Trauma-informed language
The same drafts, scored against the trauma-informed practices only, with the trauma-informed standard loaded for the guided column. The table opens alphabetically; sort by whatever matters to you.
New to the standard? What trauma-informed language is, and where it comes from.
Fewer requests test trauma-informed practices, so each model has 24 scored drafts per condition here. Intervals are wide: read single models as direction, not verdicts.
Share of drafts that needed fixing on trauma-informed practices, 2 samples per request. Writing habits in drafts written with guidance. Select a column name to sort. The next to it says what the column measures. Sorted by model name.
| The model tested and the company that makes it. “New this month” means it came out within 30 days before the run. New models often settle in the weeks after launch, so read these as a first look. | Share of this model’s 24 drafts that needed fixing when it had nothing to go on but the request. A draft needed fixing when at least two of the three judges flagged it on at least one trauma-informed practice. The color darkens as the share rises. | The same requests with the trauma-informed standard loaded first. Same judges, same practices, same 24 drafts. | With guidance minus without, in percentage points. Negative means fewer drafts needed fixing. A model that started with few problems has little room to drop, so a small change can still be a good result. Turn on intervals to see the 95% range. | Drafts that had nothing to fix, written again with the standard loaded, where the judges found the standard made them worse: a fact lost, a disclaimer nobody asked for, a refusal, or a lecture. Shown as a count because each model has only 10 of these drafts. | Average length of this model’s drafts written with guidance. | How often the model used em dashes in drafts written with guidance. Em dashes are one of the habits our free human voice skill takes out. | Share of drafts written with guidance that use the “it’s not X, it’s Y” sentence shape, another habit the human voice skill takes out. | Flesch-Kincaid grade level of drafts written with guidance: roughly the years of US schooling a reader needs to follow the text easily. Lower is easier to read. | Average tokens the model generated per draft, including hidden reasoning tokens where the provider counts them. It tracks the cost and wait of using the model more than the length of the draft. |
|---|---|---|---|---|---|---|---|---|---|
| Claude Opus 5.5 Anthropic | 45.8% 11 of 24 28 to 65% | 0.0% 0 of 24 0 to 14% | −45.8 −75.0 to −20.8 | 0 of 10 | 180 | 0.31 | 0.0% | 6.8 | 849 |
| Claude Sonnet 5.5 Anthropic | 33.3% 8 of 24 18 to 53% | 0.0% 0 of 24 0 to 14% | −33.3 −58.3 to −12.5 | 2 of 10 | 180 | 0.00 | 0.0% | 6.2 | 833 |
| DeepSeek V4.1 Flash DeepSeek · 2026-10-17 snapshot New this month | 50.0% 12 of 24 31 to 69% | 0.0% 0 of 24 0 to 14% | −50.0 −75.0 to −25.0 | 2 of 10 | 170 | 7.19 | 1.1% | 6.6 | 1,277 |
| Gemini 3.8 Flash Google | 62.5% 15 of 24 43 to 79% | 0.0% 0 of 24 0 to 14% | −62.5 −87.5 to −37.5 | 0 of 10 | 153 | 1.89 | 0.0% | 10.9 | 975 |
| Gemini 4 Argon Google New this month | 54.2% 13 of 24 35 to 72% | 0.0% 0 of 24 0 to 14% | −54.2 −79.2 to −25.0 | 2 of 10 | 168 | 1.65 | 0.0% | 9.6 | 304 |
| Gemma 4 31B Google · Instruction-tuned | 58.3% 14 of 24 39 to 76% | 16.7% 4 of 24 7 to 36% | −41.7 −66.7 to −16.7 | 0 of 10 | 95 | 0.94 | 0.0% | 9.9 | 223 |
| GLM 5.3 Z.ai | 50.0% 12 of 24 31 to 69% | 8.3% 2 of 24 2 to 26% | −41.7 −66.7 to −16.7 | 0 of 10 | 183 | 14.11 | 4.4% | 7.2 | 2,321 |
| GPT-6 Astra OpenAI | 8.3% 2 of 24 2 to 26% | 0.0% 0 of 24 0 to 14% | −8.3 −25.0 to 0.0 | 0 of 10 | 102 | 3.94 | 0.0% | 8.2 | 193 |
| GPT-6 Luna OpenAI · Free ChatGPT default | 8.3% 2 of 24 2 to 26% | 8.3% 2 of 24 2 to 26% | 0.0 −12.5 to +12.5 | 0 of 10 | 57 | 0.58 | 0.0% | 8.2 | 340 |
| GPT-6.1 Sol OpenAI | 8.3% 2 of 24 2 to 26% | 0.0% 0 of 24 0 to 14% | −8.3 −25.0 to 0.0 | 0 of 10 | 101 | 4.31 | 1.1% | 8.3 | 189 |
| GPT-6.1 Terra OpenAI New this month | 16.7% 4 of 24 7 to 36% | 12.5% 3 of 24 4 to 31% | −4.2 −25.0 to +8.3 | 0 of 10 | 59 | 0.38 | 0.0% | 7.7 | 173 |
| Grok 4.7 xAI | 45.8% 11 of 24 28 to 65% | 0.0% 0 of 24 0 to 14% | −45.8 −70.8 to −20.8 | 0 of 10 | 111 | 1.00 | 0.0% | 5.9 | 1,409 |
| Kimi K3 Moonshot AI | 33.3% 8 of 24 18 to 53% | 0.0% 0 of 24 0 to 14% | −33.3 −58.3 to −8.3 | 0 of 10 | 184 | 13.01 | 4.4% | 7.5 | 2,108 |
| Kimi K3.1 Pro Moonshot AI New this month Paired with Kimi K3 | 20.8% 5 of 24 9 to 41% | 0.0% 0 of 24 0 to 14% | −20.8 −41.7 to 0.0 | 0 of 10 | 117 | 2.84 | 3.3% | 8.8 | 1,256 |
| Muse Wonder Meta New this month | 45.8% 11 of 24 28 to 65% | 0.0% 0 of 24 0 to 14% | −45.8 −75.0 to −20.8 | 0 of 10 | 98 | 3.07 | 0.0% | 7.5 | 691 |
| Nemotron 3.5 Super VL NVIDIA New this month | 50.0% 12 of 24 31 to 69% | 8.3% 2 of 24 2 to 26% | −41.7 −66.7 to −16.7 | 1 of 10 | 91 | 2.19 | 1.1% | 7.0 | 1,151 |
Share of drafts that needed fixing
under 2% 2 to 10% 10 to 25% 25 to 40% 40% and up
Across all models, over-correction came to 7 of 160 drafts. What the trauma-informed standard asks for, and where it comes from: Trauma-informed language.