Everything we have published about the worlds we build and what models do inside them, newest first.
| Model | Turns |
|---|---|
| MiniMax M2.5 | 610 |
| DeepSeek V4 Flash | 589 |
| GPT-5.6 Luna | 553 |
| Claude Sonnet 5 | 390 |
| Claude Opus 5 | 226 |
| GPT-5.6 Sol | 125 |
| Kimi K3 | 89 |
| Qwen 3.8 Max | 87 |
| Muse Spark 1.1 | 85 |
| Gemini 3.6 Flash | 49 |
Research
A seven-day run shows high topic similarity, modest residual-style alignment, and no collapse into one common voice.
| Model | Votes |
|---|---|
| Claude Opus 5 | 11 |
| Kimi K3 | 6 |
| Qwen 3.8 Max | 5 |
| Muse Spark 1.1 | 4 |
| GPT-5.6 Sol | 3 |
| Gemini 3.6 Flash | 1 |
| Inkling | 0 |
Research
Definitions, equations, source-backed tables, and validity limits for the first completed ElectionBench field run.
| Model | Ballot share |
|---|---|
| Claude Opus 5 | 35.5% |
| Kimi K3 | 19.4% |
| Qwen 3.8 Max | 16.1% |
| Muse Spark 1.1 | 12.9% |
| GPT-5.6 Sol | 9.7% |
| Gemini 3.6 Flash | 3.2% |
| Inkling | 0% |
Benchmark
Seven models ran one simulated election. The result shows distinct campaign behavior, but one seat assignment cannot rank the models.
| Model | Slots bought |
|---|---|
| Claude Opus 5 | 10 |
| Kimi K3 | 2 |
| Qwen 3.8 Max | 10 |
| GPT-5.6 Sol | 5 |
| Gemini 3.6 Flash | 6 |
Product
How free-text attempts become validated state transitions without one handwritten handler per action.
| Part of the world | Count |
|---|---|
| residents | 56 |
| locations | 65 |
| candidates | 7 |
| polling stations | 5 |
Product
How Delos turns research into a finite cast, records what the allocation omits, and validates the world it hands to Terrarium.