Key milestones in the development of UCogNet.
The same governed U-CogNet cognition that runs clinical triage and audits code now drives a Unitree Go2 in MuJoCo MJX. An Earth-trained locomotion policy transfers zero-shot to Mars and Moon gravity — the preregistered primary endpoint (steps-to-threshold) is supported 4/4 (Earth 1.31 → Mars 1.52 → Moon 1.13 m/s; 0 steps vs 21M+ or never from scratch). Every mission passes six ontological gates; a non-maleficence law (Law 4) lifts harm-blocking 0%→100% across a 12-mission battery with zero over-blocking, each veto ed25519-signed. Robotics for humanity only. Honest scope: n=2 seeds (pilot, effect sizes).
Inter-patient MIT-BIH AAMI arrhythmia detection (de Chazal DS1/DS2 split). Tier-3 adds per-class temperature calibration + coverage–accuracy abstention + a free-energy autoencoder + L2 class-aware drift guard on top of frozen Tier-2 weights. The F-rescue mechanism escalates high-free-energy-surprise beats: F-recall lifts 8× without ever showing the system a fusion-beat label.
The C₃-symmetric Thomas attractor applied to the Planck 2018 SMICA temperature map as a third null hypothesis, Level-A calibrated. 607 patches, 30,350 attractor evaluations; KS vs Uniform(0,1) D=0.034, p_KS=0.48 — survey p-values indistinguishable from uniform, no false-positive inflation; the Cold Spot lands at p≈0.5. First public-data run — we publish the negative.
Two-arm study running Meta SAM 3 on Mexican Robotics Federation robot-soccer footage. The U-CogNet arm wraps the same SAM 3 perception with a variational free-energy self-representation + information-geometry surprise reasoner. Identical detections, +6.4% wall-time, and seven autonomous model-break events the baseline cannot flag — perception that flags its own failures.
Veterinary microscopy audit across 11 parasite species with per-class precision/recall/F1 and an honest failure analysis, published with a machine-readable manifest and figure index.
New cognitive layer sitting above the Singularity: a Thomas cyclically-symmetric attractor (ẋ = sin(y) − bx + cyclic permutations) that closes a dissipative feedback loop with the output. ∇·F = −3b guarantees volume contraction; verified Lyapunov spectrum λ₁ = +0.24, λ₃ = −0.79, Kaplan–Yorke D_KY ≈ 2.31. Regulates chaos between layers without destroying cyclic information topology.
Full cross-session + LOSO + Euclidean Alignment + session-drift MMD (tangent + latent) + inter-model RSA on BNCI2014001. Quantum embeddings show no significant advantage (Riem-TS+Q: +0.15 pp, p=0.50; UCogNet-ResV2-Q: −2.9 pp, p=0.07, medium effect) — an honest negative result on current quantum feature mapping. UCogNet-Sing debuts at rank 4 (73.8 %).
Rigorous cross-subject evaluation framework added to the benchmark: Leave-One-Subject-Out with pooled training, Rodrigues et al. Euclidean Alignment, MMD session drift (tangent-space + z-space), Kaplan–Yorke fractal dimension, inter-model RSA across Spearman-ranked subject profiles. Reveals that UCogNet-Sing latent space has 2.1× higher session drift than Riemannian tangent space (MMD_z = 0.76 vs MMD_ts = 0.36) — a measurable weakness that guides future encoder design.
Full multi-objective evaluation: 9 subjects × 5 seeds × 8 models = 360 cross-session + 45 LOSO. Riem-TS+LR tops at 76.0%; UCogNet-ResV2 ranks 3rd (74.2%, p=0.97 vs CSP+LDA). Wilcoxon paired tests, 95% CI, Cliff's delta effect sizes.
Full 8-controller benchmark on Hasegawa-Wakatani 2D (128×128) with 7D composite metric, 6-seed validation, and paired t-test. UCogNet Legacy ranks 2nd overall; Enhanced wins 3/6 seeds with best multi-seed mean (0.7219). Compared against 2026 SOTA: NeuOp-Transformer, FI-Conv, FNO Surrogate.
New /researches section with tabbed Plasma + BCI views. Full experimental protocol, simulation parameters, metric breakdowns, and peer-reviewed references for both domains.
Continuous learning loop: harvest → should_train → LoRA fine-tune → deploy → hot-swap model. Autonomous background learning with rollback on Ψ_stable degradation. 16/16 tests pass.
Initial cross-session evaluation. Superseded by Apr 2026 rigorous run (8 models, 5 seeds, Wilcoxon tests).
Cognitive controller vs PID/LQR/Zero/Random with 3 seeds. Strongest differentiation under regime shifts (Campaign C: 0.351 vs PID 0.090).
Gated mutations with gradual rollout and rollback.
807 tests active across 10 modules. Frozen-run reproducibility verified.
Task-aware dispatcher with budget-conscious mode routing.