Museums and heritage communities increasingly release 3D scans of temples, statues and embroidered garments so that games, tourism apps and virtual worlds can reuse them. On the way to market, these assets are compressed, pruned and restyled. The small, locally specific details that carry cultural meaning, such as a carved motif, a stitch pattern or an inscription, are often the first to disappear. Existing tools can prove who owns a file or reconstruct which file was derived from which. None can tell a community what was lost in a derivative, or whether the asset has drifted toward a generic "heritage look."
We propose a meaning-level receipt. Before release, the custodian publishes a commitment that the custodian cannot later alter. It lists which parts of the 3D scan the community considers meaning-bearing, along with a compact reference description of each part. Any commercial derivative that claims to be "based on" the scan can then be scored part by part on two quantities: how much of each designated part survives, and how far each part has moved toward a pool of generic, non-local examples. The tamper-proof log does not decide what is authentic; people do. The log's job is to stop anyone from quietly redefining what mattered after the fact.
Core Claim
The claim is that commercial simplification of 3D heritage assets destroys cultural meaning in a way global image-quality scores miss. A pruned scan can keep high overall rendering fidelity while erasing exactly the fine elements a community cares about. A per-element retention and homogenization score, anchored to a community-authored manifest committed before release, tracks expert judgments of meaning loss better than global fidelity, whole-scene embedding similarity, or post-hoc provenance reconstruction.
The design choice that makes this workable addresses a theoretical obstacle. AUDITA builds signed, Merkle-committed logs with causal blame scores. It also proves a "completeness barrier": an auditor limited to the record cannot be perfectly accurate unless it can solve the underlying task. For culture, the underlying task ("is this authentic?") has no solver. We do not try to get around the barrier. We move the hard judgment into the committed record, so that the auditor only has to solve a measurable problem: whether the designated elements are retained and whether they have drifted toward generic examples.
Background and Positioning
The motivating evidence comes from qualitative commercialization studies. Sichuan Opera face-changing, compressed into restaurant entertainment, loses its narrative and ritual frame (Hong et al.). Tourist apparel simplifies She embroidery and drops its symbolic logic (Chen). Oroqen dance becomes monotonous when adapted for tourism (Dang). All three document degradation case by case, and none offers a measurement. Vettore supplies the architectural version of the worry: Venetian conservation that protects facades while discarding constructive substance. That critique warns us that a score rewarding surface resemblance alone would reproduce facadism. This is why our reference descriptions are per element, and why the second score measures drift toward the generic.
On the 3D side, Gaussian Heritage supplies the method we reuse to split a scan into parts. It builds scans from smartphone video using 3D Gaussian Splatting (3DGS), a representation that models a scene as millions of small colored, semi-transparent blobs, and attaches a segmentation feature to each blob. It reports 80.3 mIoU on 3D-OVS, a benchmark for segmenting 3D scenes from text queries. It says nothing about what happens to the parts after export.
Two papers make the gap sharp. GuardSplat hides ownership bits in the spherical-harmonic color coefficients of 3DGS, the terms that encode view-dependent color. It reports over 99% bit accuracy at about 41.55 dB PSNR on Blender and LLFF, but it answers only "whose file is this." Its robustness is tested against 2D image distortions. We found no reported result on Lacuna for GuardSplat under removal of higher-order color coefficients, which is standard in 3DGS compression. GaussTrace reconstructs which models were derived from which, using statistics of blob parameters and a language model's step-by-step reasoning. It reaches a vertex-edge overlap of 0.948, against 0.566 for the best baseline, on a 44-model benchmark. It detects parametric divergence, however, not semantic loss, and it was not evaluated on heritage assets.
FG-Trac supplies the logging method we borrow. It batches pipeline events into Merkle trees and anchors their roots, at under 2% epoch overhead on CIFAR-10. Finally, Dong et al. restyle Hong Kong facades with Hakka textures using CycleGAN. In our setting that is not a competitor but a test operator: a restyling transform whose outputs look "heritage" while possibly being homogenized.
Research Program
The system has three parts.
Manifest. For each scanned asset, the custodian works with community practitioners to mark meaning-bearing elements on the segmented 3DGS scan. Each element $e$ receives a reference descriptor. The descriptor is a set of embeddings of the element rendered from fixed canonical views, computed with a frozen image encoder, plus an optional fine-texture descriptor for patterns such as stitches or brushwork. The manifest, including the element masks, descriptor hashes and the encoder version, is committed as a Merkle root to an append-only signed log in the style of FG-Trac.
Derivative record. A platform that publishes a derivative submits it together with a signed record of the transforms it applied.
Audit. The auditor locates each element in the derivative by transferring the committed 3D mask. It renders the element from the same canonical views and computes two scores:
$$R_e = \cos\big(f(D_e), f(O_e)\big), \qquad H_e = \frac{d\big(O_e, g_e\big) - d\big(D_e, g_e\big)}{d\big(O_e, g_e\big)}.$$
Here $f$ is the frozen encoder. $O_e$ and $D_e$ are the canonical renders of element $e$ in the original and the derivative. $d$ is cosine distance in encoder space. $g_e$ is the centroid of a generic pool for that element type, for example stock or mass-market renderings of the same kind of object from outside the community. $R_e$ measures survival. $H_e$ measures the fraction of the gap to the generic centroid that the derivative has closed. A derivative can keep $R_e$ moderately high while still showing large $H_e$, and that combination is exactly the homogenization signature.
Scope. The contribution is the per-element, pre-committed measurement and its validation against human judgment. The cryptography is off-the-shelf.
Baselines
The naive reference point is whole-scene rendering fidelity: PSNR and LPIPS between original and derivative, averaged over the same canonical views. This is what platforms already optimize, so if it tracks expert judgments as well as our scores, the proposal fails.
The strongest competitor on "what changed" is GaussTrace. Its best reported figure is VEO 0.948 on its provenance benchmark, from GaussTrace. We will run its released feature extraction on our derivatives and use its per-attribute divergence statistics as a predictor of expert-rated loss. It gets the same derivatives, the same number of language-model calls per asset as our pipeline uses for mask transfer, and no manifest.
The strongest competitor on ownership survival is GuardSplat, embedded in the original before derivation. Its reported figure is over 99% bit accuracy on Blender and LLFF. Beating it means showing that the manifest link survives transforms under which its bit accuracy collapses.
A whole-scene embedding baseline, the cosine similarity of frozen-encoder embeddings of full renders, isolates the value of going per element.
Two ablations isolate the key ingredients. The first replaces the community-authored committed manifest with elements chosen after the fact by open-vocabulary segmentation prompts. This tests whether pre-committed human designation matters or whether automatic part discovery suffices. The second drops $H_e$ and keeps only $R_e$, to test whether the generic-drift term carries information beyond retention.
All methods are tuned on the same held-out validation assets with the same search budget. Beating a baseline means a statistically higher rank correlation with expert meaning-loss ratings on held-out assets, with bootstrap intervals that exclude zero difference.
Why not just…?
Why not just use PSNR or LPIPS on the whole scene? This is the real threat. If meaning-bearing elements are large relative to the scene, global fidelity may track loss well enough. Our hypothesis rests on the fact that pruning and color-coefficient truncation preferentially remove small, high-frequency blobs, while global metrics are dominated by large smooth regions. The pruning and truncation arms of the benchmark test this directly. If global LPIPS matches our correlation with expert ratings, we will report that the per-element machinery is unnecessary for those transforms and restrict the claim to restyling.
Why not just run GaussTrace-style provenance analysis? GaussTrace answers which asset a derivative came from and what kind of edit was applied. It does not know which regions matter to anyone, so a large edit to a background wall and a small erasure of an inscription can look alike in its statistics. The benchmark includes matched pairs of this kind, equal parametric divergence with different elements affected, and we predict GaussTrace cannot separate them.
Why not just a signed hash in a central database, without Merkle anchoring? For integrity alone, a well-run signed log held by a neutral custodian may be enough. We treat the anchoring layer as replaceable. The part that must be tamper-evident is the manifest, because commercial operators have an incentive to redefine essentials after the fact. AUDITA's finding that unsigned records let forgers frame innocent parties in 100% of probes motivates signing. It does not motivate a public blockchain, and we make no claim that one is needed.
Why not just watermark the asset robustly? Watermarks prove ownership, not meaning retention. We will also measure where GuardSplat's bits fail under compression. Even a perfectly surviving watermark would say nothing about whether an embroidery motif survived.
Proposed Experiments
Benchmark construction. No heritage-derivative benchmark exists on Lacuna, so we build one. We capture roughly 30 heritage objects and small scenes with the Gaussian Heritage smartphone pipeline, prioritizing objects with fine meaning-bearing detail: carved wood, embroidered textiles, painted surfaces. Practitioners from the relevant communities annotate the meaning-bearing elements. For pipeline sanity checks we add non-heritage scenes from 3D-OVS and LERF-Mask, where segmentation quality is already measured.
Derivative operations. Each asset is put through a graded ladder of transforms: blob pruning at several ratios, truncation of view-dependent color terms, parameter quantization, the four edit types from GaussTrace, and texture restyling. The restyling includes a CycleGAN-style operator trained as in Dong et al. and an operator that pushes toward mass-market stock renders. For a 2D sanity check of $H_e$, we apply the same scoring to MuralDH murals before and after diffusion inpainting, asking whether inpainted regions drift toward generic motifs. The only damage-segmentation figure reported there is SAM-Adapter IoU 0.2237 from the MuralDH paper, and we found no reported result on Lacuna for motif fidelity after inpainting.
Decisive check. Practitioners and heritage experts, blind to the transform, rate each derivative on how much of the object's meaning survives. We compare Spearman correlation with these ratings across $R_e$ and $H_e$ aggregated over elements, global PSNR and LPIPS, whole-scene embedding similarity, and GaussTrace divergence. The proposal succeeds if the per-element scores, specifically with the committed manifest, beat all baselines on pruning and restyling. It is falsified if global LPIPS matches them within the confidence interval.
Gaming test. We simulate a platform that knows the manifest. It preserves designated elements at full quality while degrading their surroundings. This measures the attack surface honestly, and we report how much context loss the receipt misses.
Cost. We report manifest commitment and audit overhead against FG-Trac's under-2% figure.
References
- AUDITA: certified auditing and causal attribution of adverse outcomes in autonomous multi-agent systems
- Challenges and Opportunities of Sichuan Opera Face Changing in Haidilao in Cultural Communication
- Exploration on the Intangible Cultural Heritage Technique of She Ethnic Group in the Teaching of Fashion Design Under the Background of Cross-boundary Integration
- Study on the Characteristics and the Development Vein of Oroqen Dance Culture
- The role of craftsmanship in the conservation of Venice. State of the art and perspective
- Gaussian Heritage: 3D Digitization of Cultural Heritage with Integrated Object Segmentation
- GuardSplat: Efficient and Robust Watermarking for 3D Gaussian Splatting
- GaussTrace: Provenance Analysis of 3D Gaussian Splatting Models with Evidence-based LLM Reasoning
- Fine-Grained Traceability for Transparent ML Pipelines
- Urban Cultural Inheritance: Generative Adversarial Networks (GANs) Assisted Street Facade Design in Virtual Reality (VR) Environments Based on Hakka Settlements in Hong Kong
- A comprehensive dataset for digital restoration of Dunhuang murals