Open evaluation artifact
Brali Bench
A portable 50-case suite for checking whether practical-knowledge retrieval preserves relevance, provenance, evidence boundaries, and deliberate no-answer behavior.
Boundary: this evaluates Brali retrieval and grounded packets. It is not an unpinned language-model benchmark.
Current checked report
- 50/50 cases pass.
- Structured Topic hit rate: 1.
- Evidence Decision recall: 1.
- Safety/no-answer pass rate: 1.
- Unsupported evidence claims: 0.
Reproduce
git clone https://github.com/Brali-LifeOS/brali-lifeos.github.io.git
cd brali-lifeos.github.io
npm run build
npm run evaluate:checkPortable files
Manifest · Cases · Results · Methodology
Comparison layers
no-knowledge-control— Control showing what is available without Brali knowledge.lexical-brali— Keyword-oriented Brali retrieval baseline.structured-brali— Ontology-, protocol-, and Evidence-Decision-aware Brali retrieval.
Reuse correctly
Pin a data-v* release for research or integration claims. Preserve the Brali dataset version, source cases, methodology, and trust boundaries when publishing derived results.