Citations name one page, even when the chunk spans several

See overview and citation provenance.

Context

A citation stores one printed page, taken from where a retrieved chunk begins, but a chunk can span several pages. A spot-check found Volume 3 cited page 64 when the answer's text was on page 66, from a three-page chunk. Measured across the corpus, roughly half of all cited chunks span more than one printed page, so the error this exposes is bounded by the span, typically zero to three pages, not the eleven-to-forty-two page drift v1 shipped with.

A follow-up spot-check opened all eight cited pages across eight books and found seven exact; the only miss was the Volume 3 case. Every page_offset checked out; the drift comes only from chunks that span pages, not from a second, independent bug.

Decision

Do not cite a page range. Close the issue as accepted, not fixed: the measurement stands, the cost of the error is judged acceptable.

Alternatives rejected

• Cite the range (p. 64-66), truthful and needing no re-ingest. Rejected as unnecessary precision for a reader who never relied on exact pages in the first place.
• Locate the exact sentence the model used and store per-page text so the citation can point at the page actually carrying it. Precise, but touches generation and re-runs the whole answer set.
• Chunk on page boundaries to remove the ambiguity at the source. Left open for the eval sweep to weigh as a retrieval-quality question, not a citation-accuracy one.

Why

Editions differ: the same paragraph sits on a different printed page across different printings of the same book, and the nine books in this corpus are not the nine printings on any given student's shelf. That variance is already larger than the zero-to-three page error a spanning chunk introduces, and a reader already treats a cited page as an estimate. A range would formalise a precision the product never had, at the cost of a migration, a display change, and a column on every citation, to buy honesty nobody was missing. Feedback on citations as shipped has been good.

Consequences

The underlying measurement is not withdrawn: about half of citations still point at the start of a chunk that may extend a page or more past it. Chunking on page boundaries stays available to the retrieval sweep on its own merits.

Status

Closed 2026-09-10, product decision, no code changed. Reopens only if a reader reports opening the named page and not finding the claim.