Built at the intersection of AI, neuroscience, and complex science — an infrastructure for cognition that develops over years, and a platform for the disciplines that study how.
The infrastructure designs the brain. The virtual person validates the design. The virtual world gives the person a place to act, perceive, and be observed. Each pillar drives the next.
The platform designs the brain; the virtual person validates the design; the world is the repeatable test bench where both are observed. Same seed, same initial state, change one variable — neither real animal experiments nor the real world can do this. Here it is routine.
Building all of this required a platform that did not yet exist — where designers from neuroscience, complexity science, and AI can compose, debug, and experiment with brains as fluently as engineers compose modular code.
Most of the work to date has been on this infrastructure. The rest of this site mainly describes it. Through this platform, scientists can probe the underlying principles of three disciplines at once — making Evolvra a research instrument as much as a product.
A response to the limits of today's dominant AI paradigm — LLM-centered artificial neural networks.
The infrastructure does not yield a single product — it yields a class of intelligent systems that do not exist today.
Neurons are organised into overlapping zones. Activity emerges from local rules — there is no central conductor. Toggle the layers, hover a neuron to inspect it, click a zone in the legend to isolate it.
Every spike, every connection, every frame is recorded. The same compiled brain can be replayed from any saved frame with an arbitrary perturbation — silence a neuron, modify a weight, alter a neuromodulator level. Causal experiments that are infeasible in vivo become routine here.
Evolvra runs as a continuous frame-loop physics simulation. Each frame is one tick of brain-wide simulation, ~0.2 ms long; a single interaction spans thousands to tens of thousands of frames. The simulation advances in discrete ticks, but there is no central cognitive clock and no single processing wave moving through the system.
A signal is the source neuron's identifier — nothing else. Intensity is not on the signal; it emerges through frequency coding (a strongly activated neuron fires across more consecutive frames). Semantic content is read by downstream neurons directly from the source's state when needed.
Perception, emotion, memory, and reasoning update each other in loops until firing-rate statistics stabilize. There is no forward pass and no fixed processing order.
When stability is detected, the system emits a response and the cycle ends. This is pragmatic. The intended direction is a frame loop that runs continuously: what we now call convergence becomes the moment a distributed firing pattern across many zones stabilizes — closer to how neural populations distributively encode information during decision-making than to a switch flipping in one region. Formalizing that decision-point without breaking continuity is an open problem.
All synchronous physics — propagation, decay, dendritic integration, composition, threshold, post-fire — fits in a single sub-millisecond tick. Slow operations are submitted asynchronously and rejoin the substrate later as sustained plateau voltage.
Most ML architectures treat a neuron as a single weighted sum followed by a nonlinearity — a passive integrator. Real cortical neurons compute nonlinearly inside their dendrites: each branch is its own coincidence detector, and the soma sees a composed result.
Each Evolvra "neuron" maps to a cortical column-scale functional unit rather than a single cell. But we keep the dendritic arbor as a structural idea, because the computational property it gives us — branch-local coincidence detection plus role-typed composition — is what makes cortex more than a feed-forward sum.
Direct trees bring evidence. Modulatory trees bring context that amplifies or weakens it. Gating trees are the veto — inhibition is multiplicative, so even strong evidence can be silenced by a well-placed gate.
A segment fires only when both conditions hold inside its coincidence window: enough distinct upstream sources, and enough total weight.
activate()
When the soma crosses threshold, the dendrite hands activate()
a causal snapshot — the segment-level coincidences
that caused this neuron's recent firing. Downstream computation can
reason about why.
All neurons share this same underlying structure: dendritic arbor for input
integration, soma with voltage and threshold, axon with outgoing connections,
the same gate cycle every frame. What distinguishes the three categories is
whether the neuron carries an additional
activate()
module — a per-neuron computation that runs only when the dendritic gate
releases the neuron for firing — and what that module does. The long-term
trajectory is to replace Category B and C modules with trained clusters of
Category A units, leaving the signal protocol and the physics layer unchanged.
No activate() module.
The dendritic arbor is the whole computation. Direct evidence, top-down
modulation, and gating inhibition are composed before the soma decides
whether to fire. Most neurons live here because this path is cheap,
local, and inspectable.
The brain has no scheduler that says "run STDP now." What it has instead is a set of substances, each with its own production rule and decay rate. Learning timescales emerge from substance dynamics. Every frame, the physics runs; learning happens because the physics runs.
A concept in Evolvra is not a node and not a record — it is a pattern of co-firing across many neurons. Competition selects which neurons win; sparsity makes the winners distinct; co-firing makes the pattern itself the memory.
A concept is not a single node. It is a distributed pattern — the same input produces similar but not identical activation, and that pattern itself is what means something.
When many candidate representations activate at once, lateral inhibition forces them to compete. The role is not "to look more brain-like" — it is to select. Competition is the mechanism that turns a wash of weak simultaneous activations into a small set of decisive winners.
What competition produces is sparsity. The system pays a lower energy cost, suffers less crosstalk between concepts, tolerates more noise, and leaves clean boundaries between assemblies. Similar inputs produce similar but not identical patterns — the basis of generalization without collapse.
A group of neurons that fires together represents a single piece of information. If they fire together repeatedly, STDP and other plasticity mechanisms strengthen their mutual connections. Over time they become a cell assembly. The same physical event that means something also causes learning — representation and learning are not separate operations on separate schedules.
A concept lives in the pattern of which neurons fire and how — not in any one of them. The same concept survives the loss of individual neurons; partial activation can reconstruct the rest by spreading through strengthened connections. Recall becomes pattern completion, not retrieval from a store.
How do you design something as large and complex as a brain without losing yourself in the wiring? Five composable object types, deterministic compilation, a zone topology that mirrors how cortical regions actually overlap — and the freedom to hand-edit anywhere in the generated result.
threshold: 0.80
direct + apical + gating
~100 neurons
seeded generation
regions + links
Zones in Evolvra are not a fixed enum. A zone can nest ("V1-like" inside "early-visual" inside "PERCEPTION"), and the same neuron can belong to multiple zones simultaneously — a neuron in "V1-like" might also belong to a cross-cutting "fast-pathway" zone that overlaps with "MOTOR". This is closer to how cortical regions actually overlap than to a clean hierarchy.
This is also the same principle that complexity science calls modularity and hierarchy — overlapping, multi-scale modules that exchange information without collapsing into a single partition. Brains have it. Cities have it. Designers should be able to express it.
Evolvra trades cellular and molecular realism for total observability at the system level. Every spike, every segment, every weight, every frame — exact. The kind of access in vivo recording can only approximate, applied to the questions complexity science and neuroscience most want to ask.
Frame-by-frame replay. Any saved frame can be replayed with a perturbation: silence a neuron, modify a weight, raise a threshold, alter a neuromodulator. Causal experiments not possible in vivo become routine.
The substrate couples positive feedback with multiple independent brakes — engineered to push the network toward the edge between "too rigid to learn" and "too chaotic to remember." Whether it lands there is empirical, and observable.
Cascades of consecutive firings — their size and duration distributions are the standard signature of criticality. A clean power law indicates the regime; a deviation indicates departure.
The network rests in semi-stable activity patterns, then transitions. Each pattern is a candidate attractor; the dwell-time distribution and transition graph are directly extractable from the frame stream.
Phase coherence across zones, information flow between regions, band-specific oscillations — measurable on the same exact firing data that in vivo recording can only sample.
Four global dimensions — dopamine, norepinephrine, serotonin, acetylcholine — gate whether STDP produces LTP or LTD, and at what rate. Their distribution, release rules, and effect on plasticity are inspectable per frame, per region.
Export at every granularity — neuron, segment, zone, frame — in formats ready for offline scientific analysis. Which analytics matter most is one of our open questions.
Current status — May 2026. The upper-structure goals — full virtual person, eventually the voxel world — are real plans but depend on the substrate being further along.
The architectural choices that enable scaling — sparse-matrix connection graph, vectorised numeric state, one-integer signal protocol — were made from the start. The path past the hundred-thousand-neuron threshold does not require changing the routing API.
The current demo runs small by design — bounded by compute, not architecture. The choices that make scaling possible were made from day one. The choices that make embodiment real are the next horizon.
Sparse-matrix connection graph from Phase 1. Vectorised numeric state in numpy arrays. One-integer signal protocol — routing expressed as sparse matrix multiplication. The path past the hundred-thousand-neuron threshold does not require changing the routing API.
Text alone is too narrow a substrate for any of this to be more than gestural. The real academic value of the virtual world is not that it looks pretty — it is that it transforms what the brain can be studied as.
A perception–action loop replaces the text I/O channel. The brain sees, moves, and is changed by what it does.
Concepts and episodes attach to places, objects, and routines. Recall becomes situated, not symbolic.
Predictions are tested by changing the world and observing the result — not by training on a logged dataset.
Same seed, same initial state, same brain — change one variable, run again. Real animal experiments cannot. Real worlds cannot. Here, it is routine.
The virtual world is not a frontend; it will become part of the system. Today the brain is reached only through text. Eventually she will live in a multimodal world — walking through it while we watch the brain change.
Evolvra sits at the intersection of three disciplines. The questions we cannot answer alone are the ones that come from where those disciplines meet.
Evolvra is designed around STDP, short-term plasticity, heterosynaptic competition, intrinsic excitability, synapse birth and death, conduction-speed plasticity, neuromodulation, and metaplasticity — together with several stabilizing "brake" mechanisms. Some of these are biologically rich, and the right abstraction level for an artificial substrate is not yet clear.
We need to decide which mechanisms are essential, which can be simplified, which should be delayed, and which parameters should be exposed for experimentation. A possible direction is a mechanism-control panel where specific mechanisms can be enabled, disabled, or tuned, so researchers can observe how each one changes the system's dynamics.
Evolvra can expose every neuron, dendritic segment, connection, weight, voltage, threshold, firing event, and frame. But total observability does not automatically mean useful data. We need to decide what should be logged, summarized, visualized, and analyzed at higher levels.
Candidates include criticality signatures, avalanche size and duration distributions, metastable state transitions, phase coherence across zones, information flow between regions, stability of distributed representations, and before/after perturbation comparisons. We would value guidance on which measurements are meaningful for neuroscience and complexity-science research — and which would only be visually interesting but scientifically weak.
Because Evolvra is an artificial abstraction, there is a risk that its activity may appear brain-like without capturing the dynamics that matter. We need criteria for validation: what would count as evidence that the system is approaching useful regimes of distributed coding, plasticity, criticality, or memory consolidation?
Conversely, what observations would indicate that the abstraction is wrong, too shallow, or missing essential mechanisms? We would value help defining falsifiable tests, comparison baselines, and failure conditions — not only success metrics.
If any of these resonate with your work, we would like to hear from you.
Get in touch →