Both reference backends reproduce their own outputs
A bounded diagnostic evaluated each selected local state twice in each solver backend, with unchanged inputs and a fresh process for every call. Every backend reproduced its own signed species energy flux exactly. The earlier differences between backends persisted, including the largest discrepancy in a control state.
| evidence | result |
|---|---|
| saved-output repeatability audit | 521 saved-output checks pass; every within-backend fresh-process repeat is exactly equal and agrees with its consumed output, while all four cross-backend discrepancies persist; native accounting completes within the reviewed bound and the offline archive copy matches its checksumC-201 |
| reference physics and coupled acceptance | remain unresolved |
This narrows the numerical investigation. The discrepancy is reproducible, so another identical collection would supply little new evidence. It still does not identify whether the difference comes from build arithmetic, internal inputs or eigenvalue search behavior. Numerical refinement and implementation comparison are needed before choosing a replacement reference.
The archive tracking correction also worked in the completed batch job. The checksum-bound archive preserves relative paths for every replicate; its one offline tracking copy matches the original. Earlier failed attempts stay failed and retained. This new successful diagnostic is also consumed and will not be submitted again.
The literature-based signed reference specification stays conditional. Exact repeatability is distinct from numerical convergence, physical validity and the temperature acceptance gate. The current surrogate and guard remain selected, and the uncovered-face ceiling stays open.
What’s next. Review numerical refinement and a conservative stable energy closure before another coupled reference. No model replacement or artifact synchronization occurred.
Technical details → Verification › reference repeatability
Decisions → OQ-33 · OQ-34