAI/mmproj token, media or PA mismatch	multimodal integration bisection and multimodal contracts
Unsupported weight type	quantization.md; the fix belongs in `src/quant/`, not a builder
Build, run tests, regenerate fixtures	testing.md
Run or resume GGUF validation across OpenVINO, GenAI and llama.cpp	validation.md; use existing GHA workflows and focused local tests, preserving revisions, coverage and failures
Preflight test selection, fixtures or an overlay link	executable checks; verify required cases and link inputs before a broad batch

ov-gguf