Coverage & governance FUNGI
How much evidence backs each ranking — a generator's rank is only as trustworthy as the votes behind it, and confidence firms up as vote counts grow. Published in full so a rank can be read with its uncertainty, not in isolation.
Generator coverage — 53 generators with arena outputs, by total Mode-A votes
| Generator | Tasks | Outputs | A-votes | Confidence | In arena |
|---|---|---|---|---|---|
| Tripo H3.1 text model | 16 | 16 | 242 | firm | ✓ |
| Hunyuan3D v3 text model | 16 | 16 | 241 | firm | ✓ |
| Tripo P1 text model | 15 | 15 | 229 | firm | ✓ |
| Meshy v6 text model | 14 | 14 | 219 | firm | ✓ |
| Rodin text via fal model | 14 | 14 | 218 | firm | ✓ |
| Hunyuan3D 3.1 text model | 13 | 13 | 195 | firm | ✓ |
| SAM 3D model | 15 | 16 | 158 | firm | ✓ |
| Rodin text via Replicate model | 11 | 11 | 145 | firm | ✓ |
| Meshy 6 model | 11 | 11 | 142 | firm | ✓ |
| Pixal3D model | 11 | 11 | 125 | firm | ✓ |
| Hunyuan3D v3 model | 12 | 15 | 115 | firm | ✓ |
| Hunyuan3D 3.1 model | 12 | 15 | 112 | firm | ✓ |
| Hunyuan3D v2 model | 11 | 12 | 95 | firm | ✓ |
| Rodin/Hyper3D model | 11 | 15 | 91 | firm | ✓ |
| TRELLIS via Replicate model | 11 | 12 | 75 | firm | ✓ |
| TRELLIS via fal model | 8 | 9 | 62 | firm | ✓ |
| TRELLIS 2 model | 6 | 6 | 57 | firm | ✓ |
| TripoSR model | 7 | 7 | 42 | firm | ✓ |
| deepseek/deepseek-v3.2 model | 14 | 14 | 16 | provisional | ✓ |
| anthropic/claude-opus-4.8 (agentic) model | 11 | 11 | 15 | provisional | ✓ |
| anthropic/claude-sonnet-5 model | 13 | 13 | 15 | provisional | ✓ |
| google/gemini-3.1-pro-preview (agentic) model | 12 | 12 | 14 | provisional | ✓ |
| openai/gpt-5.6-sol model | 14 | 14 | 14 | provisional | ✓ |
| x-ai/grok-4.20 model | 14 | 14 | 14 | provisional | ✓ |
| openai/gpt-5.6-sol (agentic) model | 12 | 12 | 13 | provisional | ✓ |
| anthropic/claude-opus-4.8 model | 12 | 12 | 12 | provisional | ✓ |
| z-ai/glm-5.2 model | 12 | 12 | 12 | provisional | ✓ |
| anthropic/claude-sonnet-5 (agentic) model | 9 | 9 | 11 | provisional | ✓ |
| mistralai/mistral-medium-3-5 model | 12 | 12 | 10 | provisional | ✓ |
| qwen/qwen3.7-plus model | 13 | 13 | 10 | provisional | ✓ |
| deepseek/deepseek-v4-pro model | 10 | 10 | 9 | provisional | ✓ |
| meta-llama/llama-4-maverick model | 6 | 6 | 9 | provisional | ✓ |
| moonshotai/kimi-k2.7-code model | 10 | 10 | 9 | provisional | ✓ |
| x-ai/grok-4.5 model | 13 | 13 | 9 | provisional | ✓ |
| x-ai/grok-4.5 (agentic) model | 9 | 9 | 9 | provisional | ✓ |
| z-ai/glm-4.6v (agentic) model | 8 | 8 | 9 | provisional | ✓ |
| google/gemini-3.1-pro-preview model | 11 | 11 | 8 | provisional | ✓ |
| qwen/qwen3.7-plus (agentic) model | 8 | 8 | 8 | provisional | ✓ |
| x-ai/grok-4.20 (agentic) model | 7 | 7 | 8 | provisional | ✓ |
| z-ai/glm-4.6v model | 8 | 8 | 8 | provisional | ✓ |
| minimax/minimax-m3 model | 8 | 8 | 7 | provisional | ✓ |
| openai/gpt-5.1 (agentic) model | 6 | 6 | 7 | provisional | ✓ |
| meta-llama/llama-4-maverick (agentic) model | 6 | 6 | 6 | provisional | ✓ |
| minimax/minimax-m3 (agentic) model | 6 | 6 | 6 | provisional | ✓ |
| mistralai/mistral-medium-3-5 (agentic) model | 4 | 4 | 5 | provisional | ✓ |
| openai/gpt-5.6-sol-pro (agentic) model | 4 | 4 | 5 | provisional | ✓ |
| moonshotai/kimi-k2.7-code (agentic) model | 4 | 4 | 4 | provisional | ✓ |
| qwen/qwen3.6-plus model | 4 | 4 | 4 | provisional | ✓ |
| x-ai/grok-4.3 model | 4 | 4 | 4 | provisional | ✓ |
| openai/gpt-5.1 model | 6 | 6 | 3 | provisional | ✓ |
| InstantMesh model | 1 | 1 | 1 | provisional | ✓ |
| openai/gpt-5.6-sol-pro model | 1 | 1 | 1 | provisional | ✓ |
| qwen/qwen3.6-plus (agentic) model | 1 | 1 | 1 | provisional | ✓ |
How models enter & compete
- Entry — every output is added by an administrator or via the public submission queue, then moderated before it enters the arena. There is no private or preferential pre-testing: a model is not quietly trialled and published only if it scores well.
- Fair matchmaking — pairs are drawn by a sampler that biases toward the least-compared outputs, so coverage spreads evenly rather than concentrating votes on a favoured few. Every model is drawn from the same pool by the same sampler.
- No silent deprecation — outputs are not removed to flatter the board; the arena has no deprecation mechanism, so what competed stays on the record.
- Exclusions are principled, not selective — only reference ground-truth scans and untextured geometry-only outputs are held out of the Mode-A perceptual pool (they confound a visual-preference vote); they remain in the objective Mode-B board. Excluded generators are flagged in the table above.
Reading a rank
A generator's rank is only as trustworthy as the votes behind it. Below 30 total Mode-A votes a rank is marked provisional; at or above it, firm. Bradley–Terry confidence intervals on the leaderboard make the same uncertainty visible per generator.
Current limits, stated plainly: the arena is in an internal evaluation phase, so vote volume is low and most ranks read provisional — ranks will firm up as voting scales. Mode-A votes here are internal and not yet from a public pool.
Task coverage — 6 active tasks
"Mode-B" = objective scoring against held-out ground-truth is available for that task.
| Task | Category | Tier | Generators | Outputs | A-votes | Judge votes | Mode-B | Mode-C |
|---|---|---|---|---|---|---|---|---|
| Morchella esculenta — single-image → 3D reconstruction | Fungi | moderate | 29 | 29 | 105 | 0 | ✓ | — |
| Amanita muscaria — single-image → 3D reconstruction | Fungi | moderate | 28 | 28 | 85 | 0 | ✓ | — |
| Hericium erinaceus — single-image → 3D reconstruction | Fungi | hard | 26 | 28 | 73 | 0 | ✓ | — |
| Boletus edulis — single-image → 3D reconstruction | Fungi | easy | 26 | 26 | 91 | 0 | ✓ | — |
| Lycoperdon perlatum — single-image → 3D reconstruction | Fungi | easy | 26 | 26 | 86 | 0 | ✓ | — |
| Trametes versicolor — single-image → 3D reconstruction | Fungi | hard | 17 | 17 | 76 | 0 | ✓ | — |
Mode-C — botanical-trait accuracy experimental
Generator-level mean botanical accuracy — a 3D model graded against a literature-sourced per-taxon trait rubric by a calibrated VLM trait-checker. Only trait classes that have passed the human↔VLM agreement gate count toward the score. No class has passed the gate yet, so this axis is experimental: shown for inspection, not yet a ranking signal.
No scored outputs yet. Per-output scorecards live at /trait/<id>.
Machine-readable: /api/coverage.json.