My current research is organized around a working hypothesis: intelligence — in biological cognition, neural networks, and engineered systems — may be better understood as the discovery, preservation, and revision of finite structure than as optimal search over an unconstrained space. I am interested in the computational theory implied by that view, and in building systems that expose where it succeeds, where it fails, and what architectural commitments it requires.
I treat the projects below less as separate topics than as stress tests of the same structural hypothesis across cognition, machine learning, computer vision, graphics, and language.
Theoretical base
The theoretical base began with a simple observation: single-scalar models of memory often conflate two logically distinct quantities — how much of a trace exists, and how faithfully it preserves its original encoding.
Separating these as structural accumulation and representational fidelity, then coupling them through a small dynamical system, produced the first version of the Structural Crystallization framework. The goal was not merely to fit a particular memory paradigm, but to ask whether a richer representation of memory state could explain phenomena that are difficult to express through unitary-strength models alone: the temporary character of extinction, longer-lasting change following retrieval-extinction, differences in persistence across memory conditions, and boundary conditions in spacing and reconsolidation.
A manuscript developing this framework was submitted to Psychological Review and received two substantive external reviews before being declined. The reviews challenged both the framework's theoretical positioning and several of its empirical and methodological claims, particularly its relationship to existing memory models, the justification of its assumptions, and the scope of its necessity and sufficiency claims. Those criticisms materially changed how I approach later work: I now try to distinguish more carefully between an interesting mechanism, an architectural commitment, and a claim that the mechanism is uniquely necessary.
A later manuscript, Crystallon: Learning by the Growth and Dissolution of a Discrete Knowledge Carrier, pushes the same idea toward learning systems. Instead of assuming that learning fills values into a fixed carrier — weights in a fixed architecture, or rules over a fixed vocabulary — Crystallon treats the carrier itself as dynamic. Discrete structural units can be created, deepened, or dissolved through access-driven consistency dynamics.
The claim is deliberately narrower than in the earlier work. I am not proposing a new learning paradigm so much as asking what becomes possible when the existence and persistence of representational structure are themselves part of the learning process.
Rendering, vision, and structural signals
A second line of work grew out of computer vision and rendering.
While investigating failures of a learning-based stereo system on transparent surfaces, I became interested in whether the failure was really a learning problem at all. In simulation, transparent regions exhibited markedly different Monte Carlo variance structure from non-transparent regions. That observation became a science-fair project and later led to broader questions about whether standard preprocessing and network components can systematically erase physically meaningful statistical structure before a model has any opportunity to use it.
This developed into work on rendering difficulty and light-transport structure. One recent paper questioned the use of finite-sample per-pixel variance as a stable ground-truth measure of rendering difficulty under heavy-tailed light transport, and proposed a mechanism-grounded representation based on path structure instead.
The paper was reviewed at Pacific Graphics 2026 and was not accepted in that round, but the committee referred it to Computer Graphics Forum for fast-track resubmission with major revision. The reviews were particularly useful because they did not reject the central observation outright; instead, they challenged the stronger framing around “ground truth,” the distinction between mechanism and difficulty, the statistical interpretation of variance instability, and the lack of a sufficiently concrete downstream use.
That process has pushed the revised version toward a more careful question: not whether transport mechanism is rendering difficulty, but whether stable structural descriptors can serve as useful priors or explanatory variables where noisy scalar targets are unreliable. That revised manuscript has since been submitted to Computer Graphics Forum, where it is currently under review; a preprint is available here.
A related vision manuscript was also submitted to BMVC 2026. Although it was not accepted, the review and rebuttal process provided a second source of technical feedback, and I have been using that material to rework the paper rather than treating the rejection as an endpoint.
Structural constraints in learning systems
The same structural perspective appears in several system-oriented projects.
In CrystalCache, I explored whether long-context KV-cache eviction could benefit from separating notions analogous to accumulation and representational impact rather than reducing importance to a single local score. The goal was less to invent another eviction heuristic than to test whether a structurally informed score could preserve useful context under constrained memory.
In CrystalASR, I approached a different question: what happens when higher-level linguistic knowledge is allowed to influence recognition without being allowed to overwrite lower-level acoustic evidence? A strict upward information-flow constraint produces a narrow regime in which lexical information can resolve ambiguity while remaining subordinate to phonetic evidence. Beyond that regime, language priors begin to dominate, and recognition quality deteriorates sharply.
I am interested in this pattern because it mirrors the vision work from the opposite direction. In one case, ordinary architectural operations may erase structure that should have been preserved; in the other, an explicit architectural constraint protects lower-level structure from being overwritten.
Toward a runnable theory of emergence
My current exploratory work tries to make the opening hypothesis more literal.
The Crystal Lattice Intelligence Engine simulates crystallization dynamics on a fixed two-dimensional torus. Concept domains form connected structures; repeated access can stabilize them; sufficiently strong unreferenced access can deform their boundaries; and collisions between structures create the possibility of new coherent organization.
The experiment is intentionally simple. I am not trying to reproduce intelligence in miniature. I am trying to find out whether a small set of local structural rules can produce forms of organization that were not explicitly specified in advance — and, equally importantly, what kinds of novelty they cannot produce.
That negative side matters to me. Much of the most useful feedback I have received has come from places where a proposed mechanism failed to support the claim I originally wanted to make.
How I work
Much of this work has been conducted independently, with experiments supported through self-funded compute.
I tend to work by building a runnable system early, then using failures to decide what the actual research question is. I keep detailed development and experiment logs so that failed paths, parameter changes, conceptual revisions, and implementation decisions remain inspectable rather than being rewritten away after the fact; the longest of these logs has grown to roughly 17,000 lines.
I have also increasingly tried to use peer review as part of the research process rather than only as a publication filter. Submissions to Psychological Review, TMLR, BMVC, and Pacific Graphics have exposed different weaknesses in the same broader habit of thought: overextending a structural observation into a stronger claim than the evidence supports. Learning to narrow those claims without abandoning the underlying question has become an important part of how I work.
My interests remain centered on computation, but the questions often cross disciplinary boundaries. I am especially interested in representation, memory, structural constraints, learning dynamics, computer vision, rendering, and the points where computational models become claims about cognition. I welcome correspondence from researchers whose work intersects any of the above: [email protected].