| license: mit | |
| library_name: onnxruntime | |
| tags: | |
| - tool-calling | |
| - pathway-routing | |
| - onnx | |
| # Sonus Needle Pathway Shadow | |
| Fine-tuned Cactus Compute Needle checkpoint for Sonus request-scoped pathway | |
| tools. The model emits bounded `route_NN`, `slot_NN`, confirmation, or `stay` | |
| calls; the Sonus runtime validates every call against the active graph. | |
| This checkpoint is **shadow-only**. It must not control caller state without an | |
| independent acceptance gate and authoritative fallback. | |
| ## Evaluation | |
| - Independent stress/Noura full-schema gate: 41/77 (53.2%). | |
| - Guard-first, MiniLM top-two request shape: 50/77 local result accuracy. | |
| - High-confidence route acceptance (`margin >= 0.15`): 27/48 accepted, | |
| 27/27 correct on the disjoint gate. | |
| - Local ONNX latency: approximately 254 ms median, 339 ms p95 on the evaluation | |
| host. | |
| The full-schema score is not sufficient for authority. The staging deployment | |
| keeps Ministral authoritative and records only value-free Needle comparisons. | |
| ## Files | |
| - `encoder.onnx` | |
| - `decoder_step.onnx` | |
| - `needle.model` | |
| Training data was generated from Sonus-owned synthetic/test pathways and | |
| contains no call recordings or customer transcripts. | |