Meaning tier

Checks that compare meaning with sentence embeddings: fastembed's ONNX build of all-MiniLM-L6-v2, on CPU, with no torch. Every finding is a judgment question, because a model's similarity is evidence, not proof.

They run on every push; the model is cached between runs. Install with pip install -r checks/meaning/requirements.txt; without it the tier is skipped with a notice.

File	Role	Guarantee
document_mode.py	Document mode: a long page leans clearly to one of the four modes of documentation	Semantic, so judgment.
embeddings.py	The sentence-embedding model the meaning checks share, loaded once, on CPU
name_content.py	Name and content: what a file's name means is close to what its opening says	Semantic, so judgment.
paraphrase.py	Paraphrase: the same fact said twice in different words, which the word-run duplication check cannot see	Semantic, so judgment.
section_cohesion.py	Section cohesion: no section of a file is far, in meaning, from all the others	Semantic, so judgment.