Measured evidence needs source diversity.
R18 adds a second independent benchmark source where the exact current model/configuration can be matched, while preserving older and neighboring model results as visible but excluded corroboration.
Current pilot confidence
| System | Measured score | Confidence | Exact source families | Exact corroboration | Related excluded |
|---|---|---|---|---|---|
| Claude | 74.4 | 72.0% | 1 | 0 | 2 |
| ChatGPT | 66.1 | 72.0% | 1 | 0 | 3 |
| Grok | 62.8 | 84.0% | 2 | 1 | 1 |
| Gemini | 58.9 | 72.0% | 1 | 0 | 2 |
| Qwen | 48.9 | 72.0% | 1 | 0 | 3 |
| DeepSeek | 43.9 | 84.0% | 2 | 1 | 2 |
| Cohere Command | 30.6 | 84.0% | 2 | 1 | 1 |
What changed
Exact-current ToneBench
Grok 4.6 high, DeepSeek V4 Pro 0813 max and Command A+ have an exact-current second independent benchmark source. Their R17 measured scores do not change, but confidence rises from 72% to 84%.
Related evidence is visible
SWE-bench Verified, BFCL V4, OSWorld, and model/config variants are attached to the relevant record but marked Related / Excluded when they do not match the current representative model.
No benchmark soup
Writing, coding, tool-use and computer-use raw percentages are not averaged into the 70/15/15 R17 score simply because they exist.
Next confidence target
Find exact-current second sources for ChatGPT, Claude, Gemini and Qwen, then add domain-complete independent coding/tool/agent cohorts before changing the score formula.