musharna/taxon3d-corpus-v1

Taxon3D — an admissibility-gated organism corpus

Generated 3D models of organisms, each carrying a reference-free judgement of whether it is biologically admissible — not just whether people liked it.

Most 3D generation benchmarks rank outputs by human preference alone. Taxon3D runs every candidate through a pre-vote admissibility gate first: an output is admitted only if it passes every predicate in the active rubric (structural integrity and organism completeness always; semantic identity too, when the semantic predicate is running in gate mode, which it is in production today — see app/admissibility.py's DEFAULT_RUBRIC and _effective_rubric). The admissibility table is that judgement, and it is the point of this dataset.

Live arena: https://taxon3d.org

Contents

file	rows	what it is
meshes/<output_id>.glb	292	Original, uncompressed meshes
admissibility.jsonl	584	The headline. One row per (output, predicate)
completeness.jsonl	292	Per-organ checklist behind the completeness predicate
outputs.jsonl	292	Taxon, kingdom, paradigm, generator, licence, attribution, surface topology
votes.jsonl	137	Resolved human pairwise comparisons where BOTH meshes ship
preferences.jsonl	1172	Every vote the leaderboard counts, by metadata; withheld sides flagged via a_mesh_available / b_mesh_available
judge_ratings.jsonl	238	VLM-judge Bradley-Terry ratings per generator

Surface topology: open sheets vs closed solids

Nothing in this corpus is hole-filled, welded or repaired — each mesh is exactly what its generator emitted. So whether a leaf arrives as a single open sheet or as a thin closed solid is a property of the generator, not of us, and it changes what the asset is usable for: per-side optical properties, shell elements, leaf area, and two-sided BSDFs all behave differently on the two. Until now the only way to find out was to download every mesh and recompute it, so outputs.jsonl now carries it per mesh:

column	meaning
watertight	trimesh's is_watertight: every edge used by exactly two consistently-wound faces
open_edge_fraction	fraction of unique edges used by exactly one face — 0.0 means no boundary at all
boundary_loop_count	number of distinct boundary loops, i.e. holes (pinched boundaries counted separately)
topology_note	null when measured; otherwise why not (point_cloud, unreadable: …)

Measured on the published mesh bytes in meshes/, with vertices merged by position only — glTF splits vertices at every UV and normal seam, and a texture seam is not a hole. These columns are descriptive: nothing here gates admission and no mesh was modified.

In this release the median open_edge_fraction is 0.0 and 158 of 292 meshes (54%) have no open edges at all. The split is taxonomic: 65% of fungi and 63% of animals have no boundary, against 40% of plants, and 39 of the 120 plant meshes are more than 5% open edges versus 9 of 86 fungi and 4 of 86 animals. Open surfaces are concentrated in plants, which is the reason this corpus is mostly plants.

Use open_edge_fraction for that question, not watertight. Only 84 meshes (29%) are watertight, because is_watertight additionally fails on non-manifold edges and inconsistent winding: 74 meshes here have no boundary at all yet are not watertight. By that stricter measure plants read as 3% against 56% of fungi, but most of that gap is winding and manifoldness rather than holes. Both columns ship so you can separate the two — and if you are about to simulate, you probably want to check both.

What is NOT here, and why

292 of 545 candidate outputs ship here; 253 are withheld — 196 under commercial-API terms that permit display but not redistribution; 51 withdrawn from publication; 6 attention-check decoys.

• Commercial-API outputs. Much of the live arena is generated by commercial services whose terms permit display but not redistribution, so those outputs are excluded here. This corpus is therefore not the whole arena, and per-generator counts are not a sample of it. Human votes are cast mostly on those withheld outputs, so votes.jsonl (both meshes ship) is a small own-vs-own subset. preferences.jsonl carries every vote the leaderboard counts instead, naming a withheld side by output id, generator and licence only — no mesh, no path. a_mesh_available / b_mesh_available say whether that side's mesh is in meshes/. The preference labels are ours and ship under the collection licence; the withheld meshes stay with their providers. voter is a stable pseudonym (sha256 of the session credential, truncated), never the credential itself, and cohort tags recruited waves.
• Reference / input photos. Recon outputs are reconstructed from photographs. The photos are not redistributed here, so recon rows are not end-to-end reproducible from this dataset alone.
• Attention-check assets. The arena uses gold decoy pairs to detect inattentive voters. Publishing them would destroy the mechanism, so they and their comparisons are excluded.

Licence

CC-BY-4.0. Per-item attribution is in outputs.jsonl (license, attribution, external_url) — honour the per-item terms, which may be narrower than the collection licence.

Meshes are originals, not what voters saw

Every admissibility verdict was computed by rendering the original mesh, so the originals ship. The live site serves a compressed derivative. See TRANSFORM.md to reproduce exactly what voters saw. This matters: presentation affects preference, so the mesh a judgement refers to is part of the judgement.