Recorded by commits 1fa1848 through b3f8762
Experiment Note
The first repeated experiment was consolidated into five-seed Kovasznay raw and summary CSVs. The later manuscript reports that VARA reduced the five-seed means across the analytical quantities by approximately 33-46 percent, including velocity components, pressure, vorticity, residual, and objective metrics. This was the first evidence that the controller effect was not a single lucky run.
The benchmark surface was expanded with lightweight rectangular manufactured flows, boundary-condition stress tests, channel inflow/outflow, double-vortex flow, aspect-ratio sweeps, and lid-driven cavity support. Several metric bugs or ambiguities appeared immediately. Relative error is ill-defined when a reference field is identically or nearly zero. Rectangular reference metrics also needed correction. Guardrails were added so invalid denominators or absent reference channels could not silently create persuasive-looking numbers.
Ghia et al. centerline data were wired into the lid-driven-cavity evaluator. An early profile-aware cavity control path was also attempted.
Technical Essay
This phase exposed the difference between an analytical benchmark and a reference-backed engineering benchmark. In Kovasznay flow, the exact field can support full-field relative errors. In the lid-driven cavity, Ghia profiles are sparse centerline evaluation data rather than a full training field. Treating them as interchangeable would blur the scientific claim.
Metric robustness became part of the method, not just reporting polish. A zero reference norm must lead to a guarded alternative such as absolute RMSE, MAE, or an explicitly unavailable relative error. Similarly, a controller should not gain access to a test reference simply because the evaluator already knows it. This separation later became a hard V2 invariant.
Reading Note
The cavity work drew on the classical Ghia et al. benchmark for centerline velocity profiles and on later cavity literature for vortex topology and higher-Reynolds interpretation. The reading lesson was methodological: a known benchmark profile is valuable for evaluation, but using it inside adaptation would turn a physics-informed training claim into reference-guided tuning.
The rectangular tests also reinforced the importance of manufactured solutions. They allow exact forcing, boundary conditions, and reference fields to be written down, which is ideal for debugging whether an apparent controller gain is real or an artifact of metric construction.
Progress Update
Kovasznay supplied the first multi-seed success. The repository began moving from a single demonstration to a professor-facing benchmark suite. At the same time, the work uncovered a major future constraint: controller inputs, training signals, and final evaluation references had to be explicitly separated.