Detection

Status: Current implementation contract for table and dimension evidence

Authority: Current contract (2026-08-12)

Last reviewed: 2026-08-14

Question

What does Detection receive, do, produce, and refuse to claim?

Short answer

Detection finds independently observable evidence. It does not assign finished engineering meaning. A detector or model score is not semantic truth.

Receives

• Extraction inputs with page coordinates
• Cached OCR, arrowhead, geometry, and line results when available
• Human review state for accepted, adjusted, rejected, and missing regions

Does

• Proposes table regions through named providers
• Runs a shared table verifier on finite rules and compatible evidence
• Finds OCR readings, symbols, and arrowheads as separate evidence families
• Finds geometry carriers from Line/Arc, Vector Lines, ridge, or fast page-wide lines
• Groups mechanical clues into possible dimensions
• Keeps comparison methods independently visible

Produces

• Table proposals and verification outcomes
• OCR, symbol, and arrowhead evidence
• Geometry evidence and possible-dimension bundles
• Provenance, cache identity, and stage timing

Does not claim

• Table type, headers, values, or title-block fields as finished understanding
• Dimension type, units, view ownership, or measured-feature ownership
• Confirmed dimensions
• Validated accuracy for tables or dimensions

Next stage receives

Understand receives Detection evidence plus append-only human review. Understand assigns ownership, relationships, and engineering meaning.

Product emphasis

Table Detection and Dimension Detection are first-class Detection sections. See Table Detection Overview and Dimension Detection Overview.

Open Questions

• Which Detection caches must always warm before a sheet enters Understand?

Sources

• Current contract: table-dimension-detection-architecture.md (2026-08-12)
• Status vocabulary: Status Vocabulary