# MARION MORANETZ (541) 507-7549 | Marion.m.moranetz@gmail.com | linkedin.com/in/melmarion github.com/melmarion | Open to Relocation --- Applied behavioral scientist who builds detection and measurement systems for online influence. Architect of Persuasion-Max, a 5-layer pipeline that decomposes persuasive content into measurable components — identifying manipulation techniques, predicting recipient vulnerability, and quantifying context-dependent harm. 126K human interactions processed. Looking for Trust & Safety roles where understanding influence mechanics prevents real harm at scale. --- ## INFLUENCE DETECTION & MEASUREMENT ### Persuasion-Max — Influence Detection & Prediction Engine | 2026 *Open-source 5-layer pipeline that detects persuasion techniques in content, models recipient vulnerability, and predicts behavioral impact. Grounded in cognitive appraisal theory (Smith & Ellsworth 1985) and somatic marker hypothesis (Damasio).* - **Built 40-technique detection layer** using Zeng SemEval taxonomy: identifies emotional appeal, fear/threat framing, logical fallacies, social proof manipulation, authority exploitation, scarcity tactics, whataboutism, and 33 other influence techniques in arbitrary text. - **Engineered 16-dimension recipient vulnerability model** (Big Five personality, Moral Foundations, cognitive style, ideology profile) that predicts susceptibility to specific techniques. 21.8pp compliance spread across persona types — the same manipulative message affects different people dramatically differently. - **Quantified defensive technique risk**: crisis PR analysis found that whataboutism triggers 46.5% retaliation probability. Defensive framing backfires more often than it deflects. This type of finding directly informs content moderation policy. - **Validated domain-specificity of harm models**: cross-domain weight transfer degrades AUC by 10-13pp, confirming that manipulation detection models trained on one context (e.g. political) perform poorly in another (e.g. commercial) without retraining. Universal "toxicity" models miss context-dependent harm. - **Discovered that technique binary detection adds zero incremental signal** — detecting THAT a technique is present matters less than detecting HOW it's deployed (linguistic surface features). This finding challenges common content moderation approaches that flag technique presence without context. - **329 automated tests, 302 parameters with provenance** (1% fitted, 61% literature-constrained, 37% uncalibrated). Full ablation analysis. Paper-shaped documentation. ### Influence Detection Tools (Browser-Based) | 2025-2026 - **ClearFrame**: propaganda/influence detection interface that surfaces rhetorical techniques in real-time content. - **Dead Internet**: social network narrative manipulation detector — identifies propaganda playbooks (Wag the Dog, Overton Window), velocity spikes, and linguistic fingerprints suggesting coordinated inauthentic behavior. - **SNA Framework**: social network analysis toolkit for mapping influence propagation patterns. ### Content Strategy Platform — Measuring Influence at Scale | 2026 - **Built 12 community linguistic norm profiles** (reading difficulty, emotionality, pronoun ratios) for Reddit communities. Content deviating >1.5 sigma from norms flagged automatically. This same detection logic applies to identifying astroturfing and coordinated manipulation that breaks community norms. - **284 quality-gated training examples** with triple quality gate: >60% predicted effectiveness, organic-passing on influence detector, platform-appropriate linguistics. The "organic-passing" detector is itself a trust & safety tool — it identifies content designed to evade detection. ### Fractionation Research (Persuasion-Max Core Thesis) | 2026 - **Documented the mechanism of algorithmic emotional sequencing** (fractionation: rapid cycling of emotional states A-J-A-R) and its effect on user decision-making capacity. Research bridges tech (engagement metrics), neuroscience (prefrontal cortex disengagement), and economics (irrational purchasing). - **Identified that awareness does not prevent effect** — users who understand they're being manipulated are still behaviorally affected. This finding has direct implications for disclosure-based safety interventions. --- ## PRODUCT DEVELOPMENT - **80 repositories** including 34 iOS apps (persuasion trainers, narrative games), 39 browser projects (influence detection tools, engagement systems), and 7 Python systems. - **27 behavioral design frameworks** (Claude Code skills) codifying engagement mechanics, gamification architecture, and neurochemical targeting — understanding how these systems work is prerequisite to detecting when they're weaponized. --- ## TECHNICAL PROFICIENCY **Detection & Analysis:** NLP feature engineering, technique taxonomy classification (Zeng 40-technique), linguistic norm profiling, coordinated behavior detection, statistical significance testing (bootstrap CI), ML evaluation (AUC/ROC), interaction term discovery **Behavioral Science:** Cognitive appraisal theory, Elaboration Likelihood Model, Moral Foundations Theory, fractionation/emotional sequencing, Big Five vulnerability modeling, persuasion resistance research **Languages & Frameworks:** Python, JavaScript/React, Swift/SwiftUI, scikit-learn, FastAPI, Canvas API, SQL **ML/AI:** LLM fine-tuning (Ollama/7B), Claude API, logistic regression, NLP, prompt engineering **Tools:** Git (80 repos), VS Code, Claude Code, MCP protocol --- ## EDUCATION **Southern Oregon University** — B.S. Innovation & Leadership (3.76) | Minor: Rhetoric & Reason | 2025 Coursework: Persuasion & Negotiation (A), Data-Driven Decision Making, High-Stakes Communication **Southwestern Oregon Community College** — A.S. General Studies (3.8, top 5%) | 2023 --- ## KEY METRICS - 40 persuasion techniques detectable with Zeng SemEval taxonomy integration - 21.8pp compliance spread demonstrating recipient-dependent vulnerability - 46.5% retaliation probability on defensive whataboutism — quantified backfire risk - 10-13pp AUC degradation on cross-domain transfer — domain-specific detection required - 126K human interactions processed with full provenance tracking - 414 automated tests across primary codebases