Data and Code for "Knowledge Grounding as the Bottleneck in Text-to-3D Generation for Digital Twin Deployment"

# Data and code release — "Knowledge Grounding as the Bottleneck in Text-to-3D Generation for Digital Twin Deployment" This archive accompanies the manuscript submitted to *Advanced Engineering Informatics*. It contains everything needed to reproduce the cross-backbone KCS validation, the SG-CP causal-isolation experiment, the digital-twin readiness measurements, the attention-mechanism proxy, and the human-agreement check reported in Section 9 of the paper. ## What's in this lightweight package - `code/KCS_v3_Clean.ipynb` — the full experiment notebook: the 60-prompt dataset and hand-authored ground-truth scene graphs (Part 1), the KCS metric (Part 2), the generation/scoring pipeline (Parts 4–8), and the figure/table generation (Part 10). Includes a bootstrap cell that restores session state after any kernel restart by replaying the tagged definition cells, so the notebook is runnable end to end without external scaffolding. - `results/` — every result table and figure-input CSV/JSON referenced in the paper: cross-paradigm KCS summary (original and alternate weighting), SG-CP causal-isolation results, DT-readiness pass rates, attention-proxy scores, KCS weight-sensitivity sweep, calibrated-simulation fit, generation log, manifest, the manual-verification HTML scoring sheet, and the two LaTeX result tables used directly in the manuscript. - `human_verification/` — the 54-row stratified manual-verification sample with the human reviewer's scores and notes, the AI-assisted spot-check scores kept in a separate clearly-labelled column (never merged with the human columns), and the computed human-vs-automated agreement detail and summary statistics reported in Section 9.7. - `vlm_cache/` — every cached GPT-4o and Claude scene-graph extraction call (JSON), keyed by `<backbone>_id<NNN>__<judge>.json`. This is what the KCS scores in `results/` were actually computed from, and lets anyone re-derive every number in the paper without re-querying either API. ## What's in the companion large-asset archive (linked separately) The generated 3D outputs themselves are too large for a standard code-hosting package (multiple gigabytes) and are released as a separate archive: - `generation/` — every backbone's generated mesh/point-cloud and six-view renders for all 60 prompts (Shap-E, Point-E, TripoSR; SF3D excluded per the manifest's documented incompatibility). - `sg_cp/` — the 120 SG-CP causal-isolation outputs (vanilla / structural / ontology conditions), each with its composed mesh, renders, and `sgcp_meta.json` generation record. - `attention_proxy/` — the 2D cross-attention-proxy renders. [Companion archive DOI/URL: PLACEHOLDER — fill in once uploaded to Zenodo.] ## Reproducing a result from scratch 1. Open `code/KCS_v3_Clean.ipynb`, run the bootstrap cell (restores the dataset, ground truth, KCS implementation, and judge extractors). 2. Point `DATA_ROOT` at the large-asset archive's `generation/` directory (or `sg_cp/` for the causal-isolation experiment). 3. Run the scoring cells. Judge calls hit the cache in `vlm_cache/` first, so no API key is required to reproduce the published numbers — only to re-score genuinely new outputs. ## Citing this release If you use this dataset or code, please cite the manuscript above. A DOI for this archive will be added here once minted.