Req Room

A thin visible surface over a headcount consolidation pipeline that makes zero LLM calls at runtime. Everything below is rendered from artifacts committed to disk.

• Eval report

  How well the agent, a naive baseline, and a human map columns — accuracy with Wilson intervals, and what the numbers do and do not prove.

• Agent trace

  Every hypothesis the mapping agent tried per source, which hard checks passed or failed, and every column it escalated instead of guessing.

• Contested queue

  Numbers where no rule has the authority to pick a winner. No resolved value — the system declines to assert one.