Priority Record
Public disclosure date: April 20, 2025
This site documents a multimodal cognitive architecture I first published as a 46-page patent draft on April 20, 2025. I chose to release that version — and all subsequent versions — freely to the public.
I did this because the implications of proprietary, corporate-controlled systems built on visual thought / internal 3D simulation as a core of cognition are too large to leave exclusively in private hands. The work is therefore placed in the public record so that the chronological priority and the original architectural claims remain visible.
The current final specification is the 471-page NMCA blueprint published in 2026. Earlier versions remain preserved as separate historical records.
Attribution of the 2026 NMCA
The 2026 NMCA specification is released under CC BY-NC-SA 4.0 with an additional attribution condition stated in the source document.
All derivatives must prominently cite:
“Neurosymbolic Multimodal Cognitive Architecture (NMCA) - by Derek Van Derven (2026).”
Cite the original disclosure:
Van Derven, D. (2025). NMCA Original Blueprint (46 pages).
https://doi.org/10.5281/zenodo.21972901
This requirement applies to derivatives of the 2026 work. Earlier publications may have different licensing terms; the license attached to the specific version should be consulted.
Publication Chronology
| Date | Publication | Record |
|---|---|---|
| April 20, 2025 | Original 46-page NMCA blueprint | Zenodo DOI 10.5281/zenodo.21972901 |
| 2026 | Expanded NMCA architecture | Zenodo DOI 10.5281/zenodo.20212241 |
| 2026 | NMCA Architectural Addendum | Zenodo DOI 10.5281/zenodo.22352561 |
A multimodal system centered on an Internal 3D Scene Builder as a core substrate of thought, with persistent visual scene memory, observation-driven updating, mental simulation, contradiction handling, and the explicit loop:
Input → Internal Simulation → Action → Reflective Update
What was disclosed on April 20, 2025
- Internal 3D Scene Builder used to simulate tasks visually and model interactions
- Persistent internal visual/3D scene representation for reasoning
- Memory encoding that stores visual scenes for future recall
- Internal visual simulation synchronized with observation
- Symbolic scene representations tied to memory and belief structures
- Contradiction detection and belief updating
- Planning inside the 3D spatial model
- Simulation of routes and actions before execution
- Feedback from real observation that triggers replanning
- Multiple simulation branches
- Explicit computational loop: Input → Internal Simulation → Action → Reflective Update
An agent is told to get a banana from the refrigerator. It loads the refrigerator location from prior visual memory, internally simulates the full route and object interaction, executes the plan, discovers the refrigerator is empty, and generates a new subgoal.
Original 46-page document (Zenodo):
https://zenodo.org/records/21972901
DOI: 10.5281/zenodo.21972901
Why these pages exist
Later technical papers have described closely related functional capabilities — persistent 3D scene memory, multi-hypothesis imagination of unobserved regions, sequential belief updating, and mental simulation for embodied planning — without citing the April 20, 2025 disclosure.
These pages exist so that the public record of what was published, and when, remains easy to find and hard to erase. They are documentation of chronological priority, not legal accusations.
Comparison with later work (selected)
| Capability | April 20, 2025 Disclosure | 3D-Belief (arXiv:2605.11367) Submitted May 12, 2026 |
|---|---|---|
| Persistent 3D / spatial representation | Internal 3D Scene Builder as core component of thought | Spatially consistent 3D scene memory using explicit 3D Gaussians |
| Memory of visual / spatial scenes | Memory encoding that stores visual scenes for future recall | Maintains coherent 3D memory of previously observed regions |
| Unobserved regions represented | Internal simulation of routes, interactions, and unobserved content; multiple branches | Multi-hypothesis belief sampling of unobserved 3D regions |
| Sequential updating from new observations | Observation-driven synchronization; contradiction triggers replanning | Sequential / online belief updating as new observations arrive |
| Mental simulation before action | Explicit simulation of routes and actions before execution | Mental simulation / imagination of 3D scene completions used for planning |
| Core loop | Input → Internal Simulation → Action → Reflective Update | Observation → belief update → re-plan using revised 3D belief |
Remember → construct/maintain internal scene → imagine → observe → update
3D-Belief contributes substantial engineering (diffusion models, 3D Gaussian Splatting, benchmarks, and robot experiments). The historical question is narrower: whether the high-level architectural combination already appeared in the April 20, 2025 disclosure.
Additional post-April-20 papers with related capabilities include:
• Learning 3D Persistent Embodied World Models (arXiv May 5, 2025)
• Video World Models with Long-term Spatial Memory (arXiv June 5, 2025)
• PERSIST – World Models with Persistent 3D State (arXiv March 2026)
3D-Belief: Detailed Publication Record
3D-Belief: Embodied Belief Inference via Generative 3D World Modeling was submitted to arXiv on May 12, 2026. The authors describe a generative 3D world model that maintains explicit 3D beliefs from partial observations, updates those beliefs as new observations arrive, maintains spatially consistent scene memory, generates multiple hypotheses for unobserved regions, and supports embodied reasoning and planning.
Authors: Yifan Yin, Zehao Wen, Suyu Ye, Jieneng Chen, Zehan Zheng, Nanru Dai, Haojun Shi, Aydan Huang, Zheyuan Zhang, Alan Yuille, Jianwen Xie, Ayush Tewari, Tianmin Shu.
The authors describe several capabilities that overlap with architectural functions documented in the April 20, 2025 NMCA disclosure, including persistent spatial scene memory, internal 3D scene representation, imagination of unobserved regions, sequential belief updating, contradiction-aware state maintenance, and embodied reasoning and planning.
3D-Belief also contributes substantial later engineering, including generative modeling, 3D Gaussian Splatting, benchmarks, and robot experiments. Those implementation details are not being attributed retroactively to the April 20, 2025 NMCA disclosure.
Additional post-April-20 papers with related capabilities include:
• Learning 3D Persistent Embodied World Models — arXiv May 5, 2025
• Video World Models with Long-term Spatial Memory — arXiv June 5, 2025
• Beyond Pixel Histories: World Models with Persistent 3D State (PERSIST) — arXiv March 3, 2026
Timeline (selected)
- April 20, 2025 — Original 46-page NMCA architecture publicly disclosed and subsequently preserved on Zenodo
- May 5, 2025 — Learning 3D Persistent Embodied World Models (arXiv:2505.05495)
- June 5, 2025 — Video World Models with Long-term Spatial Memory (arXiv:2506.05284)
- March 3, 2026 — Beyond Pixel Histories: World Models with Persistent 3D State (PERSIST) (arXiv:2603.03482)
- May 12, 2026 — 3D-Belief: Embodied Belief Inference via Generative 3D World Modeling (arXiv:2605.11367)
Statement of intent
I spent a year developing this architecture. I released the original version and all later expansions freely so that the ideas could not be locked behind proprietary control and so that the public record of their origin would remain clear.
These pages are part of that record.
The architectural concepts listed above were publicly disclosed by Derek Van Derven in the April 20, 2025 NMCA blueprint, before the May 12, 2026 submission of 3D-Belief and before the other post-April-20 papers listed here. This page records that chronological fact. It does not claim that later authors copied the disclosure.
Derek Van Derven
Original disclosure: April 20, 2025
Zenodo DOI: 10.5281/zenodo.21972901
Site: visualthoughtagi.com
