V6 · VALID PAIRED CONTINUATION
Raise the pressure-solver cap
The cap constrained some inner pressure solves, but removing that cap did not materially improve the coupled target residuals; all three target gates remained failed.
AERO / HARCC S4 METHANE COOLING / PART II
CFD NUMERICAL FAILThe solver ran. The CFD did not pass.
A bounded investigation tested whether OpenFOAM solver controls explained the frozen v5 residual failure. One valid intervention helped only some fields; a configuration mistake was excluded; the stop rule ended the branch without turning a mixed result into a pass.
v6 pressure-cap test · v7 invalid test disclosed · v8 corrected paired test · no v9
Disposition
Investigation: complete for the bounded solver-control branch. Numerical credibility: not established. The v8 two-corrector treatment lowered the final `h` and solid-`e` residuals but raised `U`; all three target residual criteria still failed.
This is diagnostic evidence, not experimental validation. Energy closure, mesh independence, wall-treatment suitability and design readiness remain open.
The frozen failure that started the investigation
OpenFOAM Foundation 10 chtMultiRegionFoam reached pseudo-iteration 5000 and ended normally. Normal process completion did not satisfy the pre-existing residual limits.
| Equation | v5 residual | Frozen limit | Gate |
|---|---|---|---|
| U | 1.1485e-4 | 1e-5 | FAIL |
| h | 6.7264e-5 | 1e-5 | FAIL |
| solid e | 1.8235e-4 | 1e-6 | FAIL |
| p_rgh | 1.1583e-5 | 1e-4 | PASS |
| k | Below limit | 1e-4 | PASS |
| omega | Below limit | 1e-4 | PASS |
The case was a 70,400-cell reconstructed single-pitch segment, not a complete as-built assembly. Interface y+ ranged 36.2–289.8 (mean 100.6); no wall-treatment suitability gate had been established.
Three attempts, with the invalid one kept visible
V7 · INVALID FOR HYPOTHESIS INFERENCE
Catch a setup mistake
The treatment value was written to a region-local dictionary the solver did not use. Both runtime logs reported one outer corrector. The error was detected and the results were excluded from inference.
V8 · VALID PAIRED CONTINUATION
Correct the outer-corrector test
One versus two global outer correctors, both continued from iteration 5000 to 5200. Runtime settings were checked in each solver log.
v8: partial movement is not convergence
Final residuals at iteration 5200. Positive change means the treatment residual is higher than control.
| Field | 1 outer corrector | 2 outer correctors | Change | Both gates |
|---|---|---|---|---|
| U | 1.0619e-4 | 1.0755e-4 | +1.29% | FAIL |
| h | 6.6143e-5 | 6.5062e-5 | -1.64% | FAIL |
| solid e | 1.7854e-4 | 1.7480e-4 | -2.10% | FAIL |
| p_rgh | 1.1285e-5 | 1.1034e-5 | -2.22% | PASS |
| k | 4.7785e-5 | 9.5225e-5 | +99.28% | PASS |
| omega | 3.7510e-6 | 7.4665e-6 | +99.06% | PASS |
More solver work
400 → 800 pressure solves. Runtime rose from 69 s to 116.4 s.
The extra work is not evidence of greater physical accuracy.
The stop rule held
Since improvement was not consistent across `U`, `h` and solid `e`, this solver-control branch stopped at v8. No v9 was launched.
What this does and does not show
- Shown: the pressure iteration cap affected inner solves; its removal did not materially repair the coupled target residuals.
- Shown: the v7 treatment setting was not read by the solver, and its branch comparison was correctly excluded.
- Shown: the corrected v8 intervention changed `h` and `e` favorably but `U` unfavorably; frozen gates remained unchanged and failed.
- Not shown: a unique failure cause, mesh independence, energy closure, matched experimental agreement, physical validation or design readiness.
The heat-input prose and facewise boundary field conflict; native author CAD and complete assembly geometry are unavailable in this reconstruction. These uncertainties were retained, not resolved after the runs.
Reference boundary and model comparison
The initial GPT-5.6 Luna diagnosis used a clean, read-only context with publication results and the retrospective Part-I report withheld. It could see the original case evidence and a methods-only source extract. That is reference-withheld diagnosis, not a fully hermetic blind benchmark. Follow-up v6–v8 reviews were sequential in the same Luna context.
The fresh GPT-5.6 Sol comparison is now complete. Sol received only the original v5 packet and criteria, without Luna’s hypotheses or later outcomes. It proposed an unchanged, fixed-length diagnostic continuation; no Sol-selected solver experiment was run.
Retained public run evidence
The v8 solver logs are byte-preserved and their SHA-256 values are listed in the results JSON. The complete mesh and field/checkpoint archive remains in local research storage; this public release is not a turnkey rerun package.
Download control solver logDownload treatment solver logResults + hashes (JSON)Full public run noteSource publication DOI
Artifact hashes establish identity, not scientific validity. No full-text paper, figures or experimental tables are reproduced.