AURA - Adaptive User-Responsive Architecture

Seven participants played a custom-built game (CollectGame) across three game modes in a within-subjects randomised design. A NeutralityScore metric (weighted composite of sparsity, death rate, and standard deviation) identified Mode 1 as the neutral baseline environment. Calibration artefacts - centroids, scaler parameters, and initial multiplier - are frozen permanently and never retrained with adaptive data, preserving the integrity of the behavioural reference.

An 8-notebook Python pipeline processed 45 filtered players across 3,240 30-second telemetry windows using a 12-feature schema (10 raw + 2 derived: damage_per_hit and pickup_attempt_rate). K-Means clustering (K=3, silhouette score = 0.3752) identified three behavioural archetypes: Combat, Collection, and Exploration. Soft membership is computed via Inverse Distance Weighting (IDW) against cluster centroids. An ANFIS model was trained offline, and an MLP surrogate (architecture 6→16→8→1, ReLU hidden layers, linear output) achieved Test R² = 0.9264 and Test MAE = 0.0127 - comfortably exceeding the 0.90 threshold.

A Node.js/TypeScript backend on Render implements the full inference pipeline: telemetry normalised → archetype activity scores → IDW soft membership → window-to-window behavioural deltas → 6-dimensional state vector into MLP surrogate → neutral-centred calibration → multiplier clamped to [0.6, 1.4]. Temporal smoothing (α = 0.2) prevents abrupt jumps. A forward-only constraint ensures multipliers affect only future PCG content. MLP inference latency under 1ms; full backend round-trip under 200ms on Render's free tier.