Content Faithfulness
Say what the document says — no more, no less, and nothing it doesn't.
Content faithfulness is the property that an AI system's output accurately reflects its source document: extractions state what the document states, summaries preserve its claims without additions or distortions, and answers to questions are supported by what the pages actually say. It is the document-AI-specific framing of the hallucination problem — a system can be fluent, plausible, and confident while asserting things the document never said, and in document work that failure mode is uniquely dangerous because the output looks like it came from the source.
Faithfulness failures have characteristic shapes worth distinguishing. Fabrication: a value or claim with no basis in the document (a language model completing a pattern — inventing a plausible invoice date). Distortion: the document's content altered in meaning — a negation dropped ("no evidence of default" becoming "evidence of default"), a hedge removed, a number's unit changed. Omission-as-assertion: a summary that silently drops the exclusion clause changes what a reader believes the contract says. And contamination: the model blending its general training knowledge into what presents as a statement about this document — describing what lease agreements typically say rather than what this one says.
Faithfulness is enforced and measured, not assumed. Enforcement mechanisms include citation-grounded outputs (every claim pointing to supporting text, automatically checkable), extraction constrained to document spans where the task allows, and prompting or fine-tuning that penalizes unsupported assertions. Measurement uses entailment-style checks — does the cited source actually support the claim? — via automated verifiers, LLM-as-judge protocols with human calibration, and targeted test sets seeded with documents that contradict common-knowledge priors, which catch contamination that ordinary benchmarks miss. For regulated use, faithfulness metrics belong beside accuracy in the evaluation stack: an extraction system is not "99% accurate" in any meaningful sense if its errors include confidently fabricated values.
Every answer comes with a receipt — the page, the region, the exact words it was extracted from.
The answer must live in the pages — grounding keeps AI outputs tethered to what the document actually says.
Not 'how good is the model' but 'how good is it on our documents, per field, against ground truth.'
Proof Perimeter runs document AI inside your own perimeter — with a provenance record on every field.
Book a demo