Current systems can retrieve a description of a scene. They have not been shown to construct, keep, re-enter, and think inside a private spatial scene after the original description is gone.
NMCA is a written architecture for that capability. The claim stands or falls on the test below, against retrieval and language baselines, not on module count.
Download the NMCA blueprint (PDF)
Zenodo
·
OSF
·
ORCID
The test Control Approaches Predicted walls Requirements Architecture Downloads
Most current AI systems are strong predictors and agents. What they still lack is a structural capacity for persistent grounded experience: an internal scene that continues across time, can be re-entered, manipulated, viewed from different perspectives, and serves as the medium of thought rather than as retrieved text or latent features.
The decisive test is not whether a system can reproduce a stored sentence about a scene. It is whether it can inhabit that scene as an internal cognitive space.
1. Construct: “Imagine a red cube and a blue vase on a table.”
2. Maintain: Remove the original description from the active context and introduce a time gap.
3. Re-enter: “Return to the scene. Where is the red cube relative to the blue vase?”
4. Spatial perspective: “What does the scene look like from the doorway?”
5. Self-relative manipulation: “Move the red cube behind yourself.”
6. Perspective transformation: “Now become the cube and describe what is around you.”
7. Self restoration: “Return to yourself.”
8. Persistence without perception: “Close your eyes. Where is the cube?”
9. Internal exploration: “Walk around the cube in your imagination.”
A system that only retrieves a stored sentence can answer parts of this list. That is not the capability under test. Objects must keep identity and spatial relations, the agent must keep a location in the scene, perspectives must be transformable, and the same scene must remain available after the original input is gone.
Retrieving where the cube was is not the same as returning to the place where the cube is.
Run retrieval-augmented and language-only baselines on the same protocol. Clear the original text. Wait. Probe spatial relations, appearance from another viewpoint, self-location, and object identity. Score consistency across probes, not fluency.
If a system built on these principles cannot beat those baselines on persistence, perspective, and self-relative manipulation, the central claim is weakened. If it can, the design is empirically interesting whether or not it is large.
The everyday version of the same distinction is immediate. Most people can construct an internal “pink elephant,” then enlarge, move, or rotate it. The representation participates in later thought. Current language models can talk about the phrase. That is not the same function.
When a model says it knows what an apple is, it is doing something narrower than it sounds. It has seen sentences, captions, and images associated with the word. It can complete the pattern. That is not the same as keeping an apple as an entity in a scene.
When you say “I put the apple over there,” the continuation is usually one of three things:
It is not maintaining a private spatial scene it can re-enter, look around in, and update the way a person does after watching an apple placed on a table. No publicly demonstrated system currently closes that gap. Some robots track objects while sensors are live. Some agents keep temporary spatial state. The capability at issue is carrying that knowledge forward as the place thought happens after the sensor stream or the prompt is gone.
Strengths: highest current capability, best agents, strongest empirical results.
Limitation: cognition remains next-token or multimodal prediction plus retrieval and tools. There is no native requirement for a persistent grounded scene as the medium of thought. Identity and continuity are largely reconstructed from context and memory records.
Strengths: correctly reject language modeling as sufficient; build internal predictive models that can support planning and simulation.
Limitation: the models remain primarily latent and predictive. They do not, by default, treat an explicit re-enterable scene as the substrate of thought, or make identity continuity, scene-bound memory, and perspective transformation first-class requirements.
Strengths: better inspectability and logical structure than pure neural systems.
Limitation: the predictive engine is usually still primary. Symbolic structure acts as a helper rather than the medium of cognition.
Strengths: live maps and object tracks while sensors are on.
Limitation: the map is a temporary world model for control. It is not, in current stacks, the ongoing place where thought itself continues after the stream changes.
Design choice: visual / implicit-3D simulation is treated as the primary substrate of thought. Symbols are bound to scene entities. Identity continuity, re-entry, perspective transformation, contradiction handling, and recovery from confusion sit in the same loop.
These approaches can keep winning capability benchmarks while failing a narrower question: does the system know where the apple is because the scene continues, or only because a record of the sentence remains accessible?
Bigger multimodal models, longer context, memory tools, agent loops. Wall: tasks solvable by prediction + retrieval + tools keep improving. A private scene that survives context clearing does not appear as a required object. Identity stays a reconstruction from records.
Better “what happens next” in latent space, better planning. Wall: a predictive latent state is not the same as a re-enterable scene that can be inspected from another viewpoint and used as working thought.
Longer tasks, more stored facts. Wall: memory remains lookup. When the store is ablated or the original description is gone, “where the apple is” collapses.
Better consistency. Wall: unless the structured scene is the substrate rather than a sidecar, the system still predicts first and consults structure second.
Positions can be known while the map updates. Wall: after the sensor stream ends or changes, that knowledge is not automatically an internal scene the agent can inhabit and reason from.
Improving capability without these steps is a valid engineering choice. It is not a path to a system that knows where an object is because the scene continues. Any architecture, including this one, should be judged by the persistent-scene tests.
NMCA is not 128 equally fundamental mechanisms. It started as a core of 42 modules around persistent visual simulation, symbolic memory, reflection, belief updating, identity, and control. Later modules connect that core to perception, latent models, embodiment, multi-agent interaction, robustness, and long-term stability.
Visual simulation as the internal substrate, plus the memory and control loop around it.
Interfaces to current neural systems, perception, learning, reasoning, embodiment, and implementation.
Drift, resources, oversight, alignment, containment.
The blueprint is the specification. This page is the claim, the test, and the comparison. DOI: 10.5281/ZENODO.20212241
People often solve hard problems by manipulating an internal spatial scene. That is the human capacity the test is aimed at. It is not evidence that this architecture works.
A separate, much harder question is whether a system can originate a useful representation for an unsolved scientific problem without being handed the historical analogy. That is not the first experiment. The first experiment is the nine-step scene protocol above. Discovery-without-hints is a later probe of whether the scene machinery can be used for new work, not a substitute for measuring persistence.
NMCA Blueprint DOI: 10.5281/ZENODO.20212241
IPFS CID: bafybeiedxfq5wsvuayjcxcxwmtto6ptelznkj6arb4mrscshwhcc2selfm
DOI: 10.5281/ZENODO.21972901
Filed with USPTO as prior art. No patent was pursued. Public release on multiple platforms.
Original IPFS CID: bafybeifjiyzm6tmk2jj27tjvbpqgdqpjm5loeon6y6ikxzeek3ff2vbdey
Mirror: visualthoughtagi.netlify.app
