command-center/agent/memories/sharpshooter/wt-20261013-185330-6d1fc1f-1k10ki5io1fbr/architecture.md

Runtime Context

• Tool call execution: user-perceived failures require confirmation; when regression=true is flagged, re-execute with contextual resolution.

• Power failure: system reboot caused by power failure.

Model Configuration

• KV context: 4×200,000 fp8 quantized kv slots.

• Calibration context: fp8.

• Concurrent streams: 3.

• Expert configuration: 49,152 GPTQ w4 experts (10M-token calibration on engine 1.2.0) + 24 bf16-protected experts/layer (energy-share pattern: bf16-protected experts and their resource allocation within a card's VRAM budget) within 29,120 MiB/card ceiling; 47.7 GiB n-gram copy pinned to host RAM.

Deployment Testing Policy

• Emergency bypass: E2E gate may be skipped with --skip-e2e flag during outages; normal flow requires all three gates including the final Playwright e2e suite.