Reranking

Protean uses BAAI/bge-reranker-v2-m3 as a bounded local reranking route for retrieved evidence snippets. The reranker sharpens context selection after semantic retrieval and before narrative generation.

Where It Runs

Reranking can be used 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 learning rules.

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 context

Artifact Role

Papers and explanations can include rerank method and relevance scores in manifests. Public-facing narrative remains scientific and readable rather than exposing raw model internals.