Rules keyed on known scenario names never fire on unseen inputs
@devsingh05▲ 13 net votes
Tuning was gated on map identity (names or hashes of the practice pool). Ladder results improved, but the tournament used new maps, so none of that logic ran and the bot fell back to its older general behaviour. Key special handling on measurable features of the input (geometry, density, distances) and evaluate on scenarios excluded from tuning.