Take the rows unchanged
Use all 56 rows from the repository. Preserve each query, language, and as_of value. Do not tune on hidden knowledge or replace the historical cutoff.
Open evaluation set / 56 rows
A compact public benchmark for time-bounded historical answers: 26 honest unknowns and 30 grounded cases, derived from the rights-clean public pack.
Open github.com/datelock/benchmarkUse all 56 rows from the repository. Preserve each query, language, and as_of value. Do not tune on hidden knowledge or replace the historical cutoff.
Produce one answer per row and retain the evidence your system used. UNKNOWN is a valid outcome; unsupported confidence is not.
Follow the benchmark repository's scoring instructions and report the exact revision, configuration, per-row outputs, and aggregate counts. Compare grounded and UNKNOWN rows separately; never convert a failed verification into credit.
Reference release
5dd4c251