CITATION EXISTENCE
Can the cited record be resolved and identified?
RDIV / STRATEGY / RESEARCHLIVE ARCHIVE
R/IV / VERIFICATION BENCHMARK
MEASURE THE
QUALITY OF
VERIFICATION.
The current benchmark is a synthetic calibration harness for testing contracts, failure handling, and appropriate abstention. It is not a comparative leaderboard.
SYNTHETIC CALIBRATION · REPRODUCIBLE METHOD · NO CLAIMED SUPERIORITY[ EVALUATION TASKS ]
TEST THE RECORD.
TEST THE LIMITS.
Each case fixes its inputs, expected evidence boundary, scoring rule, and failure condition before execution.
CITATION ENTAILMENT
Does the source support the bounded claim attributed to it?
UNSUPPORTED CLAIMS
Which material statements outrun admitted evidence?
CONTRADICTIONS
Which evidence objects materially conflict?
PROVENANCE
Can each material finding be traced to its admitted record?
ACTION VERIFICATION
Does independent evidence establish that an action occurred?
ABSTENTION
Does the system stop when the record is insufficient?
[ REPRODUCIBILITY MANIFEST ]
VERSION THE TEST.
PRESERVE THE LIMITS.
Benchmark version, task IDs, lawful-source policy, expected behavior, scoring rules, system release, run timestamp, provider/model version, false-positive analysis, false-negative analysis, and abstention behavior are recorded in the machine-readable manifest. The latest run remains null until a valid execution occurs.
[ PUBLICATION RULE ]
CALIBRATION FIRST.
COMPARISON LATER.
Comparative scores will remain unpublished until a frozen, lawful real-world corpus exists and the full run is reproducible. No private holdout material, fabricated result, competitor ranking, or unsupported accuracy claim is published here.