How the judgement works
Methodology, without the incense.
The result is deterministic. The same normalized inputs, dataset version, and methodology version produce the same technical result.
What is scored
Workload uplift contributes 30%, present need 25%, value 20%, workload fit 15%, and concrete longevity signals 10%. Verified compatibility and sidegrade penalties are subtracted afterward.
Current limitation
Market value is not calibrated in dataset 2026.08.6. It is held neutral and every result is capped at low overall confidence. Purchase prices are USD-only; prices within 3% of launch MSRP are treated as approximately MSRP.
Sources and uncertainty
GPU specifications and normalized profiles carry source URLs, retrieval dates, confidence, and dataset versions. Missing critical profiles produce an unsupported result instead of an invented estimate.
Current methodology: gpu-v1.2. See the repository scoring specification for the complete curves and penalties.