Summarising a long document in chunks with locators a reader can check

Goal

Produce a summary of a document too long to read in one pass (a standard, a contract, a long report) in which every statement can be traced to a place in the source and checked there in seconds.

Prerequisites

The document as text with stable locators: page numbers for a PDF, section numbers for a specification, line numbers for a transcript. A definition of what the summary is for, since the same document has different summaries for different questions. The methodology on this wiki for summarising a source without distorting it covers fidelity within one chunk; this one covers the assembly across chunks.

Steps

1. Split along the document's structure (chapters, numbered sections, pages), not at a fixed token count, and record the locator range of each chunk. Keep chunks small enough that the relevant passage is never in the middle of a long input; the cited paper reports that models use information in the middle of long contexts worse than information at the ends.
2. For each chunk, extract claims as records: claim, locator, quote (a short verbatim span), strength (states, recommends, requires, reports). Do not summarise yet.
3. Verify each quote by exact string match against the chunk text (after whitespace normalisation). A claim whose quote is not found is dropped or re-extracted, never kept on trust.
4. Merge the claim records: remove duplicates, group by topic, keep the document's own emphasis (a requirement stated once in a normative section outranks a remark in an appendix).
5. Write the summary from the merged records only, and keep the locator on every sentence: "(section 4.2)", "(p. 17)". A sentence that cannot be given a locator is either a synthesis, which is labelled as such, or removed.
6. Where the provider offers document citations, use them for the per-chunk step: the cited vendor documentation describes citation blocks that carry the cited text with start_page_number for PDFs and character offsets for plain text, which gives locators without asking the model to type them.
7. Deliver the summary together with the claim records, so a reader can spot-check any sentence.

Expected result

A summary in which each sentence resolves to a locator and a verified quote, plus a table of claims that doubles as an index into the source.

Limits and test basis

The quote check catches fabricated or misplaced citations, not misreadings that quote correctly and conclude wrongly; step 4 is where a second reader is worth the cost. Scanned documents with OCR errors defeat exact matching and need fuzzy comparison. No accuracy figure is claimed.