Training Data and the Maladaptive Mind: A Computational Framework for Developmental Trauma

Abstract Traditional trauma theory frames adverse childhood experiences as damaging events that require healing. We propose a computational reframing: trauma represents maladaptive learned patterns arising from suboptimal training environments, functionally equivalent to problems in machine learning systems trained on poor-quality data. Key Contributions Four training data problems: Direct negative experiences (high-magnitude negative weights), indirect negative (noisy signals), absence of positive (insufficient examples), limited exposure (underfitting) Dissociation as meta-learning: Second-order learning where the system learns that cognitive engagement itself predicts overwhelm, producing preemptive processing suppression PTSD vs CPTSD distinction: Catastrophic single-event learning (localized weight perturbation) vs chronic adversarial training (systematic weight landscape distortion) Engineering solutions: Increased caregiver diversity and community-based child-rearing as training data diversification Status Preprint. Working Paper DP-2501. Links Research Square: doi.org/10.21203/rs.3.rs-8634152/v1 Code & Data: github.com/studiofarzulla/trauma-training-data ASCRI: systems.ac/5/DP-2501 Research Lab: Dissensus AI v1.1.0 (2026-07-11): Single-seed honesty correction: ±SD claims implying multi-seed runs removed (headline results are single-seed and now labeled as such); label-smoothing ablation wording corrected (the released artifact contains no systematic loss-function ablations; preliminary exploration suggests loss-dependence); three contradicted or grey-literature citations dropped.