Evaluate Later-Stage Criminal Tasks Under Incomplete Evidence

Determine whether criminal process reconstruction and sentence prediction using the Profiling, Investigation, and Judgment benchmark retain their reported performance when both tasks operate on incomplete evidence and receive only the outputs available in a pure pre-arrest pipeline.

Background

The main evaluation protocol gives criminal profiling only fragmentary scene evidence and witness testimony, while criminal process reconstruction additionally provides the known offender and sentence prediction provides further post-arrest information such as defence arguments, confessions, and charges. Consequently, the reported later-stage scores do not establish how well those tasks would perform when the entire pipeline remains subject to pre-arrest information constraints.

The authors identify this as unresolved by constructing a pure pre-arrest pipeline in which crime-process reconstruction is conditioned only on criminal-profiling outputs and sentence prediction is conditioned on the outputs of both preceding tasks. Their supplementary comparison shows that reconstruction and summarization remain nearly intact, whereas sentence prediction degrades substantially, motivating further determination of whether the main-protocol results generalize to incomplete-evidence settings.