When a multimodal document model misreads small text, inspect the image it actually receives, not only the clarity of the original PDF. Rendering, upload or preprocessing may shrink a page until strokes and marks are no longer distinguishable. A longer prompt cannot guarantee a unique reconstruction from that same ambiguous input.

The Nemotron 3.5 Super VL announcement discusses documents, charts and OCR. Using that release as a starting point, this article examines resolution, local crops and evidence location, with an eight-check mathematical counterexample for information loss. No OCR performance of the model was measured. 中文版.

What the announcement establishes

The NVIDIA technical post discovered through Google News was published October 15, 2026. It describes Nemotron 3.5 Super VL combining visual and text understanding with reasoning and tools, and discusses charts, tables, OCR and multipage documents. The event date is October 15, rather than this article’s October 18 date.

The official model card lists text, image and video inputs, with documents supplied as page images plus accompanying text. Support for that input path does not establish correct recognition of every scan at any scale. No advertised throughput or benchmark score is converted into accuracy for your documents here.

The resizing, cropping and acceptance methods below are engineering analysis. In particular, the four-cell averaging operation demonstrates a many-to-one mapping. It is not asserted to be Nemotron’s preprocessing algorithm.

Two different patterns collapse to one value

Take a 2×2 grayscale block whose values are zero or one. A has ones on its main diagonal; B has ones on its other diagonal. Their arrangements differ, but averaging all four cells into one number produces 0.5 for both. That reduced result cannot uniquely identify which original block existed.

The example does not show that every resize destroys text, and it supplies no minimum readable font size. Its claim is narrower: once a transformation maps different originals to the same representation, a later model using only that representation cannot correctly recover the original distinction in every case.

Nor does it prove that arbitrary real documents become identical under a particular production resizer; inspect the processed page to determine whether the relevant distinction survived. In technical documents, a distinction might be a decimal point, minus sign, subscript or thin connection line. Context can support a guess, but a contextual guess is different from reading visual evidence. Allow uncertainty for such fields and retain the original region for inspection instead of treating fluent completion as proof of recognition.

Eight checks reproduce the information collision

The code handles known 2×2 synthetic blocks with exact fractions. Five distinct inputs produce four distinct outputs, with A and B colliding. No imaging library, OCR system, vision model or real document is used.

# A many-to-one teaching transform, not image/model benchmarking.
from fractions import Fraction as F

def average_2x2(grid):
    return sum((F(x) for row in grid for x in row), F(0))/4

a = ((1,0),(0,1))
b = ((0,1),(1,0))
cases = [
    (a,F(1,2)), (b,F(1,2)),
    (((0,0),(0,0)),F(0)),
    (((1,1),(1,1)),F(1)),
    (((1,0),(0,0)),F(1,4)),
]
for grid,expected in cases:
    assert average_2x2(grid) == expected
assert a != b
assert len({grid for grid,_ in cases}) == 5
assert len({average_2x2(grid) for grid,_ in cases}) == 4
print("8 downsampling checks passed")

Eight local checks cover five expected averages, distinct A and B, the input count and the output count. They establish non-invertibility for this specific transform. They do not constitute a Nemotron failure report or support an estimate of character error rate from a handful of blocks.

Inspect the final input, not just the source file

An original PDF may contain a text layer, vectors or scanned pages. The model may receive rendered page images, resized images or images combined with extracted text. Record the actual path instead of grouping all of these operations under the label PDF parsing. Different paths expose different evidence to the model.

Useful diagnostic artifacts include the final page image, pixel dimensions, rendering and processing settings, page ID, crop coordinates and versions. If an SDK or server performs additional processing, state which stage is observable. A local preview is not proof of the exact tensor eventually consumed by a vision encoder.

Specify whether crop coordinates refer to the original page, its rendering or a resized image. Retain the mapping through rotation, scaling and padding. Otherwise, a precise-looking bounding box can point to the wrong row. Saving the page ID, coordinate system and processing version together lets a reviewer return to the same evidence.

File size is not equivalent to readable detail either. Compression, format, resolution and character contrast affect different things. A smaller file may reflect more efficient encoding or discarded visual information. Identify the change before deciding that increasing the byte budget is the right repair.

Whole-page location and local reading can cooperate

A whole page preserves layout, headings and table relationships; a local image may preserve finer strokes. One testable workflow locates the relevant area, provides a context-preserving crop and binds the answer back to page and region. That is a workflow option, not a guarantee of superiority over whole-page input.

A crop that is too tight can remove column headers, units, legends or cross-row relationships. Keeping a number while removing its column’s unit can produce a correctly recognized string with an incorrect interpretation. Preserve the context required by the question rather than mechanically taking the smallest bounding box.

Enlarging a crop from an already blurred thumbnail only enlarges the available representation. It does not restore details lost earlier. When fine evidence is needed, return to the high-quality original page to generate the region, then confirm that the new image actually makes the relevant distinction visible.

Measure the input budget under the actual configuration

More resolution, regions or pages can add encoding, prefill or context work. The relationship depends on the model and service’s processing, tiling and token rules. Pixel count is not a universal visual-token formula, and this article does not supply a Nemotron-specific conversion that has not been verified.

Record processed shapes, region count, reported input usage and end-to-end latency. Compare whole-page and crop strategies with the same question, evidence and output requirements. Include the location stage in the cost; timing only the final model call omits the first half of the proposed workflow.

One successful extraction does not define a general budget. Table density, font size, scan noise and rotation can change difficulty. Include both clear and difficult pages, then categorize errors rather than relying on a total average that cannot explain which documents are failing.

Valid JSON is not verified field evidence

Separate three questions for structured extraction: does the result parse, are its field values correct, and are those values tied to the right evidence? Valid JSON can contain a wrong number. A correct-looking number can also come from the wrong page or from completion rather than extraction.

Public or synthetic technical documents can form a validation set with page and region labels for important fields. When the evidence is unreadable, allow missing or uncertain values and route them to human review or a higher-quality input path. Automatically filling every blank with a plausible number makes errors harder to detect.

This connects to the earlier testable AI requirements article. Acceptance criteria should identify inspectable fields and evidence. Asking a prompt to be accurate does not replace the ground truth used to establish whether the requirement was met.

Separate an input failure from a model failure

Fix one page and one specific question. Preserve the original page, the resized version and a context-preserving crop, then compare outcomes. Observe whether errors track the processing change before investigating preprocessing, prompts, model configuration or task capability. Changing model and resolution together obscures which intervention mattered.

If the decisive evidence remains clear but recognition still fails, input quality is not the only explanation. Conversely, once the evidence is no longer distinguishable, a longer prompt is not the only place to look. A useful investigation retains these different conclusions instead of assigning every failure to one favorite cause.

If the information-collision example and document-validation method help locate a recognition error, voluntary support through the section below is welcome.

Sources and validation scope

Checked October 18, 2026 against the Google News item’s October 15 NVIDIA post and official model card. Both diagrams, the averaging model and eight checks are original teaching material. No Nemotron, real OCR, image preprocessor or performance benchmark ran. No particular resolution, token cost or recognition rate is promised.