Skip to content

Topology and View Candidates

Status: Whole-sheet topology retained as protected baseline evidence; not accepted as semantic truth

Authority: Accepted decision DL-010 and evaluation evidence

Last reviewed: 2026-08-14

Question

What did topology and view-candidate work contribute, and why did it stop as a primary route?

Short answer

Topology helped diagnose contacts, intersections, and loops. View candidates helped expose page-decomposition failures. Most candidates were wrong or incomplete for product use. Projection identity remains an Understand concern.

What we tried

Endpoint contacts. Intersections and connectivity. Loops, frames, grids, and provisional view candidates on the whole sheet.

Why it helped

The work created diagnostic seeds and reviewable structure hypotheses. The review showed that many failures came from page context, not one local geometry bug.

Why it failed or stopped

Review evidence included 2/12 accepted emitted view candidates and 10/12 rejected. 39/40 CVAT-assisted complete views were missed by the product in the cited review. 11/12 blinded topology sites landed on irrelevant text, dimension, or title-block material. Topology p95 13.942 s exceeded the protected 10 s ceiling. Bounded geometry or topology correction was not accepted as the immediate next phase.

What we retained

  • Phase 2 primitives and Phase 3 topology as protected reusable baseline evidence
  • Optional seeds only, not automatic semantic views
  • Human review and region-first accounting before local specialized routing
  • Projection identity as Understand work, not Detection truth

YOLO note

Owner labeling later rejected local YOLO models for the active information-island route. Earlier full-page-versus-crop YOLO ideas remain available only for a later separately approved experiment. Accepted decision DL-010 supersedes that active next-step idea.

Open Questions

  • Which topology seeds remain useful once human-scale regions pass the usefulness gate?

Sources