v0.28.0

Relevant first.Explain every miss.

Graph popularity cannot lift a weak match above a substantially stronger lexical match. decided diagnose shows why a named decision did or did not appear.

What shipped

  • relationship-graph boosts apply only within lexical near-ties
  • candidates below 85% of the strongest BM25 match receive no graph boost
  • find --explain reports graph_floor_ratio and an applied or clamped gate result
  • decided diagnose distinguishes no match, partial match, filtering, lower rank, and result-window truncation
  • diagnosis uses the same tokeniser, filters, and ranking path as production retrieval

The change keeps retrieval predictable as a decision corpus grows. It also gives authors a deterministic trace when an expected record is missing, without adding a second heuristic or changing the result set.