This page measures, it does not replay. The run below is real: the golden set was
put to
shopfloor, an on-prem documentation assistant, and every answer and
retrieved chunk was recorded. What your browser is doing right now is the other half —
running the same deterministic checks and the same retrieval metrics as the CLI,
ported line by line from numbers.py, checks.py and
metrics.py. Nothing here asks a model to judge another model.
checking the port against the harness…
Retrieval and grounding, by category move k and the table recomputes
| category | n | recall@k | MRR | prec@k | grounded | abstention |
|---|
top-5 retrieved