SIDScope is a diagnostic tool for evaluating the quality of semantic-ID interfaces in generative recommendation systems. These interfaces connect item tokenizers to recommenders, but often lack information about their structure and performance. SIDScope addresses this by providing a set of tools to normalize and verify these interfaces, profile their structure, and compare revisions. It also accounts for path-to-item outcomes in generated traces. The authors tested SIDScope on nine source-traced tokenizer exports from seven families and found that interface health is multi-signal rather than scalar. They also identified three hidden gaps: prefix alignment strongly tracks held-out candidate exposure when retrieval consumes SID prefixes, but weakens as scoring becomes prefix-independent; a valid target path can survive without uniquely retrieving the target item by 1.2-3.0 percentage points; and repairing the mapping does not by itself restore an inherited generator; model reuse requires a separate handoff check. --- Why it matters: This matters to researchers in AI because it provides a new tool for evaluating the quality of semantic-ID interfaces, which are crucial for generative recommendation systems. SIDScope's findings also highlight the importance of considering multiple signals when assessing interface health and the need for separate checks before model reuse. Source: https://arxiv.org/abs/2608.18779