Beyond Visual Similarity: Auditing Structural Editability in Diagram Reconstruction
Abstract
An image-to-diagram reconstruction can be visually convincing while remaining difficult to revise: connectors may be detached, labels may be outlines, and an apparently native document may contain little more than a raster image. This paper studies how to audit the exported artifact without access to a reconstruction system's proprietary internals. We distinguish semantic recovery, package-level object structure, and executable editing behavior, and define an evaluation vector with explicit missing-denominator rules. The empirical material comprises eleven publicly downloadable demonstration cases with VSDX, draw.io, and PowerPoint counterparts, including English flowcharts, architecture diagrams, and a low-resolution Chinese example. On 21 October 2026, all eleven VSDX downloads matched the corresponding local public assets byte-for-byte. A fresh structural audit found 300 VSDX Shape elements, 173 shapes containing text, and 228 endpoint connection records, all referencing existing shape identifiers. These counts establish inspectable structure, not semantic accuracy. Thirty-three saved draw.io variants provide file-level metamorphic checks, and 200 semantic fixtures expose complementary metric failure modes. Visual examples show why rerouting, branch interpretation, and format-dependent decomposition require separate examination. The resulting protocol supports reproducible artifact assessment while withholding production prompts and implementation. The study is a descriptive, author-affiliated convenience-case analysis; it does not rank converters or establish human editing performance.
Unreviewed research preprint, version 2. The author develops LayerBack (https://layerback.com/) and discloses a commercial interest. AI assistance supported drafting, analysis code and figure preparation. Public demonstration cases and controlled diagnostics do not establish converter accuracy, comparative superiority or human editing performance. This deposit contains the manuscript only; proprietary implementation and customer data are not included.
Related
- Mechanical property reconstruction and damage evolution of rock-like materials for deep argillaceous sandstone under multiple similarity constraints
- Lee-Yang zeros of modulated XY spin chains with Dzyaloshinskii-Moriya interaction: zero-contour topology and quantum phase-diagram reconstruction
- Blind-spot maps: auditing what a structural equation model’s design could not have detected
- Cluster-based Structural Similarity for Dataset Visualization and Data Selection for Machine Learning Interatomic Potentials
- Local atomic environments reveal structural controls on MgSiO3 liquid-crystal similarity
BibTeX
@misc{chen2026beyond,
title = {{Beyond Visual Similarity: Auditing Structural Editability in Diagram Reconstruction}},
author = {Davie Chen},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23556459},
url = {https://zenodo.org/records/23556459},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Beyond Visual Similarity: Auditing Structural Editability in Diagram Reconstruction
AU - Davie Chen
PY - 2026
DA - 2026/10/22
PB - Zenodo
DO - 10.5281/zenodo.23556459
UR - https://zenodo.org/records/23556459
AB - An image-to-diagram reconstruction can be visually convincing while remaining difficult to revise: connectors may be detached, labels may be outlines, and an apparently native document may contain little more than a raster image. This paper studies how to audit the exported artifact without access to a reconstruction system's proprietary internals. We distinguish semantic recovery, package-level object structure, and executable editing behavior, and define an evaluation vector with explicit missing-denominator rules. The empirical material comprises eleven publicly downloadable demonstration cases with VSDX, draw.io, and PowerPoint counterparts, including English flowcharts, architecture diagrams, and a low-resolution Chinese example. On 21 October 2026, all eleven VSDX downloads matched the corresponding local public assets byte-for-byte. A fresh structural audit found 300 VSDX Shape elements, 173 shapes containing text, and 228 endpoint connection records, all referencing existing shape identifiers. These counts establish inspectable structure, not semantic accuracy. Thirty-three saved draw.io variants provide file-level metamorphic checks, and 200 semantic fixtures expose complementary metric failure modes. Visual examples show why rerouting, branch interpretation, and format-dependent decomposition require separate examination. The resulting protocol supports reproducible artifact assessment while withholding production prompts and implementation. The study is a descriptive, author-affiliated convenience-case analysis; it does not rank converters or establish human editing performance. Unreviewed research preprint, version 2. The author develops LayerBack (https://layerback.com/) and discloses a commercial interest. AI assistance supported drafting, analysis code and figure preparation. Public demonstration cases and controlled diagnostics do not establish converter accuracy, comparative superiority or human editing performance. This deposit contains the manuscript only; proprietary implementation and customer data are not included.
ER -