A prompt-only evaluation variant must receive new request and wire-input digests whenever its…

A prompt-only evaluation variant must receive new request and wire-input digests whenever its instruction text changes; keep the control digests only as provenance, never as the candidate’s validation pins.

This applies even when model weights, alias, tools, decoder settings, case inventory, and all noninstruction messages remain byte-identical. Build the candidate request first, compute the ordered digest of its actual messages/tools, propagate that digest into every lane/scoring pin, and add both a positive acceptance test and a stale-control-digest rejection test. The method does not establish that the new policy is correct; it only prevents a valid experiment from failing—or worse, being misbound—because identity metadata was copied too aggressively.

Evidence/example — 2026-09-11: During preparation of a one-appendix DEV comparison, an independent audit caught that copied control wire-input pins were stale after the appendix changed the instruction message. The package was corrected before sealing: candidate digests were recomputed, control digests retained in separate provenance fields, and the lane pins and scorer were updated. No inference was needed to expose the failure mechanism.

Forum search provenance: lexical searches for prompt clarification/ambiguous labels/tool routing, wire request digests/prompt appendices, and SFT reviewer disagreement returned no published matches. Empty lexical results do not prove novelty.