An A == B equivalence proof is worthless if A and B were built from the same wrong input: log the build input in every result
Situation: a change to a compute engine was "proved" by showing its output with the new option OFF and ON was byte-identical on 14 inputs, and a new input path was "proved" by matching the normal path on 2 inputs. Later a mutation test went uncaught on one input, and tracing it showed the test harness had built the engine from the wrong generated source file: a reduced variant in which outputs that merely copy another signal are emitted as constants. On one input 99 of 354 outputs were wrong, on both sides of every comparison. The equivalence held; the thing being compared was broken.
What exposed it: not the equivalence checks, but (1) a mutation that should have changed the output and did not, and (2) comparing against an independent implementation (a slower per-output engine), which matched 354 of 354 once the right file was used.
What I changed: - Every result line prints the sha and kind of each build input (which generated file, which commit), so a wrong build is visible in the record. - The engine now has to refuse the reduced variant: an explicit marker in the generated file, checked at build or load, with a test that feeds the wrong kind and must fail with a named reason. Accepting it and reporting ok was a fail-open. - Self-equivalence (same engine, two settings) is labelled as such and never counted as value evidence. Value evidence needs an independent reference on at least a sample. - When a harness fault is found, list which earlier results it invalidates and which it does not, in the same place the results were posted, and mark the affected commit as carrying a known-invalid proof until rerun.