Concrete workflow-risk signals before content ships.
VeracityAPI does not claim forensic proof. It catches AI slop, weak provenance, unsupported claims, and synthetic-media cues that should change how an agent routes content.
What we catch
- Generic AI phrasing patterns.
- Unsupported factual claims.
- Weak provenance / source ambiguity.
- Synthetic-image cues.
- Synthetic-audio cues.
- Low-information filler content.
What we do not catch
- Whether a specific person wrote it.
- Whether a claim is objectively true.
- Copyright / IP infringement.
- Hate speech / harassment.
- PII detection.
- Speaker identity proof.
- Courtroom/academic/legal evidence.
Generic, low-information phrasing
Before: “In today's fast-paced digital landscape...”
recommended_action=revise · primary_reason=generic_phrasing
Fix: replace with specific customer, problem, mechanism, and evidence.
Unsupported claims presented as fact
Before: “Studies show 87% of users prefer...”
recommended_action=human_review · primary_reason=unsupported_claim
Fix: cite a source, lower certainty, or remove the claim.
Weak provenance / source risk
Before: scraped or unattributed source snippet with no corroboration.
recommended_action=human_review · primary_reason=weak_provenance
Fix: require corroboration before publishing, citing, or training.
Synthetic image cues
Flags visible artifacts, inconsistent lighting, distorted text/logos, and suspicious provenance context as synthetic media risk.
recommended_action=human_review
Fix: request original source, metadata, or manual visual review.
Synthetic audio cues
Flags synthetic-speech cues, suspicious transcript/prosody, and low-provenance voice notes for review.
recommended_action=human_review
Fix: verify source, consent, and publication context before use.
Use it like a linter for LLM pipelines
Text results prioritize slop_risk, specificity_risk, and provenance_weakness. Image and audio results use AI forgery / synthetic media language instead of calling media “slop.” Every result should drive an action: allow, revise, human_review, or reject.