Accepted.
Numbers without hardware, dataset, and cache state appeared universal and could not be reproduced.
Documentation describes mechanisms and qualified complexity. Numeric baselines
require the manifest from docs/performance/benchmarking.md, raw data, and the
tested commit. Targets are labeled as targets.
- The README publishes no latency or memory figure without attached evidence.
- Comparisons keep dataset, model, lockfile, and cache state constant.
- A local result can guide investigation, but becomes a baseline only after the complete protocol is published.