The Aggregation Fallacy: Beyond the Illusion of Precision in Forensic Violence Risk Assessment

This project provides reproducibility and reporting materials supporting the manuscript “The Aggregation Fallacy: Beyond the Illusion of Precision in Forensic Risk Assessment.” The paper addresses a recurring problem in forensic mental health practice: multiple risk instruments are frequently “stacked” and narrated as convergent corroboration, even when instruments share constructs, inputs, and measurement error. In such cases, the appearance of increased confidence may not reflect increased evidentiary independence, and combination rules—especially implicit OR-logic (“high risk if any tool is high”)—may inflate false positives and distort proportionality in legal decision-making. Purpose. The project operationalizes the manuscript’s core claims in a transparent, audit-ready way. It supplies (i) a dependence-sensitive benchmark demonstrating how OR-logic inflation changes under plausible inter-tool dependence, (ii) derived outputs (tables and figures) used in the manuscript, and (iii) drop-in LaTeX appendices and reporting scaffolds designed to improve courtroom literacy and contestability. The goal is not to produce field prevalence estimates, but to provide a clear structural demonstration of how base rates, redundancy, and integration rules jointly shape error burdens that are legally material. Methods and contents. The repository includes a reproducible simulation benchmark (Box A) based on a latent-Gaussian threshold model with exchangeable inter-tool correlation (ρ). Parameters reflect an illustrative setting in which each tool has marginal false-positive rate α and sensitivity Se, and the OR-rule classifies as positive if any tool flags “high risk.” Monte Carlo simulations generate the reported false-positive and true-positive rates, confidence intervals, and fold-inflation relative to a single tool. The repository also contains scripts to regenerate the benchmark outputs and to recreate manuscript figures illustrating false-positive behavior under OR/AND rules and the break-even conditions under which OR-logic could be decision-rational given asymmetric error costs. Expected outcomes. (1) A reproducible, dependence-aware benchmark that readers can rerun, adapt, and critique; (2) derived tables/figures that match the manuscript reporting; (3) LaTeX appendix material suitable for direct inclusion in journal supplements; and (4) audit scaffolding (assumptions and changelog) to support transparent scholarly and courtroom communication. More broadly, the project aims to promote Parsimonious Validity: anchoring evaluation on one best-fit instrument for the referral outcome/horizon and integrating genuinely orthogonal evidence (dynamic change, protective factors, and disconfirmatory information) with explicit uncertainty. Ethics and data statement. This repository contains no identifiable human participant data. Examples are synthetic/illustrative. The simulation benchmark is a structural demonstration and should not be interpreted as a direct estimate of real-world violence risk or tool performance in any specific jurisdiction. How to use. Users can reproduce the benchmark and regenerate outputs by installing dependencies listed in the repository and running the included pipeline script. Derived outputs are provided for convenience and traceability.