Version: 2026.07.11

Brain Router, Arbiter & GPU Flow

Local agents enter at 127.0.0.1:8788/mcp. The proof tool is retrieval_arbitrate_tool; raw recall, direct FTS, direct Qdrant and direct Map are bypasses and do not prove the product path.

Router and arbiter flow from MCP through bounded recall lanes, L2 dedup, governance scoring, CUDA reranking, budget and grounding. GRIFF-Map returns only subordinate evidence.

The semantic lane uses CPU-backed Qdrant over vectors produced by CUDA. CUDA is also proven for full-corpus semantic-edge maintenance and cross-encoder reranking. Current Qdrant must not be described as GPU search.

The 120-second request timeout is a deadlock fuse, not a latency target. Product performance is accepted per route: tuned p95 must beat the controlled baseline and MRR must be equal or better. The final five-repetition quiet-window run passed all 11 paths. Four concurrent routes completed in 2.87 seconds while a concurrent MCP initialize returned in 8 ms.

GRIFF-Map has a continuous lightweight delta watcher and an hourly full metadata-only reconciliation. The full scan does no content parsing or hashes, publishes atomically after journal replay, and runs High priority on affinity 0xC000.

Event-driven ingest queues CUDA embedding and semantic-edge maintenance while relational identity remains authoritative and Qdrant remains a CPU-backed search projection.