Exploratory modeling, draft simulations, and notes from the field

Exploration: Incentives and Compliance
Testing Goodhart’s Law in Report Creation

We simulated a broad set of incentive and policy approaches to assess whether Goodhart’s Law applied in each. For the below hypothetical case study assessing non-compliance of reporting among project leads, the model revealed that instituting outcome-aligned 'filed report reuse bonuses' (paying for reports that were valuable to others) far outperformed incentive-free cultural training. (However, bonuses for filing reports without re-use incentive tie-ins led to greatly reduced benefits.) We utilized an agent-based model with OCEAN personality traits.

Insight: Bonuses for filing reports without tie-ins to downstream utility led to counterproductive gamification and reduced benefits.

Method: Agent-Based Model (ABM) with OCEAN personality profiling.

For a small-scale online social network (school environment), we simulated structural interventions to decouple network vitality from toxic content velocity. For the below hypothetical case study on platform safety, the model revealed that a 'selective permeability' architecture — high structural barriers (echo chambers) mediated by algorithmic bridges — far outperformed standard censorship. (Paradoxically, strong homophily acted as a containment vessel for toxicity while allowing high-consensus utility to tunnel through.) We utilized a proprietary Agent-Based Model with OCEAN psychometric agents for this evaluation. (LLM-based agents were also explored in this study.) Note that the toxic content detector was naive, based on imagined variance in response times for students based on different types of content, mediated by personality type.

Insight: Paradoxically, strong homophily acted as a "containment vessel" for toxicity, preventing spread while allowing high-consensus utility to tunnel through.

Method: Proprietary Agent-Based Model utilizing OCEAN psychometric agents, also including LLM-based agent exploration.

Note: Detection thresholds were modeled based on variance in response latencies across personality types.