# Same form, three inputs: PDF, screenshot, and URL

Use one controlled intake form to compare recreation inputs, inspect the field checklist, and understand which details need human review.

Controlled benchmark using a synthetic form. Results apply only to the attached fixture and recorded run conditions.

## A controlled comparison

The reference is an eight-field project inquiry: name, email, company, project type, budget range, target date, project description, and permission to reply. All three files show the same empty form. There are no real submissions or personal data.

The source files below make the test reproducible. Score labels, field kinds, required flags, option order, and overall field order separately. Count only input fields, not the heading or submit button.

## What each input preserves



Input	Information available	Review carefully
URL / HTML	Semantic labels, types, required attributes, options	JavaScript-only controls and hidden branch logic.
Screenshot	Visible labels, marks, grouping, and layout	Small text, cropped controls, hidden options.
PDF	Visible page content and sometimes selectable text	Page order, scan quality, and print-only conventions.

## How to run the test

Download the reference files. Recreate the HTML URL, upload the screenshot, and upload the PDF in separate drafts. Compare each result with the same eight-field answer key before making corrections. Record the date and any failed run.

Time from submission to editable draft; then time the cleanup separately. Do not call a faster first draft a faster completed form when it needs more correction. Inspect a published preview and submit synthetic test answers before treating the result as usable.

## What this benchmark can establish

A single simple form is a reproducibility example, not a representative accuracy study. It cannot establish performance on handwriting, complex tables, scanned legal documents, signatures, conditional branches, or competitors’ products.

The local HTML parser run on September 12, 2026 recovered 8 of 8 fields, with 8 of 8 labels, types, and required flags matching the answer key. Processing took about 8.6 milliseconds on the local test machine; this excludes network fetching, model refinement, and cleanup. Screenshot and PDF recreation were not run because the existing AI connection is not configured. The attached run report records those unavailable results explicitly. A missing or failed run remains missing or failed; no score is inferred from another input type.

## Checklist

[ ] Use identical source content for all inputs.
[ ] Count fields and attributes separately.
[ ] Record failures and cleanup, not just speed.
[ ] Test the published result with synthetic answers.

Notes: