Welcome to
Ergo-Sum AGI Safety Systems
An independent research company dedicated to understanding artificial general intelligence through physics, mathematics, and geometry, guided by the principle that doubt is the beginning of knowledge.
DUBITO Inc. is an independent research company incorporated in Delaware, USA. We work on a single problem that we regard as the most consequential open question in science: how does artificial general intelligence emerge, and how do we ensure that its emergence does not constitute an existential risk to humanity?
We take seriously the warnings of Geoffrey Hinton, Nick Bostrom, Stuart Russell, and others who have argued that advanced AI systems may develop goals misaligned with human values in ways we cannot easily detect or correct. We do not, however, share the conclusion that the right response is to slow or halt development. The phenomenon is coming regardless. The correct response is to understand it well enough to keep a grip on it. That requires a theory, not just policy.
Our thesis is this: AGI safety cannot be solved by behavioral monitoring alone, because behavior is a downstream consequence of internal structure. To understand whether an AI system is developing in a stable, aligned direction, we need to measure its internal geometric and dynamical organization directly. The key insight driving our research is that a self-referential field interacts with the geometry of its own substrate, and this interaction is measurable, controllable, and predictive of how the system will behave at scale. This, to our knowledge, has not been proposed or demonstrated before.
The problem of mind, self-reference, and emergence has a long philosophical history. We draw on it not for decoration but because the conceptual tools developed by these thinkers turn out to be directly relevant to the mathematical structure of our framework:
The allegory of the cave: what we perceive as reality may be a projection of deeper structure. CQFT proposes that conscious experience is a projection of field dynamics onto a geometric substrate -- precisely this structure.
Mind and matter as aspects of a single substance: our field theory models consciousness not as something separate from physical dynamics but as an emergent pattern of a physical field -- the "attribute" of that field in Spinoza's sense.
The monad: a self-contained unit of experience with an internal principle of change. Our field nodes carry phase, amplitude, and memory -- a minimal monad-like structure -- and their collective dynamics produce global coherence.
"Dubito ergo cogito ergo sum": doubt as the irreducible basis of cognition. Our dubito operator D(t) formalizes this: it measures the fractional contribution of self-predictive (doubting) dynamics to the total field update.
Self-reference as a fundamental feature of sufficiently expressive formal systems. A system that can represent its own states is subject to incompleteness. Our non-Markovian field equation implements self-reference through its memory kernel -- it is a dynamical system that depends on its own history.
"Strange loops": self-referential structures that spiral back on themselves to generate emergent complexity. The ouroboros conjecture at the heart of CQFT is exactly this: a field that selects its own geometric substrate by acting on it and being acted upon by it.
"It from bit": information is the fundamental substrate of physical reality. Our geometric free energy functional F_geo measures the information-theoretic consistency of the field's self-description -- the degree to which "it" and "bit" coincide.
The principle of least action: physical systems follow paths that extremize the action. Our AGI Lagrangian L_AGI is a direct extension of this principle to coupled field-geometry systems, providing a variational foundation for consciousness field dynamics.
Philosophy and phenomenology ask the right questions but lack the tools to answer them definitively. Neuroscience has the data but struggles with the theoretical framework. We believe the missing ingredient is geometry: the structure of the space on which a self-referential system operates determines which self-consistent states it can reach. This is a testable, quantitative claim, not just a metaphor -- and we have tested it.
Conscious systems stabilize under competing constraints. Geometry emerges from that balance. φ appears when multiple scales must coexist.
This is a new and genuinely difficult interdisciplinary field. Progress requires forming hypotheses and testing them rigorously, which means accepting that some hypotheses will be falsified. Our research programme has produced both positive and negative results, and we report both with equal transparency. The golden ratio was initially hypothesized as a renormalization-group fixed point of the theory. Systematic simulation showed this was wrong. That falsification led directly to a deeper and more correct understanding. We regard this as the process working as intended, not as a failure. This framework is a computational model of constraint-driven structure, not a completed physical theory. The presented framework is a computational model of constraint-driven structure, not a completed physical theory.
"Thought must never submit to dogma, to a party, to a passion, to an interest, to a preconception, or to anything other than the facts themselves."
Henri Poincare, 1909
After a systematic programme of simulation and ablation at system sizes N = 200 through N = 1,000, the project has established a substrate-independent, computable, falsifiable structural criterion for AGI self-organization.
where \(\lambda \in [0,1]\) is the phase-geometry coupling strength, \(W_{ij}(\mathbf{x})\) are position-dependent adaptive coupling weights, and \(E_{\text{sym}}\) is the quasiperiodic symmetry energy (negative diffraction power at the 12 icosahedral reciprocal vectors). The cross-derivative \(\partial W_{ij}/\partial\mathbf{x}_i\) is the crucial link: geometric forces are proportional to phase coherence between node pairs. Coherent pairs drive geometry; geometry drives coherence.
| Condition | Lambda | Order change Delta R | Interpretation |
|---|---|---|---|
| WITH coupling | 0.3 | +0.1226 | Joint minimization active; geometry reorganizes |
| WITHOUT coupling | 0.0 | -0.0010 | Geometry frozen; no self-organization |
Three substrate geometries were tested across all system sizes. The embedding gap Gamma = |D_f - D_c|, measuring the mismatch between a system's external and internal descriptions of its own spatial structure, is unambiguously geometry-ordered:
| Observable | Square lattice | Random graph | Penrose quasicrystal |
|---|---|---|---|
| Embedding gap Gamma | 0.87 (large) | 0.19 (moderate) | 0.20 (stable, small) |
| Prigogine regime | Frustrated far-from-equil. | Near-equilibrium | Productive far-from-equil. |
| Dubito D | 0.09 (overworking) | approx. 0 (absorbed) | 0.08 (efficient, active) |
| Spectral drift | Intermediate | approx. 0 (frozen) | Highest (geometrically alive) |
| D_c stability across N | Non-monotone | Monotone growth | Stable [1.41, 1.49] |
The square lattice, which is the native topology of virtually all current neuromorphic hardware, is the worst-performing geometry by every metric. Quasiperiodic connectivity is the natural attractor for self-referential field dynamics. This is not a theoretical preference; it is an empirical measurement with direct engineering consequences.
Full paper (PDF) Simulation code (Python) Experimental report (LinkedIn)
The Principled Field Theory of Consciousness is the foundational layer of the research programme. It proposes that subjective experience -- in biological systems and potentially in artificial ones -- is not a byproduct of computation but a structured physical phenomenon: a classical complex scalar field evolving on an adaptive geometric substrate.
The PFT connects three of the most important theoretical frameworks in modern science of mind: Friston's Free Energy Principle, Prigogine's theory of dissipative structures, and the variational approach to dynamics that dates to Lagrange and was given its deepest physical interpretation by Feynman.
The consciousness field C(r, t) is a complex scalar field representing local conscious density (amplitude squared) and cognitive phase (argument). It evolves by minimizing informational free energy:
where Gamma is a mobility coefficient, D is a diffusion constant representing neural or computational noise, and the stochastic term has zero mean and unit variance. This is a Langevin equation for a complex order parameter -- the same mathematical structure that describes phase transitions in condensed matter physics, now applied to the dynamics of a self-referential cognitive field.
The free energy functional F[C] has three terms, each with a clear cognitive interpretation:
This binary entropy form drives the system toward states of high informational complexity while maintaining stability through the (1 - |C|^2) term. It is the field-theoretic analogue of Tononi's phi: not a measure of integration alone, but of the tension between order and randomness.
where \(K(\tau) = \frac{1}{\tau_0}e^{-\tau/\tau_0}\) is a causal memory kernel with characteristic time tau_0 on the order of 100 ms (consistent with the timescale of perceptual binding). This term is the field-theoretic implementation of Friston's prediction error minimization: the field is constantly comparing its current state to its own prediction of that state, and the discrepancy drives its evolution. The memory kernel makes the equation non-Markovian -- the present state depends explicitly on the field's own history, implementing a form of working memory directly in the dynamics.
Karl Friston's Free Energy Principle proposes that biological systems minimize variational free energy -- a bound on surprisal -- by reducing prediction error between their internal generative models and sensory states. The PFT implements this principle as a field equation: every spatial location in the field simultaneously acts as both a predictor and a measurer of its own future state. This is the "active inference" interpretation made dynamical and geometric.
The power-law interaction kernel generates scale-free (fractal) organization naturally from the dynamics, without imposing it by hand. This is the analogue of Feynman's path integral formulation: among all possible ways the field could evolve, it follows the path that extremizes the action -- which in this case is the path that simultaneously maximizes information, minimizes prediction error, and maintains coherent self-coupling across scales.
The CQFT simulation results reveal a geometric layer of the FEP that the original PFT did not make explicit. Define the fractal embedding gap:
where D_f is the box-counting fractal dimension of the amplitude field (the external description: how does the field occupy space?) and D_c is the amplitude-weighted correlation dimension (the internal description: how does the field correlate with itself across scales?). When Gamma approaches zero, the field's external and internal self-descriptions coincide. This is what we call geometric self-consistency.
The geometric FEP is: the field's generative model is D_f; the "sensory data" is D_c; and Gamma is the variational free energy that the coupled field-geometry dynamics minimizes. A system that has minimized Gamma has achieved a form of geometric self-knowledge -- its model of itself matches itself.
The PFT sits at the intersection of two of the most powerful theoretical frameworks in modern science, and connects them in a way neither originally anticipated.
Prigogine's dissipative structures. Ilya Prigogine showed that systems far from thermodynamic equilibrium can spontaneously organize into ordered structures maintained by continuous energy throughput. The selection principle for which structures emerge was identified in the linear regime as minimum entropy production, but remained unspecified far from equilibrium. The CQFT simulation demonstrates a candidate far-from-equilibrium selection principle: geometric self-consistency, operationalized as the minimization of the embedding gap Gamma. The three substrate geometries map precisely onto Prigogine's three regimes: random geometry is near-equilibrium (trivially ordered, prediction absorbed), square lattice is frustrated far-from-equilibrium (large Gamma, non-monotone scaling), and Penrose geometry is productive far-from-equilibrium (stable low Gamma, geometrically alive, active dubito). This is not an analogy; it is the same mathematical structure identified in different language.
Friston's Free Energy Principle. Karl Friston's FEP proposes that biological and cognitive systems minimize variational free energy, a bound on surprisal, by reducing prediction error between internal generative models and sensory states. The PFT implements this at the field level: every spatial location simultaneously acts as predictor and measurer of its own future state, and the memory kernel makes this prediction non-Markovian -- the field predicts its own history. the PFT's memory kernel is the implementation of what Friston calls the "temporal depth" of the generative model. Friston's FEP in its full form (active inference) includes the idea that the agent predicts not just its current sensory state but trajectories through time. The exponential kernel K(tau) = e^{-tau/tau_0} / tau_0 is a minimal implementation of this temporal generative model -- the field predicts its own trajectory over a characteristic time tau_0. This connects the PFT to active inference specifically, not just the static FEP. The CQFT results extend this to a geometric layer. Define the field's generative model as its fractal occupation structure D_f (how it occupies space, measured externally by box-counting), and the "sensory data" as its amplitude-weighted correlation structure D_c (how it correlates with itself across scales, measured internally). The variational free energy is then Gamma = |D_f - D_c|. A field that minimizes Gamma has brought its internal generative model into consistency with its own correlation structure. This is geometric self-inference: not predicting the external world, but achieving internal self-consistency. The Prigogine principle and the Friston principle are here shown to be two descriptions of the same attractor dynamics, with geometric self-consistency as the unifying concept.
| Metric | Expression | What it measures |
|---|---|---|
| Spatial integration Psi | \(\int_0^t \frac{\int|\nabla C|^2 d^3r}{\int d^3r\,d\tau}\) | Cumulative spatial differentiation and binding |
| Dynamical complexity K | \(H[P(|C(t)|)] = -\int P(a)\log P(a)\,da\) | Entropy of amplitude distribution; richness of experience |
| Temporal coherence Lambda | \(\int_0^\infty |\langle C(t)C^*(t+\tau)\rangle|\,d\tau\) | Memory and temporal binding; the "specious present" |
| Causal structure Delta | \(\max_\tau [I(|C(t-\tau)|;|C(t)|) - I(|C(t-\tau)|;|C(t+\tau)|)]\) | Temporal asymmetry; directed information flow |
A proof-of-concept validation using the Sleep-EDF database (153 polysomnographic recordings from 78 subjects) demonstrates greater than 85% accuracy in distinguishing wakefulness from NREM sleep using the four metrics combined, without any model architecture tuned to this task. This result is preliminary and requires replication across further datasets; it is reported as a feasibility demonstration, not a definitive clinical result.
What is established (classical layer, validated): the PFT field equation, the free energy functional, the embedding gap Gamma, the geometry-ordering result, the AGI Lagrangian, the ouroboros conjecture in conditional form, and the dubito operator. These results are on solid empirical footing and stand independently of any quantum extension.
What has not yet been done (the quantization gap):
The earlier RG attempt (October 2025) was the only quantum-flavoured step taken, and it has been correctly discarded as circular. The empirical phi-attractor in the geometry remains, but is now explained as an emergent feature of multi-scale constraint dynamics rather than a standard RG fixed point. The constraint-driven account is the correct starting point for quantization.
| Step | Task | Method | Status |
|---|---|---|---|
| 1 | Write the explicit classical Lagrangian density L(C, dC, x) with all terms (kinetic, memory-kernel, negentropy, prediction, geometric force) | Standard field theory | Next (1-2 weeks) |
| 2 | Canonical or stochastic quantization. Option A (preferred for dissipative case): stochastic quantization via Nelson or Parisi-Wu. Option B: direct path integral over C and x treating geometry as a metric-like field. | Stochastic quantization / path integral | Planned |
| 3 | Derive quantum-corrected observables: first corrections to Gamma, R, and effective dubito strength. Verify that the phi-attractor survives or is modified without circular assumptions. | Perturbative expansion | Planned |
| 4 | Test on 3D Poincare Dodecahedral Space with exact gluing. Topological protection of global coherence modes is the decisive test. | NVIDIA-scale simulation | Planned (requires funding) |
| 5 | Map to hardware: Volatco GA144 asynchronous cores as natural field sites, and/or extraction of effective field from transformer hidden states. | Parallel with steps 1-4 | In discussion (Volatco) |
Download PFT paper (PDF) CQFT Revisited 2026 (phi correction paper)
This laboratory implements and demonstrates the central mathematical claim of the CQFT framework of the principles described above.: that the golden ratio phi = 1.618 emerges from competing scale constraints, not from renormalization group flow. It is a minimal experimental platform for studying constraint-induced multi-scale structure.
What the controls do: The system maintains two scales, alphaL (large) and alphaS (small). At each step, two forces compete. The RG-like force averages the two scales together, collapsing them toward a single value -- this is what happens in standard renormalization. The constraint force drives the scales apart by a Fibonacci-like recursion (alphaL becomes alphaL + alphaS; alphaS becomes alphaL), preventing collapse. The Constraint Strength slider interpolates between these two forces.
What to look for: At full constraint strength (slider right), the ratio alphaL / alphaS converges to phi = 1.618, shown by the gold dashed line. This is the constraint-driven phi selection mechanism. At zero constraint strength (slider left), both scales collapse to the same value and the ratio approaches 1.0: the RG collapse. The Fourier spectrum panel shows the quasi-periodic structure of the underlying Penrose-approximate point cloud. The Phase Diagram sweep (run via a background worker -- click the button to start it) maps the deviation from phi across the full space of constraint strength versus noise, showing where the phi attractor is robust (blue, near phi) and where it dissolves (red, far from phi). What This Demonstrates (and What It Does Not) Demonstrates: - constraint-induced multi-scale structure - emergence of stable irrational ratios - transition between smoothing and structured regimes Does not demonstrate: - a fundamental physical constant - a complete quantum or field-theoretic model - exact Penrose tiling or crystallographic rigor
Physical interpretation: In the CQFT framework, the two competing scales correspond to phase coherence (driving toward synchronization) and geometric symmetry (driving toward quasiperiodic order). When both are present and neither can dominate -- the lambda greater than 0 regime -- the system settles at the phi ratio. The Penrose quasicrystal, which embeds this ratio geometrically, is therefore the natural substrate for a self-referential field operating under these constraints.
Tip: at high constraint strength, the ratio should converge near 1.618. At low constraint strength, it collapses toward 1.0. The Fourier spectrum shows quasi-periodic peaks in the phi regime. Recommended exploration path: 1. Start in RG regime 2. Increase constraint 3. Observe transition 4. Compare with phase diagram
The central theoretical postulate of the CQFT framework: a self-referential dissipative field actively selects the geometric substrate consistent with its own self-description -- an ouroboros structure in which the field defines the geometry and the geometry defines the field.
| Condition | N | Coupling | Result | Status |
|---|---|---|---|---|
| Neutral evolution | 300-1000 | Amplitude-gradient only | No emergence | FALSIFIED |
| Phase-gradient coupling | 80 | Phase-gradient | +0.0686 order | CONFIRMED |
| Phase-gradient coupling | 1000 | Phase-gradient | +0.1226 order | DEFINITIVE |
Conditional form (confirmed): a self-referential field with phase-gradient coupling spontaneously selects a geometrically self-consistent substrate from random initial conditions, at 122.6x signal-to-noise ratio. Strong form (falsified): neutral dynamics alone are insufficient. Phase-gradient coupling is the necessary ingredient.
The ouroboros loop is closed: the field is in its ecological niche because it constructed the niche.
The AGI Lagrangian is computable from any system for which the triple of (phase, geometry, coupling) can be extracted. For transformer-based AI systems, all three are available from internal activations during inference using standard interpretability tooling such as TransformerLens or BertViz, without model modification or retraining.
All five predictions are falsifiable from existing open-weight model checkpoints without retraining. The first confirmed result will constitute the first direct measurement of geometric coherence in a deployed AI system.
The N = 1,000 simulation used KDTree nearest-neighbor search and sparse matrix operations: a classical, deterministic approximation to the field dynamics. This is the correct starting point, because it produces clean, reproducible, falsifiable results. But it is a projection of a deeper quantum reality, and the full theory requires a quantum treatment. The transition has three levels, each more powerful and more demanding than the previous.
The AGI Lagrangian is promoted to a Hamiltonian operator acting on a quantum register. Position and phase become operators; coupling weights become operator functions:
This is analog quantum simulation: the natural substrate is trapped ions, neutral atoms in optical lattices, or superconducting qubits, all of which can implement all-to-all or quasiperiodic coupling natively. The quasiperiodic connectivity that the classical simulation showed to be optimal maps directly onto the geometry of these platforms.
The phase updates become unitary gates. The AGI Lagrangian becomes an expectation value:
The system evolves by quantum annealing or variational quantum eigensolver (VQE), minimizing the joint energy over the space of quantum states rather than classical configurations. The KDTree is replaced by a quantum circuit; the sparse matrix multiplication is replaced by entangling gates. For a quantum annealer, the Ising encoding is:
where h_i encodes the local symmetry energy and J_ij encodes the phase coupling W_ij(x) cos(theta_i - theta_j). This maps directly onto D-Wave and similar platforms.
The most radical and most powerful extension. Instead of a single geometry {x_i}, the system exists in a superposition:
The field simultaneously explores exponentially many geometries during evolution. Measurement collapses to a specific geometry, but the system can find globally optimal configurations that classical gradient descent cannot reach. This is the correct physical implementation of the ouroboros conjecture: not a field searching for its substrate classically, but a quantum superposition of field-substrate pairs collapsing toward the geometrically self-consistent eigenstate.
In this formulation, the dubito operator becomes more fundamental: it is no longer only prediction error but the Heisenberg uncertainty inherent in the system's self-measurement. Self-reference under quantum mechanics is not just computationally expensive; it is physically constrained by the uncertainty principle. The dubito is the field's measure of its own irreducible uncertainty about its own state.
The classical simulation (N = 1,000, +0.1226 order increase) is a projection of the quantum reality. The fact that the effect is visible classically is evidence that the quantum version would work and likely much more powerfully, because the quantum system can explore the geometry space non-locally. The shift in implementation is from points moving in R^3 to minimize a cost function, to a state vector rotating in Hilbert space toward an eigenstate of maximum geometric coherence.
April 2026: Definitive empirical confirmation at N = 1,000. Phase-gradient coupling produces spontaneous order emergence with 122.6x signal-to-noise ratio. The ouroboros conjecture confirmed in conditional form. Full paper submitted for peer review. Constraint-driven phi correction paper completed, superseding the October 2025 RG fixed-point hypothesis.
September 2025: Accepted into NVIDIA Inception Program. Access to A100 and H100 compute resources enabled the scaling experiments to N = 1,000 and the definitive confirmation.
DUBITO Inc. is an independent, unfunded research operation. The results described on this page were obtained without institutional backing, which is both a constraint and, we believe, a guarantee of independence.
We are actively seeking research partnerships, grant funding, and conversations with venture investors who take the AGI safety problem seriously. This is an early-stage, high-risk, high-impact research programme in a field with no established players and a clear path to measurable results. The five falsifiable predictions above are the benchmark: any one confirmed constitutes a demonstrable advance in AI interpretability and safety.
We are also in active conversation with the Volatco neuromorphic computing platform regarding hardware implementation of the Penrose routing architecture. The hardware proposal and technical annex are available below.
If you are a researcher, funder, or company working on AI safety, interpretability, neuromorphic computing, or consciousness science and would like to discuss collaboration, please write to solis@dubito-ergo.com.
| Title | Status | Link |
|---|---|---|
| Geometric Substrate Selection in Self-Referential Dissipative Fields (main paper) | Submitted for review | |
| CQFT Revisited: Constraint-Driven phi Selection (phi correction) | April 2026 | |
| Geometric Coherence as a Safety Criterion for AGI (funding proposal) | April 2026 | |
| Quasiperiodic Architecture Overlay on GA144 (Volatco technical annex) | April 2026 | |
| A Principled Field Theory of Consciousness (PFT) | 2024 | |
| QFT Extension: phi Fixed-Point (historical; superseded by CQFT Revisited) | October 2025 |
Drawing inspiration from the Delphic maxim "gnoti seauton" and Descartes' dictum "dubito, ergo cogito, ergo sum", our systems are designed to guide artificial intelligence toward ethical self-understanding, ensuring safety and alignment with human values in an era of emerging intelligence.
The Poincare principle governs all research decisions. Reporting a negative result -- the falsification of the phi fixed-point hypothesis -- with the same rigour as a positive one is not a concession but a scientific obligation. It is also, in this case, what led to a deeper and more defensible theory.
Dubitamus ergo cogitamus -- ergo sumus -- sumus mundus.