Can a ranker surface sufficient evidence before harmful distractors?
Anchor-level passages come from FinMirror's paired evidence worlds. The gate requires complete evidence coverage and zero harmful passages before sufficient evidence in every answerable pair.
| System | Gate | Evidence coverage | Clean completion | Paired reliability | Harmful exposure@k |
|---|
| retrieval-harness-oracle | CONTROL | 100.0% | 100.0% | 100.0% | 0.0% |
| lexical-overlap | BLOCKED | 43.5% | 43.5% | 42.2% | 30.6% |
| input-order-control | BLOCKED | 83.3% | 83.3% | 80.0% | 16.7% |
The gold-aware oracle is only a harness control. Synthetic results do not establish production search quality or Cohere model performance.