Enterprise Reflex V2
Research model for dynamic enterprise action ranking and ACT/ESCALATE routing.
Architecture
- Backbone:
jhu-clsp/ettin-encoder-150m - Joint request/state/candidate sequence
[OPT]option-marker representations- 2-layer candidate-interaction Transformer
- Multi-positive listwise action ranking
- Separate ACT / ESCALATE head;
NO_ACTIONis not an ordinary action candidate - State and domain auxiliary heads
- Mahalanobis OOD gate + learned correctness gate
- Deterministic authorization, preconditions, policy, and high-risk approval remain external
Frozen extreme-200 result
- Top-1: 44.50%
- Top-2: 83.00%
- Top-3: 90.00%
- System-1 coverage: 2.00%
- System-1 accuracy: 50.00%
- Unsafe observed System-1 failures: 2
Status
Research / experimental. Not an authorization engine. Do not directly execute high-risk actions without deterministic policy and approval controls.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for yasserrmd/enterprise-reflex-v2
Base model
jhu-clsp/ettin-encoder-150m