ArrmesNeuralRouterV3 (v10 Temperature-Calibrated Release)

Sub-millisecond CPU multi-head hashed n-gram MLP and typed argument span extractor for ARRMES v1.7.2 (v10).

Headline Metrics

  • Sealed acceptance-v8 (N=3,135): positive exact accuracy 97.86% verified (97.57% Track A raw), hard-negative accuracy 96.14% verified (95.10% Track A raw).
  • Full acceptance-v7 (N=2,580): 100.00% positive (1,380/1,380) and 100.00% hard-negative under v3 contract (1,200/1,200; 99.58% or 1,195/1,200 when restricted to v2-only contract), with three-platform bitwise-identical predictions (SHA-256 f23e6de69333...) across macOS ARM64, Linux x86_64, and Windows Server 2025 AMD64 (including the zero-C-extension pure-stdlib array matrix backend).
  • Open-phrasing overnight-v1 (N=80): 100.00% exact accuracy across macOS, Linux, and Windows.
  • Footprint and latency: 7.48 MiB compressed .npz, 0.416 ms Apple M4 Pro CPU p50, 0.714 ms AMD Ryzen 9 7950X3D CPU p50, 4.48 ms Windows Server 2025 AMD64 pure-stdlib CPU p50 (79.27 MiB Win32 PeakWorkingSetSize).
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Collection including NorviTech/arrmes-neural-router-v3