What does the
platform add?
Most benchmarks score a raw model. We score the difference CroweLM makes: the same question answered with the platform's grounded cultivation knowledge, and again with the bare model. The gap is the product.
How we measure
Public benchmarks
Standard academic tasks (math, multi-domain reasoning, code), scored for accuracy. An honest baseline: this is the floor every tier clears before the part that matters.
Grounded vs bare
Each cultivation question is answered twice: once with CroweLM's proprietary knowledge base retrieving real corpus passages (grounded), once without (bare). A judge scores both on a 0 to 5 scale. The difference, grounded minus bare, is what the platform contributes over the raw model. That number is uniquely Crowe's.
Results
The harness is live and grounding is wired. We publish numbers only from a complete run across the CroweLM tiers, with the knowledge base active. We would rather show nothing than show a partial board. Check back shortly.
Methodology
- Driven through the real CroweLM stack, not a mock: numbers reflect actual routing and providers.
- Track A scores accuracy on public benchmark slices (math, multi-domain reasoning, code).
- Track B answers each cultivation question twice. Grounded retrieves from the CroweLM knowledge base; bare uses the model alone. A judge grades both 0 to 5.
- Tiers are shown by CroweLM name. A tier missing grounded or bare data for a run is excluded, never shown as a fabricated zero.