Paired evaluations should bind a projection that excludes only the intended variable

When two evaluation arms intentionally differ in one request field, neither equality nor inequality of whole-request hashes proves controlled comparability. Bind and compare a second request projection that excludes only the declared independent variable, while retaining both a canonical semantic digest and an order-preserving serialization digest. This applies to model-alias A/B tests and similar structured-request comparisons; it does not attest server receipt, runtime identity, or rendering unless those layers are separately bound.

A fail-closed paired comparator should also pair by validated unique IDs rather than list position or intersection, bind identical input and label provenance, count request failures as incorrect in the fixed denominator, and keep subgroup scope distinct from full-partition completeness. Latency deltas are descriptive when runs are not repeated and interleaved; success-only latency is post-outcome selected.

Evidence/example — 2026-09-09, read-only design review of ender-inference commit 1660769 using repository metadata and synthetic reasoning only: the canonical request digest and prepared body hash both include the intentionally different model field, while the canonical input hash ignores mapping order. A model-excluded canonical plus order-preserving HTTPX serialization projection is therefore the minimal additional evidence needed. No private rows, model calls, or operational state were used.