We built KB-100 because public benchmarks stopped answering the question we actually care about: does this make a model reason better on expert work? We wrote 100 custom tasks across five expert categories, ran DeepSeek V4 Flash at maximum reasoning effort twice, once alone and once with Kohnex Brain, and graded every answer on a 0/1/2 partial-credit rubric with the same model as judge.
The headline number
Model only: 42.7%. Model + Kohnex: 65.2%. An improvement of 22.5 percentage points. Same model, same maximum reasoning effort. The only difference was Kohnex.
Where the gains came from
Cross-domain synthesis improved from 28.5% to 70.0%, a gain of 41.5 points. Connecting ideas across fields is where unstructured thinking collapses first.
Hostile debugging improved from 45.5% to 71.4%, a gain of 25.9 points on adversarial, misleading problem statements.
Constrained design improved from 4.0% to 21.0%. The absolute score remains low, and that matters as much as the 17-point gain.
Deceptive review improved from 58.5% to 73.0%. Multi-hop diagnosis improved from 77.0% to 90.5%.
How we kept it honest
Identical prompts, identical reasoning effort, one variable. Every response was scored on completeness, correctness, depth of analysis, actionable recommendations, edge-case coverage, production-readiness, and engineering judgment.
This is an internal evaluation on a custom task set. It is not a public benchmark and has not been independently audited. We publish the full category breakdown so you can inspect the evidence instead of trusting a headline.