The model
DeepSeek V4 Flash at maximum reasoning effort in both the baseline and Kohnex-assisted arms.
The evidence, in the open
Proof over
promises.
Same model. Same maximum effort. The only variable is Kohnex.
Start with the methodologyExpert work. Under pressure.
DeepSeek V4 Flash at maximum reasoning effort.
100 custom tasks. Same model as judge.
DeepSeek V4 FlashModel alone42.7%
Kohnex BrainSame model + Kohnex65.2%22.5percentage
points
points
Internal evaluation on a custom task set. Not a public benchmark. Not independently audited.
Every category. The full picture.
20 tasks per categoryExact scores.
Published category results| Category | Model alone | With Kohnex | Gain |
|---|---|---|---|
| Cross-domain synthesis | 28.5% | 70.0% | +41.5 |
| Hostile debugging | 45.5% | 71.4% | +25.9 |
| Constrained design | 4.0% | 21.0% | +17.0 |
| Deceptive review | 58.5% | 73.0% | +14.5 |
| Multi-hop diagnosis | 77.0% | 90.5% | +13.5 |
How the numbers
came to be.
Good evidence starts with a fair comparison. Here is exactly what these evaluations measure, and what they do not.
The task set
100 custom expert tasks, with 20 tasks in each of five categories. Both arms use DeepSeek V4 Flash at maximum reasoning effort. Each response is scored by the same model acting as judge, using a 0/1/2 partial-credit rubric across seven quality dimensions.
The comparison
Identical prompts. Identical effort. Kohnex reasoning is the single variable between the two arms.
The limits
These are Kohnex-authored custom tasks and internal scores, not independently audited results. Improvements are specific to these evaluations.
Constrained design scored 21.0% with Kohnex in KB-100. That is an improvement, and still a low absolute score. Both deserve to be visible.
Why we publish the whole picture