| Tag coverage | Only a small share of the catalogue is tagged (65 of about 8,000 games at the time of writing), so "tagged alike" can only speak about those games | More AI tagging runs and review; the holdout test shows tags are the strongest lever |
| Tag types | Tagged games are all slots, so tests on tagged games say nothing about table or live players | Tag table and live games |
| Batch recency | The batch profile decays by the last-played date, not per bet; heavy play 25 days ago plus one bet yesterday looks recent | The live path already fixes this with per-event decay |
| Evaluation | Accuracy is measured on simulated customers whose rules we wrote | Run the same holdout test on real activity |
| Player protection | A switch the recommender obeys, not a risk detector | Feed the list from the responsible-gambling system, possibly over the compacted suppression topic |
| Licensing | No country or licence rules; a game not allowed in a market could be recommended | Add availability to the catalogue and filter candidates |
| Currency | Only EUR is accepted | Convert at ingestion or keep per-currency totals |
| Crowd scale | Co-play is computed in memory on a sample (250,000 customers by default) | Batch job with ALS or sharded counting |
| Duplicate bets | The live path catches replays, not the same bet sent as two messages | Bet-id dedup in the Lua script or upstream |
| Redis crowd sampling | SCAN walks one node | Per-shard scan for Redis Cluster |
| Simulator writes | Plain messages only; ingestions with a Schema Registry cannot be written by the simulator | Registry-aware publisher |