Docs
Documentation
Descriptive guides for the evidence loop — rewritten for the web from the repository docs. Honest about what works in 0.2 and what remains experimental.
KraftBench
Real workloads, categories, reproducibility, and what we do not claim.
Optimizer
Safe mode, allow-listed parameters, hill climb, validate, commit or revert.
Statistics
Sample summaries, comparison classes, and decision-aid honesty.
Kraft Index
Weighted composite, geometric means, baseline 10,000, renormalization.
Architecture
Crate map, control flow, platform boundary, persistence, privilege.
Safety
Non-goals, guardrails, hardware boundary, thermal honesty.
Telemetry
Local minimal snapshots — never invented benchmark scores.
Hardware support
AMD-only policy, exit codes 20–25, detection, product surfaces.
Roadmap
Completed milestones, planned work, explicit non-roadmap.
Contributing
Proprietary terms, evidence rules, workflow, security reporting.
Also on GitHub
Source documents
Canonical markdown lives in the repository when you need diffs or offline reading: