Three scores. One opinion.
Every device in the registry is graded on three orthogonal axes. The Overall Score is a weighted blend (SCS 0.45 · ARS 0.35 · TLS 0.20). Weights are versioned and shown on each device page.
SCS
Spark Capability Score
Raw AI + compute capability per device. Weighted toward tokens/sec on representative open models.
- AI throughput (tok/s, Q4 70B + Q5 13B)50%
- CPU build & compile (Clang, Rust LTO)30%
- GPU rendering & RTX features (DLSS, RT)20%
ARS
Agent Readiness Score
How well a device runs as a 24/7 local agent host: thermals, stability, sleep/resume, battery under mixed load.
- 72-hour Agent Stress Suite (crash-free)40%
- Sustained thermals & throttle headroom25%
- Battery under mixed agent load20%
- Wake/sleep with resident agents15%
TLS
Trust & Lifecycle Score
Governance and longevity. Driver cadence, OS compatibility, security posture, OEM track record.
- Firmware & driver update cadence30%
- Agent stack compatibility (OpenShell, llama.cpp, vLLM)25%
- Security posture (TPM, secure boot, telemetry)25%
- OEM long-term support history20%
Agent Stress Suite v1
A 72-hour mixed-agent workload that we run on every device under test. Anything that crashes, throttles, or refuses to resume is penalized.
Agent A — Coder
Qwen3-Coder-32B + Aider
Hourly refactor loop on a 250k LOC monorepo.
Agent B — Researcher
Llama-3.1-70B + Letta
Continuous web research with a 32k context.
Agent C — Meeting
Qwen3-7B + Whisper-large-v3
Live transcription + summary every 15 min.