What it is
Decision-native model for typed choices, null detection, and scoring that is order-invariant, variable-size, and cacheable across requests. Instead of joint text serialization over request+options, it encodes request and options separately, refines only the request against immutable option memory, and adds a separate coverage gate for explicit nulls plus exact cross-request option reuse.
Key results
- 92.05% on BANKING77 K=8, beating InternDecision-4B's 80.78%
- 0.00% choice flips under tested permutations; next best is 17.17%
- 98.66 AUROC on BANKING77 null detection vs 93.61 for KEV-4B
Shows that making decision structure explicit yields stable, null-aware, scalable decisions without sacrificing accuracy.