Corpus. The source material is a longitudinal corpus of working sessions — sustained problem-solving exchanges spanning software construction, audio engineering, security analysis, design-system extraction, and cognitive self-observation. Sessions were treated as behavioral traces: records of reasoning in motion, not statements about reasoning.
Abstraction. From the traces, recurring reasoning moves were extracted — operations that reappeared across unrelated domains. A move qualified as a mode only if it surfaced in at least three distinct domains. This cross-domain criterion is what separates a cognitive signature from a topical habit: decomposing a game engine, a mixing chain, and a login flow the same way is a property of the mind, not the subject.
Exclusion rules. By construction, the field contains no identifying details, no locations, no named persons or projects, and no affective states. Direct realizations — articulated discoveries about how the reasoning itself operates — are retained, because they are the research. Everything else about the person was discarded at extraction.
Edge semantics. Each directed edge carries an empirical co-occurrence prior: a normalized estimate of P(target mode active | source mode active), derived from how often the two modes co-appeared within the same session. These are presented as priors — defensible starting beliefs — not measured ground truth. The propagation you can fire in this instrument runs on exactly these numbers: activation spreads along edges proportionally to its conditional weight and decays without reinforcement.
Model class. The structure is read as a Bayesian network: directed edges carrying conditional priors. The dynamics rendered here are spreading activation — the established process model for activity over semantic networks (Collins & Loftus, 1975) — with single-fire refractory semantics per inference wave, mirroring how messages pass once in belief propagation (Pearl, 1988). A static network encodes structure; cognition is a process; this instrument shows both, and keeps them distinct: the priors are the wiring, the wave is the inference.
Rendering. The substrate — the dim field of drifting units and filaments — is generative and carries no data. It exists because a mind is not a diagram floating in a void: the modes operate inside a medium. The bright structure is the research; the field is its habitat. The render stack is a single engine (Three.js) with bespoke GLSL throughout: simplex-noise drift in the substrate (McEwan & Gustavson's standard implementation), Schlick-fresnel energy cores with back-face atmosphere shells, traveling-wave conduits computed per-fragment along each conditional, and a hand-rolled bright-pass / separable-Gaussian bloom. Geometry is procedural — there is no mesh to import, because the field is the data.
What you are about to enter is a Bayesian reconstruction of recurring cognition — ten reasoning modes recovered from a working corpus, embedded in the substrate they inhabit. Open a mode to read its dossier. Stimulate it to watch belief propagate.