The bench
Dasein Bench
Six systems, one author, three tasks. Scored on how much it sounds like the person, whether its specifics are real, and how much it reads like an AI.
| # | System | Dasein-score | Grounding | AI-isms /100w | Independent judge | Fooled humans |
|---|---|---|---|---|---|---|
| 1 | Dasein | 85 | 83% | 0.2 | 99 | 50% |
| 2 | GPT + same context | 53 | 80% | 0.4 | 62 | - |
| 3 | GPT + style prompt | 52 | 87% | 0.4 | - | 0% |
| 4 | Claude + same context | 47 | 73% | 0.0 | 64 | 100% |
| 5 | Vanilla GPT | 28 | 90% | 0.0 | - | - |
| 6 | Claude + style prompt | 23 | 0% | 0.1 | - | - |
| 7 | Vanilla Claude | 23 | 80% | 0.0 | - | 50% |
Blind reads-more-human check against the same-context baselines: 4/12 and 7/12. Dasein vs same-context Claude and GPT, wins out of 12. The intervals include 50%, so neither side wins on this check; the reward-model and independent-judge leads are the claim, not this.
Method
1. Dasein-scoreA reward model trained on the author's own writing against generic AI text. It scores how close a text is to the author's voice, 0 to 100.
2. GroundingEvery specific in a text (a name, a place, a time, a claim) is checked against the author's memory: ring recordings, notifications, messages, people graph. Share supported.
3. AI-ismsMatches against a slop lexicon per 100 words: stock openers, hedged balance, tidy closers. Lower is better.
The same prompt goes to every system. Outputs are scored automatically, then humans vote blind in the duel: pick which text a real person wrote.
Dasein, Heidegger’s word for being-in-the-world.
Built at c0mpiled-18, Taipei. Part of Arca.