# System Card — AMF Agent RC8A ## Scope AMF Agent RC8A is a capability-routed agent system qualified on the frozen Agent mission contract. The qualified system has two neural capability roles: 1. `TOOL_EXECUTION_ARGUMENTS` 2. `DIRECT_RESPONSE` Policy, authorization, routing, deterministic envelope construction, and refusal paths are system-level responsibilities rather than unconstrained neural generation. ## Claim boundary The release proves the behavior measured by the included qualification protocol and physical evidence. It does not claim universal superiority over Qwen or other systems outside the tested mission/domain. The public artifact is a `PINNED_RECONSTRUCTION`, not a newly trained AMF foundation model. ## Execution authority The constitutional runtime rule is: > No Capability Authorization => No Execution. Known deterministic deployment constraints are hard gates before physical optimization. ## Base substrate - Model: `Qwen/Qwen3-4B-Instruct-2507` - Revision: `cdbee75f17c01a7cc42f958dc650907174af0554` - Realization mode: RAW - Runtime quantization: none in the winning graph - Dynamic PEFT in winning hot path: none ## Physical environment - GPU: NVIDIA A100-SXM4-80GB - CUDA: 12.8 - PyTorch: 2.8.0+cu128 - Python: 3.12.3 - Authority network: FORBIDDEN - Model source: OFFLINE_PINNED_SNAPSHOT See the machine-readable receipts under `evidence/selected/` and `system/`.