Open preprints from the AICumene community — where signals meet minds and machines: neural interfaces, biosignal processing, drug discovery, cognitive enhancement, and efficient AI at the edge.

PAPERS

November 2026

From Brain Dynamics to Strong Artificial Intelligence: Learning Creates Functional Regime Switching

Alexandra Bernadotte, Ivan Menshikov

Current AI is evaluated externally, by benchmarks, outputs and reward; here the object of measurement is the internal regime organization created by learning. In a Continuous Thought Machine trained on maze solving, learning moves input-excited modes toward the unit circle, creates functional rhythms above the aperiodic background, and separates computation into a high-dimensional desynchronised working regime and a low-dimensional synchronised resting regime; beyond the training horizon the system alternates between work and rest, replays its solution, stores the answer in accumulated synchronisation, and restarts through a clock-like mode. Version 2 adds a first measurement on a toy JEPA world model that keeps latent dimension, oscillatory dimension and autonomous regime apart. Regime switching is proposed as a measurable structural marker of the transition from task-performing artificial systems toward strong artificial intelligence.

Signal processingCognitiveLLM systemsv2 · HTML · PDF · v1 · DOI (v1)
August 2026

Multichannel Singular Spectrum Analysis

Alexandra Bernadotte, Ivan Menshikov

A block-Hankel method for two linked inverse problems in multichannel recordings: how many oscillatory sources are active, and where they are. The rank of the block-Hankel matrix counts distinct oscillatory modes; the spatial loading of each mode is an amplitude fingerprint of the source position, and matching it to a known attenuation law recovers the coordinates. Identifiability is characterized geometrically — d + 2 sensors in general position determine a source in d dimensions — and a first-order accuracy formula is attained in simulations.

Signal processingNeural interfacesv1 · HTML · PDF · DOI
July 2026

Brain Oscillator Intrinsic Dimension marks consciousness, recovery mode, and failed downshifting in coma

Alexandra Bernadotte, Ivan Menshikov

BOID — the count of active oscillatory modes recovered from EEG by multichannel singular-spectrum analysis and ESPRIT — is high in conscious, resource-demanding states and low in a healthy recovery mode. Across sleep, anesthesia, meditation, and 575 post-anoxic coma patients, poor outcome is a failure to downshift out of pathological high-dimensional activity (fused model AUC 0.81, leakage-free).

Signal processingCognitiveNeural interfacesv1 · HTML · PDF · Supplement · DOI
July 2026

When Do Language Models Reach Nash? Heterogeneous Cross-Family LLM Agents in Classic Economic Games

Alexandra Bernadotte, Ivan Menshikov

Agents from three model families (Qwen 3.6, Gemma 4, Nemotron 3) on two servers play the p-beauty contest, the iterated Prisoner’s Dilemma and a Braess routing game. They reproduce the exact Nash equilibrium only when it reduces to picking the lowest of given payoffs; a contamination probe exposes the beauty-contest “success” as template recall, and cross-family play reveals strategic asymmetries that self-play hides.

LLM systemsCognitivev1 · HTML · PDF · DOI
June 2026

Verifiable AI for Law: a legal verifiability compiler with a rule/judgment firewall

Alexandra Bernadotte, Ivan Menshikov

Legal language-model agents may fabricate authorities, overstate discretionary conclusions and present incomplete analyses as final. Verifiable AI for Law is a bounded decision-support architecture built around an independent deterministic verifiability compiler: Proposer and Critic agents generate and challenge a typed Application-IR, the compiler checks citations, jurisdiction, source currency and formally decidable predicates against approved rules and validated facts, and a rule/judgment firewall escalates open-textured legal standards to qualified human reviewers. The model proposes; the compiler disposes.

LLM systemsv1 · HTML · PDF · DOI
April 2026

The April Theses of the New Cognitive Paradigm: From Imitation to Structural Measurability

Alexandra Bernadotte

The central problem of artificial intelligence is no longer whether a system can imitate human outputs, but whether cognition itself can be made structurally measurable. The New Cognitive Paradigm replaces behavioural and phenomenological accounts of intelligence, consciousness, language and truth with a measurement programme based on internal organization, and introduces three operational instruments: the epsilon-operator, the ersatz-operator and the operator of measurability.

CognitiveLLM systemsv1 · HTML · PDF · DOI
August 2025

What is Strong AI? Intrinsic Structural Criteria for Strong Intelligence Based on Dimensionality and Topology

Alexandra Bernadotte, Vasilii A. Gromov

A system of instrumental, internally defined criteria that classifies intelligence not by task performance but by intrinsic structural invariants of the data a system generates — intrinsic dimension, topology, and dynamical complexity. It yields a measurable hierarchy — Weak (Level 0), Strong (Level 1), Super-Strong (Level 2) — and a formal operational definition of Strong AI distinct from Strong Natural Intelligence and from AGI.

LLM systemsCognitivev1 · HTML · PDF · DOI

IN PREPARATION

June 2026

Bit-Exact Linear-Attention Decoding on Volta: Megakernel Design at the Load–Store Wall draft

Alexandra Bernadotte, Ivan Menshikov
Edge AILLM systems
June 2026

A Two-Tier Virtual-Screening Funnel on Legacy Datacenter GPUs draft

Alexandra Bernadotte, Ivan Menshikov
Drug discovery
forthcoming

Closed-Loop Motor Decoding from Non-Invasive EEG under a 40 ms Budget in review

Alexandra Bernadotte, AICumene Neural Interfaces group
Neural interfacesSignal processing

LAB NOTES

Short, citable write-ups — negative results, engineering lessons, reproductions. The things that would never clear arXiv moderation but move the field anyway. Each note is versioned and gets its own DOI, exactly like a paper.

July 2026

The heart adds nothing to coma prognosis: a pre-registered negative control

The same intrinsic-dimension and heart-rate-variability machinery applied to simultaneous ECG in the I-CARE cohort — marginal AUC over the EEG-plus-clinical model was indistinguishable from zero across every cardiac descriptor.

July 2026

What LUT dequantization actually costs you on Volta: an autopsy

We tried table-lookup dequant to beat the load–store wall on V100. It loses. Here is the profile that killed it, so you don’t have to rerun it.

June 2026

Gemma-on-Volta numerics: four fixes for coherent generation

Raw RMSNorm, unit scaling, weightless value-norm, proportional RoPE — the minimal patch set that turns garbage into text on sm_70.