# Summaries You Can Trust: Every Fact Linked to Its Source Page

Anti-hallucination by design: generation constrained to retrieved passages, field-level provenance for structured extraction, a verification pass, and "not stated in the file" as a first-class answer.

The objection to AI summaries in serious casework is correct: a fluent paragraph that might be invented is worse than no paragraph at all. A reviewer cannot cite "the AI said so" in a decision. So the engineering question is not how to make summaries sound reliable. It is how to make every sentence checkable, mechanically, in seconds.

The answer is provenance by design: a summary is not a text, it is a set of claims, and every claim carries a pointer to the exact page it came from.

## Constrain generation to the record

The first rule is architectural. The model never writes from memory or from the whole file; it writes from retrieved passages — the specific pages the retrieval stage selected. Its instructions are strict: use only this material, cite each fact, and where the material does not answer, say so. This single constraint removes most hallucination risk, because the model is summarising evidence placed in front of it, not recalling what such files usually say.

## Two kinds of summary, two kinds of rigour

**Narrative summaries** — the record summary, the global factual summary — are generated passage by passage, with citations attached at sentence level. Each sentence resolves to a document and page. The reviewer reads the summary as an index into the file: anything surprising is one click from its source.

**Structured extraction** — claim number, worker name, date of accident, area of injury, benefits granted, decisions made — is handled field by field. Each field carries its own provenance: the value, the page it was read from, and a confidence score. Fields the file does not support are returned empty with a reason, never filled by inference. An empty field is information; a guessed field is a liability.

## The verification pass

Generation is checked, not trusted. A second pass — programmatic where possible, model-assisted where necessary — tests each cited claim against its cited passage: does this page actually support this sentence? Claims that fail are dropped or flagged for human review, never silently kept. Over time, the failure patterns teach you where your OCR, retrieval, or prompts are weak. The summary pipeline becomes self-diagnosing.

## Flags are part of the summary

A trustworthy summary also surfaces what a careful reader would flag: an occupational-disease entitlement question, a security concern, material in a language other than English or French, a gap in the treatment record. Keyword and pattern detection over the indexed text generates these flags with the same provenance discipline — each flag points at the passage that triggered it. The summary's job is not to make the file look tidy. It is to make the file legible, including its problems.

## Why this changes adoption

Reviewers do not reject AI summaries because they dislike technology. They reject them because verification costs more than reading. Provenance inverts that: checking a cited fact takes seconds, so the summary genuinely saves the hours it promises. That is the whole adoption argument, and it is an engineering property, not a marketing one. Build summaries as claims with receipts, and the trust follows.
