Partially reconstructs Table 1: most of the roughly 90 numeric values are exact, but literal 0 values in the 12-inch column become blanks, one mean cell picks up stray digits (33.0 830000), and a row-label-only section header is broadcast across all six columns in one instance.
What was measured
Table Preservation
Preserves complex table structures, including rows, columns, multi-row headers, and merged-cell relationships in markdown.
decisive for this rankingtransformation
Accurate Markdown conversion of complex PDFs depends on keeping table structure intact, not flattening it into plain text. (3 of 3 judges)
What was given, what came back
Test input: Scanned Research Paper · pdf · group: scanned-research-paper
Input — what we sent
Scanned Research Paper
An image-only scanned research paper used to stress OCR and layout recovery in a multi-column academic document with figures, charts, tables, captions, and references.
Why this input is hard
- · OCR on scanned pages
- · Multi-column reading order
- · Figure and chart handling
- · Table reconstruction from scans
- · Caption association
- · Reference extraction
- · Overall document structure retention
Output — unretouched
Also checked on this input — same tool, 7 other criteria
Advanced Features✓ WorkedSchema-driven extract.run fully recovers the corrupted Table 4 rows, returning all 12 checked rows exactly, including the 10-inch-cut 1979 MPB row whose parse.run output was badly garbled.Advanced Features◐ MixedThe confidence signal flags the faint handwritten margin note and the garbled Table 4 as low confidence, but it misses the hallucinated 'USA' word in the title, which is still rated 0.9354 confident.Complex Document Handling⚠ StruggledHandles the 12-page scan in 10.6 seconds with no truncation, but the densest 17-column table has unreadable stretches with injected glyphs and run-on numbers.Markdown Quality✓ WorkedUses no invented tags or malformed markdown; the only HTML seen is legitimate <br /> inside table cells, so the syntax stays clean even though the document does not surface real heading markup.Reading Order & Structure⚠ StruggledDisplaces the byline by a full column: the author line that sits above the two-column split in the source is emitted only after the entire left column, and its footnote markers are rendered inconsistently.Text & OCR Completeness✓ WorkedConverts all 12 scanned pages with no gaps; the page-11-to-page-12 handoff is preserved verbatim, and a separate page-marker output shows pages 1 through 12 present with no missing markers.Visual Content Retention✓ WorkedSegments a shield logo out of a single full-page raster scan and also retains Figure 1 as an image with an accurate synthesized caption; the returned logo crop is a tight 77x81px cut with legible shield text.
Provenance
- Observation
- fcad1a88-bccf-4eea-8bf5-0d8e5dbc22ee
- 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: "scanned-research-paper"
})MCP · mcp.aidemos.com/api/mcp
Free with attribution.
Same input, same check — 8 other tools
measured on Table Preservation
Adobe API✓ WorkedPreserves a multi-column scientific table of tree treatments and before/after diameters, keeping the treatment rows and inch/cm subcolumns readable.Extend AI◐ MixedKeeps the numeric table structure, but drops between-column annotations and leaves stray OCR characters in some numeric cells, so contextual information is partially lost.Landing AI◐ MixedPartially reflects the text that sits between table columns, but the intervening content still disrupts the table structure in the parsed output.LlamaParse✓ WorkedPreserves a nested treatment table with the 7-inch, 10-inch, 12-inch, 100-leave-tree, and clearcut columns and the acres/live-lodgepole rows.Mistral AI✗ FailedBroken column boundaries and disrupted value alignment make the reconstructed table significantly less faithful to the source.Nutrient.io✗ FailedBreaks more complex multi-level tables, with rows and merged cells misaligned or lost once the hierarchy becomes dense.Tensorlake✗ FailedStruggles with hierarchical scanned tables, misplacing column headers and producing unreliable reconstructions on both the multicolumn table and the denser complex table.Upstage AI◐ MixedKeeps the table values intact but reconstructs the headers incorrectly, leaving a grouped-column table with inconsistent structure.
This evidence is published in
Real inputs and real outputs, no retouching · every cell queryable via API & MCP · aidemos.com