DeepSeek V4 Flash + Kohnex: +22.5 Points on FORGE-100

We built FORGE-100 because public benchmarks stopped answering the question we actually care about: does this make a model reason better on expert work? So we wrote 100 custom tasks across five expert categories, ran DeepSeek V4 Flash at maximum reasoning effort twice — once alone, once with Kohnex Brain injected — and graded every answer on a 0/1/2 partial-credit rubric with the same model as judge.

The headline number

Where the gains came from

• Cross-domain synthesis: +41.5 (28.5% → 70.0%) — connecting ideas across fields is where unstructured thinking collapses first.
• Hostile debugging: +25.9 (45.5% → 71.4%) — adversarial, misleading problem statements.
• Constrained design: +17.0 (4.0% → 21.0%) — tiny base, but a 5x lift on the hardest category.
• Deceptive review: +14.5 (58.5% → 73.0%) — spotting what looks right but isn't.
• Multi-hop diagnosis: +13.5 (77.0% → 90.5%) — already strong, pushed near ceiling.

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 — not a public benchmark, not independently audited — and we publish the full breakdown on our benchmarks page so you can inspect every category instead of trusting a headline.

That last part is the point. Proof over promises.