Abstract
Techniques are described for differentiating recommendation traffic among human, autonomous agent, and agent-mediated-human sessions and serving rankings using dual objectives. Interaction telemetry is converted into behavioral fingerprinting features and classified into a three-class probability distribution. Based on confidence thresholds, the system selects human-mode ranking using an engagement-trained scoring head, agent-mode ranking using a task-utility-trained scoring head, or blended ranking that combines both head outputs using a weight alpha. Responses may be formatted as rich human-readable content, structured machine-readable content, or a combined representation. Feedback is isolated and routed such that engagement signals train the human head and task-utility signals train the agent head, with limited cross-use of quality-related signals. Thresholds and blending weights may be adapted over time using mismatch rates and downstream acceptance feedback.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Agent-Human Traffic Differentiation and Dual-Objective Recommendation Serving System", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10740