State

Build the full endgame role architecture now, but test one slice at a time. No important role should be approved by a single model when quality matters.

Active Models
7
GPU 3 active
CPU 4 active
2 cold standby
Quorum Roles
3-of-3
No single-model approval
Testing Slice
Homepage JS
2 roles active
Round robin off
Last Pipeline Run
Pass
Verify OK
Audit OK

Role Assignment + Quorum

One role can map to a 3-model set instead of a single approval point
Role
Primary Set
Reasoning Layer
Timeout
Output Contract
Quorum Gate
Router
Core orchestration · quorum advisory
Planner
Structured planning · independent planning trio
Python Generator
Software engineering · code quorum
C++ Generator
Native systems · compile-first quorum
JS Generator
Frontend interaction · contract-first trio
Verifier
Strict validation · never single point of approval
Auditor
Final review · separate from verifier quorum

Model Runtime Control

Start, stop, warm, unload and pin without scanning the whole page
router-a
pipe_router_gemma4b · GPU · active
Router Family
Pinned GPU
router-b
pipe_router_gemma4b · CPU · active
Router Family
Pin CPU
router-c
pipe_router_gemma4b · cold standby
Router Family
Standby
planner-a
pipe_planner_qwen3_8b · GPU · benchmark
Planner Family
Benchmark
python-a
pipe_python_coder_primary · CPU · cold
Python Family
Cold Standby
cpp-a
pipe_cpp_coder_primary · CPU · cold
C++ Family
Cold Standby