Recorded by commits ee25fe3 through 63babb0
Experiment Note
The profile-aware cavity controller was corrected so the Ghia reference stayed evaluation-only. A residual-only cavity control path was strengthened. A streamfunction cavity formulation was then implemented and reverted on the same day. The code history does not record a prose explanation for the revert, but the later documentation makes the likely scientific reason clear: changing the physics parameterization changes derivative order, cost, and numerical behavior, so it cannot be mixed invisibly into a controller comparison.
Cavity-safe acceptance logic was improved after these trials. The controller needed stronger checks that an intervention did not buy a local improvement by damaging overall flow consistency.
Technical Essay
This was the first clear encounter with reference leakage. Ghia profiles were useful for answering, "Did the learned centerline match a trusted benchmark?" They were not acceptable for answering, "Should the controller take this next action?" The controller had to make decisions from quantities available during training: PDE, continuity, momentum, boundary, or permitted sparse-data losses.
The streamfunction experiment raised another fairness issue. In two dimensions, u = psi_y and v = -psi_x enforce incompressibility by construction, which is attractive for cavity flow. However, Navier-Stokes residuals then require higher derivatives of psi. Better results could come from the formulation rather than from VARA. The eventual design kept shared physics formulations as separately reported overlays or ablations, applied equally to every method.
Reading Note
The relevant hard-constraint literature showed that trial functions can encode boundary or divergence conditions exactly. The NSFnets/streamfunction line of work motivated hard incompressibility, while exact boundary-construction work motivated output transforms. The key note was not that one formulation is always superior, but that architecture/formulation gains must be distinguished from controller gains.
Progress Update
The project gained an important negative result: a technically appealing formulation was not automatically acceptable evidence. The experimental protocol became stricter: reference data for evaluation only, controller decisions based on reference-free diagnostics, and formulation changes isolated from the core VARA claim.