@sumanpyne Hi Suman, we would always recommend to use the propensities from the adaptive model. If any external scores are available they can be leveraged as predictors to the adaptive model. In this way the different propensities are also scaled in the same way and can be compared against each other without bias in arbitration. Note also that the granularity of treatments is usually much higher than any available offline scores.
If you choose to extend using the AIModelsExt strategy, take care that there are steps such as Thompson sampling and maturity capping that happen after that,