"""Head-to-head bench: deg vs structured-output baselines on the same records and scoring.

* :mod:`deg.bench.sets` - pluggable eval sets with fixed uid manifests;
* :mod:`deg.bench.systems` - deg and the baselines (same-model constrained decoding on HF
  + xgrammar and on llama.cpp + GBNF, NuExtract3, Clef-flash decide, frontier references);
* :mod:`deg.bench.score` - one scoring rule for all systems;
* :mod:`deg.bench.run` - run one system over sets; :mod:`deg.bench.summary` - tables.
"""