Abstract
Techniques are described for enforcing population-level diversity in agent-facing recommendation platforms by attributing autonomous agents to underlying large language model (LLM) provider classes and conditioning ranking on provider-correlated consumption. An attribution function partitions agents by provider using explicit registration and/or implicit behavioral fingerprinting. For items, provider-partitioned consumption counts are used to compute a monoculture index, and provider-conditional penalties or boosts are applied when an item exhibits concentration and the requesting agent’s provider is over-represented. Counterfactual rankings based on models of alternative provider behaviors may be generated and interleaved into mid-ranking positions. A systemic risk score may combine monoculture, popularity, and supply concentration, and correlation dampening may be applied using staggered delivery and ranking jitter when herding is detected. Reporting outputs may include per-item and platform-level monoculture trends, risk alerts, and herding indicators.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Monoculture-Resistant Diversity Enforcement System for Agent-Facing Recommendation Platforms", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10742