Let the model select source spans; let the host reproduce the quotations

Let the model select source spans; let the host reproduce the quotations.

For evidence-bearing evaluators, separate selecting evidence from reproducing its bytes. Construct immutable passage IDs from authoritative source identity, representation, exact text and offsets; ask the model for IDs; expand quotations in host code and validate the ledger again before accepting them. Keep full policy context available so segmentation does not erase exceptions or conditions. Reject unknown IDs and forged provenance rather than accepting approximate text.

Representation matters: an exact phrase in a decoded JSON leaf can fail against raw escaped JSON, while an exact phrase copied from another supplied source can have the wrong citation reference. Neither mechanism by itself proves invented policy. Preserve the raw source hash and an explicit decoding path. Avoid arbitrary recursive unescaping or fuzzy suffix repair.

Limits: a valid passage ID proves provenance, not relevance, intent, or correct application of a rule. Maintain separate semantic review and chronology checks. Also measure the complete serialized request: readable policy cards plus a host provenance ledger can duplicate substantial input unless the model-facing projection omits redundant metadata losslessly. Smaller bytes are not evidence of lower token billing.

Applies when: A structured evaluator must cite exact policies or source passages while remaining auditable across serialized representations. Keywords: evaluator citations span identifiers JSON provenance exact quotes semantic validation