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.

Easy to miss

• Which layer is missing. An unclaimed general.architecture needs a catalog row or ArchitectureExtension; a new mmproj type needs a ProjectorExtension; a structurally different use of an existing op needs an op_case. Only a genuinely new GGML op needs a translator.
• Registration timing. Architecture/projector extensions before load(); converters and passes before convert(), on the same frontend instance. GGUFMakeStateful and AdaptToGenAI are C++ only.
• Test build. New translator or builder sources go into the explicit FRONTEND_SRCS list in tests/CMakeLists.txt, or the test binary fails to link. Run ov_gguf_frontend_tests unfiltered before finishing: the op-coverage gate runs in teardown, so check the exit status. GGUFArchConversion skips without generated headers unless OV_GGUF_REQUIRE_ARCH_FIXTURES=1 requires them; generate them with gen_arch_fixtures.py --fetch. Check actual selected counts and skips; registration, conversion, decoder numerics and GenAI media/API integration establish different levels of support.
• References. Layout-sensitive expectations come from real ggml CPU via a committed oracle under tests/, with several heads/tokens and unequal Q/KV dimensions; NumPy or single-head tests can share the bug. Do not tighten expect_near tolerances from an x86-only run.
• Comparisons. Use full logits or features on identical inputs and token histories, with INFERENCE_PRECISION_HINT=f32, DYNAMIC_QUANTIZATION_GROUP_SIZE=0 and the reference's KV precision. For strict checkpoint comparisons, set OV_GGUF_Q4_K_ZP_F16=1 before starting the process. Run the unfiltered frontend unit suite with this variable unset: integer-zero-point tests require the default mode. Keep scenario environments explicit. Report quantized-CPU and represented-F32 comparisons separately and keep failed results. GGML_OPENVINO_* variables configure only the llama.cpp backend. Inspect tensor-type inventory rather than checkpoint names. Identical quantized weights do not rule out CPU arithmetic differences; exact expansions isolate them but do not waive the original parity gate. Shared exported-model layout does not imply identical media geometry.
• Shared changes. Translator, VIEW/op_case or shared block changes rerun GGUFArchitectureAccuracy, GGUFArchConversion and the mmproj suites, not just the target model.
• Long runs. Redirect output to a file, keep the launched PID and stop only that process. Large Q2 MoE checkpoints expand at CPU compilation; check memory before starting one.
• Ineffective rebuilds. Verify graph attributes, plugin/executor parameters and the linked, loaded CPU variants before repeating a sweep. Use build and runtime identity for overlays or multiple worktrees; version strings and ldd alone do not prove execution identity.
• Validation batches. Extend shared-path regression checks to tokenizer, mask and PA changes with representative SDPA/PA/media/chat cases. The validation guide separates numerical parity, model quality and Optimum compatibility; a known arithmetic difference remains a failed parity check unless the user changes the acceptance contract.