Limitations

What argot deliberately does not prove: fit suitability, in-vocabulary choices, masked content, and changes outside the checked range.

argot is a probabilistic review guardrail, not a correctness oracle. A finding is evidence to inspect; a clean run means only that no configured detector found a pattern in the checked changeset.

The model needs a suitable fit

argot inspect reports whether the repository has enough eligible history and whether the fitted corpus is suitable. A Not recommended verdict means the learned model is not calibrated enough to treat its hits as strong evidence. Generated, vendored, data-heavy, unsupported, or shallow-history repositories can reduce what the model learns. Use argot init --suggest, review exclusions, or use conventional review until the repository has a sound fit.

Familiar vocabulary can still be wrong

Argot is strongest when a change introduces a pattern foreign to the repository. It does not reliably catch an error built entirely from vocabulary already used there — for example, choosing the wrong familiar exception type or calling an existing API in the wrong place. The semantic and architecture checks narrow different gaps; neither turns a clean check into proof of correctness.

Some input is intentionally masked or outside scope

The voice scorer masks prose such as comments and docstrings, and its normal source-language scope excludes configured generated, data, and non-source content. Custom rules can deliberately cover additional paths. Argot also judges the diff or range you give it: unchanged code, an omitted ref range, and events outside that range are not retrospectively validated.

Optional capability limits

Semantic checks need the semantic feature in the binary (the released binaries ship it) and a semantic index from a fit. With semantic = "off" in [rules], or a fit that did not build one, redundant and misplaced are skipped with a diagnostic while the remaining checks continue — never for want of a network, since the embedder is compiled in. Architecture and integrity checks likewise depend on their shipped feature and fit artifacts. Confirm the live rule registry and fit health with argot rules and argot inspect before relying on a specific detector.

misplaced also abstains where it cannot answer honestly: on a body that calls nothing (a property setter is written out of the names it assigns, so it reads as whatever unit owns them, not as code in the wrong place), and on any function recovered from inside a region the parser could not read. The second is not rare on every grammar — on one real 924k-line Object Pascal tree a single unparsed construct put 23% of the repository’s functions out of reach for both misplaced and redundant. Neither rule reports a guess in that state, so “nothing found” there means “not judged”, not “judged clean”.