Neural Simulator

Neural Simulator is an open research project attempting to grow an artificial brain from spiking neurons and synapses. The long-term goal is one integrated brain that starts small, develops through embodied and social experience, and eventually holds a genuinely self-directed conversation: reasoning to its own conclusions from its own perceptions, memories, needs, functional affect, and uncertainty, rather than retrieving stored text or predicting plausible-sounding text. The project's working hypothesis, sometimes called the emergentist bet, is that consciousness-relevant function emerges once a human brain's faculties and behavior are emulated completely and faithfully enough. The project does not claim to have reached that bar, or to have produced anything conscious, felt, or sentient; it is testing the hypothesis, not asserting its outcome (see Boundaries And License below).

This is not a finished mind or a chatbot wrapped in neuroscience terminology. It is a simulator, a collection of tested brain mechanisms, and a research program working to join those mechanisms into a continuously developing whole.

What The Project Is Building

The target system has five defining properties:

• Fully spiking: cognition is carried by simulated neurons, synapses, and their changing activity.
• One shared substrate: specialized regions are welcome, but perception, memory, value, emotion, thought, and language must interact within one brain.
• Grounded development: the brain learns by sensing a world, acting in it, experiencing consequences, and interacting with people. Text can be part of that experience, but isolated text prediction is not the goal.
• Causal inner state: needs, emotions, curiosity, confidence, and memory must change what the brain attends to, learns, says, and does. They are not labels added to an answer after it has been generated.
• Individual ownership: early brains should run locally and grow only as learning requires it. The engineering target is high-end consumer hardware, with sparse and local computation that could later suit neuromorphic hardware.

The intended developmental loop is:

world and people -> sensation -> internal state -> action or speech
                 -> consequence -> learning -> changed future behavior

Ordinary host code may provide the outside world, body interface, storage, visualization, and scientific instruments. It must not remain a substitute for a cognitive function that the simulated brain is meant to perform.

Read the project charter for the commitments behind this work and the technical overview for the architecture, current evidence, and major gaps.

Current State

The simulation foundation is substantial. It supports central processing unit and NVIDIA graphics processing unit execution, sparse spiking networks, multiple neuron models, declared brain regions and pathways, several forms of local plasticity, neuromodulation, homeostasis, structural change, checkpointing, experiment runners, and visualization.

The cognitive system is much earlier:

Area	Honest status
Shared brain	Multiple regions and pathways can occupy one simulation and update together. The recall composer now runs on the spiking one-brain store by default, with a hippocampus-style sharded lookup (2026-09-05) that keeps that same recall roughly 400x faster with no loss of recall or no-confabulation guarantee. As of 2026-09-05 four core thinking organs — surprise, the forward world-model, self-monitoring, and phrasing — share one literal neural pool by default (the "one-brain" merge), validated answer-preserving across 6 seeds — a genuine one-substrate step, though still co-residency (their organ routing is host Python), not itself a scaffold retirement. The single faculty whose host scaffold is fully retired is instead the recall composer's own mechanism: its host full-scan / NumPy recall path was demoted to an opt-out oracle (2026-09-02), the spiking store being the default. This is still co-residency of many other faculties sharing simulation steps, not one true substrate: most organ routing is still host Python. As of the 2026-09-05 ledger head the tracked production-integration ledger lists 66 faculties in total, 30 of them genuinely spiking and on by default, and 1 with its host scaffold fully retired (scaffold_retired: 1) — the ledger's own note states that most of the live chat's load-bearing cognition is still host, Python, or NumPy, not the shared spiking substrate.
Grounded action	Navigation, action selection, reward learning, perception, memory, and replay have working research results, usually in constrained tasks.
Grounded communication	A six-seed result joins learned visual association, hunger, a request-or-silence choice, and a world consequence. A newer six-seed result learns a tiny external vocal convention with two intents and two referents and succeeds on both untrained cross-combinations. It still uses injected motor exploration, fixed neural channels, and a host listener; it is preverbal learning, not natural language.
Conversational integration (current frontier)	Validated faculties are wired into the live chat one dependency at a time and kept only when a causal lesion changes the reply. Default-on and spiking: a claim-level no-confabulation moat that drops ungrounded content (an honest "I don't know" rather than a made-up answer); comprehension monitoring with targeted repair ("I didn't follow that", naming the unresolved role); a why / what-if causal forward model; scalar-implicature pragmatics ("some" reads as "some but not all"); a non-contradiction belief gate with in-place reconsolidation on a prediction error; self-initiated remarks on idle turns; open-ended generation that volunteers a novel, grounded, plausibility-gated, moat-checked proposition flagged as a guess; and a spiking global-workspace ignition bus that now authors the combine-and-decide step host Python used to do, abstains on a genuine multi-answer conflict, and runs re-entrant deliberation whose depth the substrate discovers itself. Honest inner-state read-outs (affect, a structural self-model, a graded certainty band) are functional statements, never felt states. As of 2026-08-19 three of the brain's own signals were made load-bearing on the conversation itself: its mood colors phrasing, its thought-swap decision steers which topic the turn engages, and its dopamine mode sets how engaged the reply is. A 2026-08-26 batch flipped ten more validated faculties to on-by-default, synced to main: empathy for another person's situation, an honest self-authorship marker on the brain's own guesses, attention as a competitive race between held discourse referents, a spiking vision-identity pipeline, a dopamine-trained value-driven choice between tied recalls, offline sleep-replay consolidation, the bounded-recall spiking mouth noted under Language, a neurally-mediated global-workspace "stop" on topic change, silent (no ongoing firing) working memory, and a basal-ganglia-style speak-or-stay-silent selector. None of the ten retires a host scaffold. Named host scaffolds remain, none is fully retired yet, and multi-fact continual learning is still open.
Language	The repository contains simple question answering and bounded spiking language experiments, plus a live conversational loop that composes honest grounded replies and honest silences. On an open-ended prompt the brain now volunteers a novel, grounded, plausibility-gated, moat-verified proposition flagged as a guess, and multi-sentence replies are planned by a neural discourse planner. As of 2026-09-04, the brain's own from-scratch spiking "mouth" is the default generation path for live chat, beating a simple word-pair baseline on simple text at a deployable size. It is honestly not yet fluent enough on broad, arbitrary-topic conversation to retire the external Qwen2.5-0.5B transformer acting as a declared articulation scaffold, which remains the tracked number-one blocker. Two candidate replacement mechanisms tried the week of 2026-09-05 (an erase-before-write "delta-rule" memory write, and a HiPPO/content-addressable-attention "hippokey" memory) were both banked, not shipped: hippokey was a no-go, and the delta-rule showed only a modest sub-bar lift on the decisive wikitext-103 test. Fluent, open-ended conversation grounded in ongoing life is not yet achieved.
Emotion and drives	Reward, value, neuromodulator signals, and curiosity mechanisms exist and several now act in the live chat. The brain reads its own live spiking mood and lets it color what it volunteers and how it phrases a reply; a six-seed result makes that mood caused by a simulated body-state read by dedicated interoceptive neurons, and cutting those synapses makes the feeling stop tracking the body. Affect is load-bearing on the surface (the difference vanishes when the embodiment pathway is lesioned), the self-selected dopamine mode sets how engaged the reply is, and a curiosity mechanism appends an honest follow-up question on an abstain about a novel topic. The feeling-to-word mapping is still a small host template, and a rich emotional system that develops broadly through experience remains open.
Memory and self-monitoring	Several memory and self-monitoring faculties now run on by default and spiking in the live chat: episodic recall of the current conversation (an honest "I recall / I don't recall discussing X" from a hippocampal pattern-completion); prospective memory that holds a deferred "remind me to X when Y" across turns and fires it on the cue; working memory holding two or more discourse referents on a bump attractor, plus an anaphora store and a discourse-event register ("who was doing it before?"); a metacognition confidence hedge whose confidence-to-correctness mapping is now learned by a local three-factor Hebbian rule (six-seed GO, type-2 AUC 0.825); an expectation-violation surprise notice; and a queryable forward world-model. The earlier episodic composition seam stays closed at readout (emergent assembly select, one-shot attractor, dendritic-plateau completion). As of 2026-09-02 the brain's full roughly 79,000-fact knowledge base, not an earlier 15,000-fact core, is the default it talks from. Some fact content and self-report wording remain host scaffolds.
Growth	Structural plasticity and capacity-growth infrastructure exist. A brain-native policy that safely grows a whole developing brain is not yet complete.

The central problem is integration. Passing a small test does not show that a mechanism can serve its role in a living brain, so the current frontier is a continuous-integration arc: validated faculties are wired into a live conversational loop one dependency at a time and kept only when the conversation itself measurably improves under causal controls and independent seeds. Running the real conversation, rather than an isolated probe, is what exposes a mechanism that was mis-scoped in isolation. A 2026-08-19 audit lesioned the thirty-one faculties then on by default and found twenty-three genuinely change the reply and none inert, with roughly fifteen genuinely spiking and lesion-load-bearing. As of the 2026-09-05 ledger head the tracked production-integration ledger counts 66 faculties in total, 30 of them genuinely spiking and on by default, and 1 with its host scaffold fully retired (scaffold_retired: 1), so this remains, by the ledger's own note, co-residency of many faculties in one loop rather than one true substrate for most of the live chat's load-bearing cognition, which still runs in host Python or NumPy. The open problems are a fluent open-ended mouth, a single true substrate for the remaining faculties, and mechanisms that emerge rather than being hand-wired.

For a dated status report, see Current State. For the planned build order, see the Roadmap.

Research Discipline

Temporary scaffolds are allowed to make a question testable, but every scaffold must remain visible, have a brain-native replacement, and have a condition for removal. Examples include host-written parsers, fixed concept codes, hand-set pathways, conventional language training, and external teaching systems. See the Scaffold Ledger.

Capability claims are expected to survive causal controls, independent random seeds when practical, and comparison with simpler explanations. Findings cite raw artifacts and provenance. Negative, corrected, and superseded results stay in the record because they constrain future work. The chronological evidence is under research/findings/.

Run The Project

Start with QUICKSTART.md for installation and backend setup.

Run a small central-processing-unit test:

SIM_BACKEND=numpy python -m pytest tests/test_strict_step_errors.py -q

Launch the interactive simulator on a configured NVIDIA system:

python neural-simulator.py

Most scientific experiments are headless modules in research/runners/. They are research instruments rather than polished product demos.

Repository Guide

Path	Purpose
sim/	Simulation engine, neuron and synapse state, learning, brain regions, neuromodulation, growth, and backends.
experiment/	Reusable stimulus, training, and readout support.
research/runners/	Executable experiments and diagnostic runs.
research/findings/	Dated interpretations of results, including failures and corrections.
research/findings/raw/	Raw measurements and run provenance.
tools/gates/	Automated checks for unsupported or stale research claims.
ui/, viz/, webapp/	Interactive control and visualization surfaces.
docs/	Current status, architecture, research standards, and historical plans.

Boundaries And License

The project studies functional mechanisms associated with cognition, affect, self-monitoring, and communication. Every self-report the brain gives is designed as an honest functional read-out of an internal spiking signal (for example, "my familiarity monitor reads this as novel"), never a claim of phenomenal experience; building and measuring those correlates is treated as a deliverable, not a caveat. It does not claim consciousness, sentience, felt emotion, human equivalence, or reliable general intelligence. Outputs from research demos are experimental and must not be treated as authoritative or as a safety-critical decision system.

The code is released under the MIT License and is provided without warranty. Contributions should preserve the distinction between measured behavior, interpretation, and speculation.