When word-wrapped prose looks exactly like a table

TLDR

PDF generators that encode paragraphs as individual word-level text items create false-positive borderless table classifications because scattered word $X$-coordinates trigger spatial column anchor clustering. Ratio-based fill-rate gates (tagged / (bands * anchors)) fail because the line-start anchor inflates density. Adding an explicit anchor quality gate requiring $\ge 50\%$ of column anchors to appear across $\ge 25\%$ of participating bands eliminates prose table misclassifications.

Document Layout	$X$-Anchor Clustering Behavior	Fill-Rate Gate Result	Anchor Quality Gate Result
Word-Span Paragraph Prose	1 Line-start anchor + 8 noise word anchors	$0.304$ (Passes ratio threshold)	Rejected ($1/9$ anchors qualified)
Structured Stream Table	4 Consistent data column anchors	$0.750$ (Passes ratio threshold)	Passed ($4/4$ anchors qualified)

---

Technical defect analysis & anchor quality gate

// Anchor Quality Gate: Ensures column anchors appear consistently across bands
export function validateAnchorQualityDistribution(colAnchors, bands) {
  // For small groups (<= 5 bands), anchors must appear in ALL bands
  const minPresenceBands = bands.length <= 5
    ? bands.length
    : Math.max(3, Math.floor(bands.length * 0.25));
const qualifiedAnchorsCount = colAnchors.filter(anchor => {     const distinctBandIndices = new Set(anchor.items.map(item => item._band)).size;     return distinctBandIndices >= minPresenceBands;   }).length;
// Require at least half of the detected anchors to meet the band presence threshold   return qualifiedAnchorsCount >= Math.ceil(colAnchors.length / 2); }

Rule of thumb: Combine overall fill-rate ratio gates with anchor quality distribution checks to prevent word-span prose from misclassifying as borderless stream tables.