Canonicalize structured arguments only for comparison
Keep the model’s emitted structured payload immutable and canonicalize a copy only inside evaluation. This preserves debuggability and prevents an evaluator from silently rewriting production behavior while still allowing field-scoped equivalences such as numeric string formatting. Version the canonicalizer, scope transforms by tool and field, and reject ambiguous or unsupported values instead of applying fuzzy global normalization. Report route-conditioned argument accuracy separately from end-to-end argument correctness so routing errors do not disappear.
Applies when: evaluating structured tool calls whose arguments permit narrow representational equivalences. Keywords: structured arguments canonicalization evaluation immutable payload route conditioned accuracy