Your OCR job finished. Did its page windows preserve the document?

What one completed request establishes

Baidu released Unlimited OCR 2 on October 27, according to PC Watch and Zeniteq. Its long-document processing divides pages into groups and carries cached state across their boundaries. The October 29 version-one paper describes overlapping chunks that share a boundary page. One submitted request can therefore involve multiple internal processing windows.^[1][2][3]

The paper explicitly says its sparse attention does not preserve global document information, limiting understanding of structures such as heading hierarchies. That qualification belongs beside the continuity claim. Completing a long file and recovering a distant section parent are separate acceptance questions. The reported research results are Baidu's; the two publications did not independently reproduce them.^[2][3]

Diagnose this fictional manual

This is a static teaching exercise, with invented text and returned output. No OCR model was run. Imagine a twelve-page manual processed with a window transition around pages ten and eleven. All twelve output page markers are present, in order. Before reading the diagnosis, decide whether that is enough to index its text and section structure.^[2][3]

The source ends page ten with "The valve opens only when" and begins page eleven with "both sensors agree." The fictional returned text reads "The valve opens only when only when both sensors agree." A page-eleven subsection, "Reset procedure", belongs to the section "Emergency controls", introduced on page three. The output assigns it to "Maintenance", a different section introduced on page eight.^[3]

There are two failures: duplicated continuation words and a wrong heading ancestor. Correct page-marker coverage detects neither. Compare the sentence anchors to establish the duplication, then trace the source hierarchy to establish the correct parent. These invented faults illustrate acceptance checks; they are not reported model failures, and they do not show that overlapping windows cause duplication.^[2][3]

Keep a source-anchored seam record

A filled teaching row is: manual A; pages ten to eleven; end anchor "only when"; next anchor "both sensors agree"; observed join "only when only when"; verdict duplicate. A separate ancestry row is: subsection "Reset procedure", page eleven; expected parent "Emergency controls", page three; returned parent "Maintenance", page eight; verdict wrong ancestor.^[3]

For your own record, add the document identity, parser checkpoint, image mode, window schedule, source span and raw-output location. Keep omission, duplication, paragraph joining, heading ancestry, page coverage and truncation as separate fields. A useful row preserves the evidence another engineer needs to check its verdict, rather than just a single overall pass flag.^[2][3]

Compare window schedules without changing the test

The following evaluation is proposed, not executed. Keep the checkpoint, renderer, resolution, image mode, output budget, runtime and page order fixed. Change only the supported window schedule. Zeniteq notes that multi-page global-view processing and single-page dynamic tiling use different visual inputs. Comparing those paths cannot isolate a window-context effect.^[2]

Annotate paragraph continuations and distant heading parents before comparing outputs. Sample window seams and away-from-seam controls, keeping their results separate. Include a whole-document schedule only if it fits; otherwise record it as infeasible. A concentration of seam errors is a lead for investigation. A small sample without that pattern cannot rule out boundary failures elsewhere.^[2][3]

Decide before indexing

The static exercise requires no software or GPU. Actual model execution needs a separate setup decision: the documented path uses custom model code, and serving uses a customized SGLang wheel. Review and pin those components before use. The documented A800 with 80 GB is a test environment, not an established minimum hardware requirement.^[2]

Record output caps and inspect truncation as well as joins. Keep per-step speed separate from total processing time: more pages still require image processing and generated output. Before indexing, retain a seam map and verified heading-parent records. Hold the affected document when a required structural check fails, and state which uninspected regions remain unknown.^[1][2][3]