attune-verify

Generation fact-checker for the attune-* family. Verifies named entities in LLM-generated content actually exist — imports import, CLI flags are real, links resolve, counts match source — so hallucinations that pass unit tests are caught before they reach a reader.

Install

pip install attune-verify

With the optional LLM semantic layer (requires attune-rag):

pip install 'attune-verify[rag]'

Quick start

from attune_verify import verify, VerifyContext
from pathlib import Path

ctx = VerifyContext(
    project_root=Path("."),
    allowed_help_cmds=frozenset(["attune"]),
)
result = verify(generated_content, ctx)
if not result.ok:
    for f in result.findings:
        print(f"{f.kind}: {f.detail}")

Part of the attune family

• attune-rag grounds generation in accurate retrieved sources (input-side)
• attune-verify checks that named entities in the output actually exist (output-side)

Together they bracket generation: rag verifies "is this claim supported?"; verify checks "does this named thing exist?"

What each checker verifies

Findings are error when a claim is refuted and warning when it cannot be checked — an unverifiable claim is never a silent pass. result.ok is False only on errors; raise_if_failed(result) turns it into a hard gate.

Status

Beta — the deterministic core (imports, flags, links, counts) is stable and guarded by a labeled precision/recall corpus (gated ≥ 0.95 each) and mutation testing (gated ≥ 0.75). The public API above (verify, VerifyContext, VerifyResult, Finding, FindingKind, raise_if_failed, and the Judge protocol) is what beta covers: it will not change shape without a deprecation in a minor release. The LLM semantic layer is optional via the [rag] extra and is the least settled part of the surface.

Links are read from prose only. Inline ([text](target)) and all three reference forms — full [text][label], collapsed [text][], and shortcut [text] — resolve against the document's [label]: target definitions. An explicit reference whose label is never defined is an error: it renders as literal text, so the link does not exist. An undefined shortcut is ordinary prose (the [3] case is not a broken link) and is skipped, as are GFM footnotes, which share the same syntax.

Both --long and -short flags are checked. A single-letter short flag is unambiguous, so an absent one is an error. A longer single-dash token is not: -xzf may be a cluster of three flags, -name a single-dash long option, and -j4 a flag with an attached value. Each reading is tried, and if none verifies the finding is a warning rather than an error — an ambiguous token is unverifiable, not refuted.

Known limitations

• A dash followed by digits (-5) is read as a negative number, not a flag, so a numeric short flag (head -5) is not checked.
• A flag written as bare `--flag` in prose is attributed to the nearest preceding word, so it may degrade to a warning rather than resolve.
• Counts are matched to a source by keyword overlap with the claim's surrounding text; a numeric claim whose context names no source is skipped rather than guessed at.

License

Apache 2.0