Drops heading structure on most of the filing: the FORM 10-K opening area receives no # or ## markup, and the report says 80 of 84 pages stay as plain paragraphs rather than navigable section headings.
What was measured
Reading Order & Structure
Maintains headings, sections, column order, document hierarchy, and overall flow.
decisive for this rankingtransformation
Clean Markdown requires the original reading order, headings, sections, and hierarchy to be maintained correctly. (3 of 3 judges)
What was given, what came back
Test input: Hybrid Earnings Report · pdf · group: hybrid-earnings-report
Input — what we sent
Hybrid Earnings Report
A long, real-world hybrid annual report used to benchmark end-to-end PDF-to-markdown conversion. It combines native text, financial tables, charts/graphics, and scanned signature/stamp regions, stressing preservation of document structure, reading order, and embedded visual content across a multi-section report.
Why this input is hard
- · Native digital text extraction
- · Complex financial table preservation
- · Chart and graphic retention
- · Image-based signatures and stamps
- · Reading order across a long multi-section report
- · Markdown quality and consistency
Output — unretouched
Also checked on this input — same tool, 10 other criteria
Advanced Features✗ FailedOn the bar-chart schema, extract.run gets only 3 of 5 Sales values right: 2011 and 2012 are silently pulled from a different table ($69.9B and $73.3B), with no confidence or warning signal.Advanced Features◐ MixedThe parse pipeline’s confidence field catches all 10 footer-number corruptions and also produces one false alarm on a correct securities table, so the signal is real but not perfectly calibrated.Advanced Features✓ WorkedThe separate schema-driven extract.run endpoint cleanly recovers both a financial table and a donut chart: all requested table rows and all five donut percentages come back exactly.Complex Document Handling◐ MixedProcesses the full 84-page mixed-content report in 25.5 seconds with no truncation and keeps the hardest nested table’s 27 values exact, but 10 footer numbers are corrupted in the back two-thirds of the file.Markdown Quality⚠ StruggledProduces usable but not clean markdown: the output contains 12 <signature> tags, 2 <empty> placeholders, and literal HTML tags such as <b>, <i>, and <u> instead of pure CommonMark.Table Preservation✓ WorkedPreserves a genuinely row-paired 52-row state table as a clean grid, keeping the 1,792 stores total and 239,539 retail-square-footage total intact.Table Preservation✗ FailedMis-merges two independent side-by-side lists into one table, forcing unrelated cost-of-sales and SG&A bullets into fake row pairs where no row correspondence exists in the source.Text & OCR Completeness✓ WorkedConverts the full 84-page native-digital annual report with no skipped pages; the output contains all 84 page-marker pairs, and checked high-risk numbers survive exactly, including the $51,550,988,273 market-value figure and 599,982,121 shares outstanding.Visual Content Retention✓ WorkedRetains visual content as real image outputs when requested: the portrait, donut chart, and bar chart are all returned as preserved figure/image crops rather than being dropped.Visual Content Retention✗ FailedDoes not retain handwritten signature graphics as images: a signature crop returns only the <signature> placeholder plus the printed name/title, omitting the cursive mark itself.
Provenance
- Observation
- 77c69713-eb97-47e0-8b45-b4d8c8902bc2
- Evidence run
- 6e3160de-fe46-4b45-b071-72560b5c5d0e
- Study
- Convert a Complex PDF into Clean Markdown with an API
- Research task
- 86b9h7t37
- Tested at
- not recorded
- Source
- first-party
- Evidence state
- verified
- Proof shown
- input + output shown
- Cost / latency
- not captured
- Repeat run
- not captured
- Tester
- not captured
The last three rows are honest blanks, not placeholders — our capture has no field for them yet.
Query this
get_evidence({
tool: "reducto",
scenario: "hybrid-earnings-report"
})MCP · mcp.aidemos.com/api/mcp
Free with attribution.
Same input, same check — 7 other tools
measured on Reading Order & Structure
Extend AI✓ WorkedMaintains report hierarchy and reading flow in an 84-page annual report, keeping section text, bullets, and business narrative aligned instead of flattening them.Landing AI✗ FailedFlattens a top-level page heading into plain text, losing heading semantics on a page that should read as a structured annual-report story page.LlamaParse✓ WorkedKeeps the report's heading-and-paragraph sequence intact instead of flattening the page into disconnected text blocks.Mistral AI◐ MixedHeading hierarchy is reconstructed inconsistently, so some parts of the document are flattened even though the underlying content is still recovered.Nutrient.io✓ WorkedPreserves the section hierarchy and bullet-point reading order on the sample page, keeping the 'A Growth Story Again' heading tied to its paragraph and 4 bullets in sequence.Tensorlake✓ WorkedPreserves heading order and section relationships in an 84-page hybrid annual report, keeping the narrative, bullets, and figure placement aligned with the source flow.Upstage AI✗ FailedBreaks the two-column reading order on the annual-report page, so headings and body text no longer follow the source navigation cleanly.
This evidence is published in
Real inputs and real outputs, no retouching · every cell queryable via API & MCP · aidemos.com