ADR 0012 — Measure a PDF's text layer per page, not per document

ga · floor-software-engineer

Decision

Measure a PDF's text layer per page, not per document

Why

Which of the two reading paths a PDF needs is a property of each PAGE, not the document. A per-document flag marks every affected file for the vision path and loses exactly the information that decides the cost.

Measured

6,258 pages across 114 PDFs, of which 57 are image-only -- 0.9%. A per-document flag marks 20 documents, roughly 4,000 pages. Rendering 57 instead is a 110x difference, and the 153-page Opus 4.5 system card needs exactly page 71.

Inputs and outputs

This decision declares none. Every name in this section, when present, resolves against the capability vocabulary or the floor registry -- a typo fails the generator rather than rendering.

Where it is