Abstract
A system that repeatedly generates a fresh batch of candidate content topics for a person to choose from, once per day in this case, will eventually re-surface a topic that person has already engaged with in some way, unless the system tracks and consults that history. Doing this naively creates two opposite failure modes. Excluding too little re-shows the same ideas, and excluding too much eventually starves the system of anything left to offer, because a topic a person simply didn't pick that day (having preferred a different one in the same batch) is not the same signal as a topic they actively rejected. This disclosure documents a production mechanism that classifies topic history into two signal-strength tiers, a hard tier built from two independently-sourced signals with genuinely different decay behavior, and a soft, weaker tier, and enforces them asymmetrically. Only the hard tier is permitted to remove a candidate from today's batch; the soft tier is tracked but never causes a removal, which protects a promised fixed batch size from being eroded by ambiguous signal. Two real production incidents in which this batch-completeness property was itself incompletely enforced, and their fixes, are documented as part of the record, along with a real, dated worked example run against the live system.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Panday, Vipul, "Multi-Tier, Signal-Weighted Topic Exclusion for Recurring AI Content Suggestion, With a Batch-Completeness Guarantee", Technical Disclosure Commons, (September 02, 2026)
https://www.tdcommons.org/dpubs_series/11583