Evidence layer·Sharpen
Reranking
The reranker (`BAAI/bge-reranker-v2-m3`) sharpens evidence ordering after semantic retrieval and before narrative generation. Bounded by design — it reorders context; it does not change candidate scores.
Where reranking runs
Reranking can be invoked before:
- candidate explanation generation
- computational assessment paper generation
- failure-aware analysis
- comparative context synthesis
The reranker only changes evidence ordering. It does not change candidate scores, validation outcomes, or bounded-feedback rules.
The bounds
Reranking is intentionally capped:
- only a limited number of retrieved snippets are reranked
- each snippet is truncated to a bounded character window
- model failures fall back to deterministic lexical scoring
- overflow snippets are never allowed to create unbounded work
semantic retrieval
-> bounded candidate set
-> BGE reranker when available
-> lexical fallback when unavailable
-> selected evidence contextArtifacts
Papers and explanations can include rerank method and relevance scores in supplemental replay artifacts when reviewed for public release. The public-facing narrative remains scientific and readable; raw model internals stay in private operator state.