OCR, Symbols, and Arrowheads

Status: Current Detection evidence families

Authority: Current contract and current OCR implementation

Last reviewed: 2026-08-14

Question

Why are OCR, symbols, and arrowheads separate Detection evidence families?

Short answer

Each family is mechanical evidence. Each family keeps its own provenance. Understand later assigns roles such as dimension text, note text, or symbol meaning. Detection must not hide that boundary.

OCR

OCR means optical character recognition. The current OCR engine is PaddleOCR. The current model pair is PP-OCRv6_medium_det plus PP-OCRv6_medium_rec. OCR readings can contain measurements, notes, or table text. An OCR reading is not yet a typed dimension or table field.

Symbols

Symbol evidence marks candidate marks on the page. Symbol evidence remains a Detection family until Understand assigns meaning. Do not treat a symbol detector score as finished engineering semantics.

Arrowheads

Arrowhead evidence supports dimension carrier search and pairing. Fast Dimension Lines uses local arrow clues after the page-wide straight-line pass. An arrowhead clue is not a confirmed dimension by itself.

Shared rule

Keep OCR, symbols, and arrowheads independently reviewable. Do not fuse them into hidden meaning inside Detection. Cache reuse should warm these families without forcing full sheet assembly.

Open Questions

• Which symbol classes become first-class Detection outputs before Understand work begins?

Sources

• Current contract: table-dimension-detection-architecture.md (2026-08-12)
• Current OCR implementation: Detection worktree paddle_ocr.py (PP-OCRv6)
• Fast Dimension Lines