MG-RESOURCE-001 · Application availability · Repaired

Why a small CSV created a huge hidden page

File size alone did not describe the amount of page construction this application performed.

A synthetic 226,863-byte cost CSV produced 437,704 page elements in Margin Guard. Most were inside closed source-detail sections. The controlled result was temporary interaction blocking, with no observed crash, code execution or data disclosure. [1]

This explains an existing repaired, low-severity application finding from Browser Busters, Ten Thirty Two’s AI-assisted review project. It is not a new browser-engine flaw or a claim that every CSV freezes a browser.

The input was wide; the hidden output was larger

The selected cost file contained 500 data rows and 410 columns. Rendering created 425,524 elements inside closed details, out of 437,704 total elements. One measured main-thread task lasted 1,099 milliseconds. These are observations from the controlled environment, not a timing promise for another machine. [1]

The important boundary was between hiding detail and postponing its construction. The original application built detailed source evidence even while the visitor had those sections closed. The interface looked compact, but the underlying document was already large. [1]

That distinction makes this useful beyond a file-size headline: a byte limit alone would not have explained the measured page expansion. It does not follow that the same source would create the same document in another CSV tool.

Bound the width and defer unopened details

The repair applies a 100-column parser limit and constructs detailed source evidence when its row is opened. The original 410-column file is rejected. Separately, the team compared a 500-row, 100-column input that remained inside the accepted boundary. [1]

Comparable accepted input: 500 rows, 100 columns [1]
MeasurementBefore repairAfter repair
Initial page elements127,70412,705
Detail constructionBuilt before openingDeferred until requested

After repair, the initial view contained zero evidence bodies. Opening one row added 230 elements and displayed 107 source fields. That check matters: reducing the initial page by removing the evidence altogether would not satisfy the same review task. [1]

The 437,704-element original case and the 127,704-to-12,705 comparison use different column counts. They must not be presented as a before-and-after benchmark of the same input. The column rejection and deferred-rendering result are two separate repair checks.

Measure the work a file causes

For a comparable review tool, use an authorized local setup with fictional data. A useful test plan examines input dimensions and the rendered result together, rather than assuming that a modest download size means a modest interface workload.

  1. Record file bytes, row count and column count alongside the visible workflow.
  2. Measure the initial document, including detail sections that are closed.
  3. Check an accepted boundary input separately from an input the parser should reject.
  4. Open one detail row and verify that the underlying evidence is still available.

For an operator, the tested repair means an over-wide input receives a validation boundary and accepted details remain available on demand. It does not guarantee instantaneous processing or explain every slow import. A slow file can require its own diagnosis.

Evidence and limits

This required a visitor to select a crafted or unusually wide local file and run the app. The counts and task duration describe the recorded environment. They do not establish a universal denial-of-service condition, customer incident frequency or performance across all devices. [1]

[1] Original application evidence. Browser Busters evidence file, finding MG-RESOURCE-001, especially observed and regression. See Margin Guard and the findings hub for the application context and coverage boundaries.