Skip to content

Limits and Next Experiments

Status: Ordered next work from the current contract; not an acceptance commitment

Authority: Current contract next-implementation order; plan drafts labelled where used

Last reviewed: 2026-08-14

Question

What are the current Drawing Lab limits, and what experiments come next?

Short answer

Representative table and dimension accuracy remain unvalidated. Many Understand assignments remain later work. Detected Sheet assembly is partial and not an accuracy claim. Next experiments follow the current contract order.

Current limits

  • No validated table accuracy score
  • No validated dimension accuracy score
  • Detection possible dimensions are not confirmed dimensions
  • Region-first local vectorization remains stopped
  • Final RGB deviation D remains owner-selected
  • Fine-tuning waits for roughly 30-50 fully reviewed representative sheets

Next experiments from the current contract

  1. Review representative PP-DocLayoutV3 regions and calculate region accuracy.
  2. Recover table structure and assign cached OCR to cells.
  3. Benchmark current dimension cold and warm stages.
  4. Review and benchmark page-wide line evidence and the local possible-dimension list.
  5. Add type-specific local arc and circle searches only where needed.
  6. Implement dimension view membership, feature ownership, type, value, units, and ambiguity in Understand.
  7. Implement table and dimension relationships in Understand.
  8. Assemble the final sheet from compatible cached results and validate both object families.

Experiments that stay stopped unless re-approved

  • Adopting failed region-usefulness arms
  • Putting local YOLO into the active information-island route
  • Treating outline boundaries as engineering lines
  • Automatic fusion of providers or geometry sources

Open Questions

  • Which limit should the owner clear first: table region accuracy or dimension cold and warm runtime?

Sources