Docs / Kraft Index
Kraft Index
The Kraft Index is a weighted composite of category scores derived from KraftBench measurements. It exists so experiments can be compared with one explainable number — without erasing the category breakdown underneath.
Definition
- Each measurement contributes a higher-is-better
score - Within a category, scores are combined with a geometric mean (limits one microbench dominating)
- Categories are blended with documented weights
- Missing categories (e.g. GPU weight 0, or failed storage) cause renormalization over present weights
- A baseline run’s mean raw composite maps to 10,000
- Later runs:
index = (raw / baseline_raw) × 10000
Default weights (M1)
| Category | Weight | Rationale |
|---|---|---|
| CPU | 0.40 | Dominant for general systems work |
| Memory | 0.20 | Bandwidth / latency sensitivity |
| Storage | 0.15 | Local I/O capability |
| System | 0.10 | Scheduler / threading proxies |
| Realtime | 0.15 | Pipeline / parse / parallel stand-ins |
| GPU | 0.00 | Reserved until a real backend exists |
Weights sum to 1.0. GPU is zero so the index is not diluted with placeholders.
CPU0.40
Memory0.20
Storage0.15
Realtime0.15
System0.10
GPU0.00
Illustrative visualization of documented weights — not a live measurement.
What gets stored
Every experiment stores the composite Kraft Index, category scores, per-benchmark components, and raw sample measurement sets. Inspect with:
kraftverk explain <id>
kraftverk history
kraftverk lineage <id>