Blog / Evaluation note

Why PageSpatial preserves both readings when native text and OCR disagree

27 Aug 2026·5 min read·Evaluation note

A PDF can contain two different truths: the text encoded inside the file and the ink visible on the rendered page. Native extraction reads the first. OCR reads the second. Either witness can fail while still returning a confident-looking value.

PageSpatial therefore keeps both raw observations. Matching, reading order, tables, Markdown, and conflict records are derived data that reference those sources. A derived view cannot erase the evidence it came from.

The measured reason

In the project’s adjudicated conflict set, OCR was right in 23 cases, native extraction was right in 17, and both were wrong in 6. The split is not a production accuracy estimate; it is evidence that “always trust native,” “always trust OCR,” and “pick the higher confidence” are unsafe policies.

adjudicated conflicts

OCR right       23
native right    17
both wrong       6
total           46

The evaluation also confirmed all 14 blocking conflict escalations that were checked. That result supports a narrow claim: when the current rule set says a disagreement is material, hiding it would be unsafe.

What the service does

The deployed parse path runs native PDF extraction and local CPU OCR. It associates observations that refer to the same ink. When a critical value remains materially different, PageSpatial preserves both readings and raises a blocking conflict rather than silently choosing one.

The system does not claim that disagreement reveals the correct answer. It claims that disagreement must remain visible.

Downstream systems can route that record to a stronger automated tier, exclude it from trusted retrieval, or inspect the cited page region. The evidence record makes those choices possible without pretending certainty.

Scope of the claim

The figures above come from a development and gold evaluation corpus, not a random production sample. Geometry precision, table structure, retrieval quality, and answer grounding still require their own measurements. PageSpatial’s contract is narrower: retain source evidence, make derivation explicit, and fail visibly when a trusted value cannot be justified.