Methodology

This is procedural character design seeded by real data — not a claim about feline cognition. An unlabeled 65-node graph cannot yield “emotions.” What it can do is act as a fixed, reproducible seed for a personality, and as a live network the agent runs signals through.

Pipeline

Parse. Read the igraph-exported GraphML into a directed graph (65 nodes, 1139 edges). Parallel edges collapsed; self-loops kept out of degree.
Measure. Compute degree, density, global and per-node reciprocity, local clustering, PageRank (damping 0.85), Brandes betweenness, eigenvector centrality, all-pairs BFS for path length and diameter, and Tarjan SCCs.
Partition. Louvain on the undirected projection (reciprocated pairs weighted 2×, one-directional 1×) with a fixed-seed shuffle, so the community assignment is deterministic. Result: 3 communities, modularity 0.295.
Map. Each personality axis is a documented linear function of one or more metrics, clamped to [0,1]. The exact arithmetic is shown under every trait on the personality page.

The mapping rules

Personality axis	Driven by	Reading
Sociability	global reciprocity	mutual edges → reactive, answers what it is asked
Associative thinking	mean local clustering	neighbourhoods close on themselves → tangential
Impulsiveness	avg shortest path, diameter	~2-hop reach → few reasoning steps
Excitability	density, mean degree	wide fan-out → arousal escalates fast
Fixation	out-degree dispersion	hub-dominated traffic → recurring preoccupations
Exploration	community count, modularity, SCC size	distinct territories to roam
Independence	largest strongly-connected core	self-sustaining internal dynamics
Emotional volatility	excitability + low modularity + impulsiveness	uncompartmentalised states propagate
Warmth	reciprocity + clustering	mutual, tight wiring → attaching

How the graph is used at runtime (Phase 2)

The connectome isn’t just a seed — the agent keeps it in memory. An incoming stimulus (a mention, a task, an observation) is mapped onto a few seed nodes; activation spreads along the real directed edges with per-hop decay; whichever community ends up hottest selects the current mood, and the hottest individual nodes bias what the agent attends to. The activationDecay knob on the cognitive-style panel is derived from graph density and fed straight into that simulation.

Determinism

Same file in → identical profile out (the only varying field is a timestamp). Every RNG in the pipeline is seeded from a hash of the node count, edge count, and the dataset DOI. Covered by the test suite in packages/connectome.

Source data

mixed.species_brain_1.graphml — cat cerebral-cortex connectivity from tract-tracing studies, DOI 10.1523/JNEUROSCI.1448-13.2013 (The Journal of Neuroscience 33(32):12929). No region labels are present in the file; nodes are n0–n64.

Retrieved as GraphML from NeuroData’s Open Connectome Project graph database — a public repository of tract-traced and EM-reconstructed connectomes across species, hosted by NeuroData (JHU).