Metacognitive AI systems that improve safely.
UCogNet (Universal Cognition Network) is a modular cognitive platform that routes tasks to the right solving mode, executes with verifiable evidence, and evolves via gated experiments under strict budgets. Built in Europe for the EU AI Act: every decision is gated, signed, and overridable by construction.
Route → Execute → Reward → Evolve
CLICK TO EXPANDWhat UCogNet solves
Three core problems in production AI systems.
Agents are hard to trust
UCogNet approachEvidence-first execution with claims, provenance and replay.
One-size-fits-all scaffolds degrade performance
UCogNet approachTask-aware routing selects minimal vs agentic modes.
Self-improvement can go off the rails
UCogNet approachGated evolution with A/B thresholds, cost caps and rollback.
Evidence across domains
UCogNet evaluated on plasma turbulence control (Hasegawa-Wakatani 2D, 8 controllers, 7D composite, 6 seeds) and BCI neural decoding (BNCI2014001, 9 subjects × 5 seeds, 360 cross-session + 45 LOSO evaluations, Wilcoxon paired tests with 95% CI). Same cognitive architecture — two scientific domains.
8
Plasma controllers405
BCI evaluations5
Validation seeds8
Models compared⚡
Plasma Turbulence Control
| # | Controller | Mean | 95% CI |
|---|---|---|---|
| 🥇 | UCogNet Enhanced | 0.7219 | ±0.015 |
| 🥈 | NeuOp-Transf. ‡2026 | 0.7285 | ±0.005 |
| 🥉 | UCogNet Legacy | 0.7311 | ±0.022 |
● UCogNet Enhanced: lowest multi-seed mean (0.7219 ± 0.015). No pairwise difference statistically significant (p>0.05, n=6). ‡ Neural operator baselines are surrogate approximations.
7D composite • 6 seeds • 95% CI (t-dist.) • surrogates disclosed🧠
BCI Neural Decoding
| Method | Accuracy | 95% CI |
|---|---|---|
| Riem-TS+LR | 76.0% | ±4.1% |
| Riem-MDM | 75.6% | ±4.1% |
| UCogNet-ResV2 | 74.2% | ±4.5% |
| CSP+LDA | 74.2% | ±4.5% |
| ShallowCNN | 73.7% | ±4.3% |
| UCogNet-Std | 71.8% | ±4.3% |
| CSP+SVM | 71.5% | ±4.7% |
| EEGNet | 71.1% | ±5.2% |
● UCogNet-ResV2 ranks 3rd of 8 (74.2%) — statistically tied with CSP+LDA (p=0.97). Significantly outperforms CSP+SVM (p=0.006) and EEGNet (p=0.09). LOSO: 64.4%.
Cross-session • 360 evaluations • Wilcoxon paired test • 95% CIKey findings
⚡
Plasma: UCogNet Legacy ranks 2nd of 8, beating both 2026 neural operator baselines. Enhanced variant wins 3/6 seeds.
🧠
BCI: UCogNet-ResV2 ranks 3rd of 8 models (74.2%) across 360 evaluations with Wilcoxon paired tests. Significantly beats CSP+SVM and EEGNet.
🔬
Cross-domain: single cognitive architecture operates both plasma turbulence control and neural decoding.
Where UCogNet applies
A general cognitive platform, validated in BCI decoding and parametric physics control.
Brain-computer interfaces
Competitive neural decoding on BNCI2014001 with representationally distinct features and robust subject coverage (9/9 threshold pass in 4-class).
Active researchPhysical control systems
Cognitive controller outperforms PID and LQR under out-of-distribution regime shifts in parametric physics simulations (Module 5, 5 OOD campaigns).
Active researchAutonomous agents
Task-aware routing, evidence-first execution, and gated self-improvement for production AI agents.
Core platformHigh-risk infrastructure
Cognitive architectures for high-stakes environments where failure modes cascade and classical controllers fall short.
Planned“When the world is on fire, you need a mind that dances with chaos.”
— UCogNet Research Center, by Brainstream • February 2026