Recorded by commits 66d8622 through 9f92e10
Experiment Note
A guarded final physics-repair phase was added to recover physics consistency after adaptive training. A separate collateral-damage guard was tried and then reverted, indicating that the first form either duplicated safeguards or did not behave reliably enough to retain. On 4 June, a residual-adaptive-refinement (RAR) ablation baseline was added.
Technical Essay
Adaptive allocation can optimize reconstruction while leaving a residual or boundary term slightly worse. A final physics repair is a plausible remedy, but it creates a second intervention whose cost and side effects must also be guarded. The scientifically safe design is to compare raw, unweighted metrics before and after repair, restore the previous checkpoint if a protected metric degrades beyond tolerance, and exclude the repair from exact equal-step studies unless every method receives the same phase.
RAR was the necessary baseline because it isolates the question: is the gain just from moving collocation points toward large residuals, or does retaining variable identity, regional persistence, action type, and rollback add value?
Reading Note
Residual-based adaptive refinement is one of the closest precedents to VARA. Its strength is spatial targeting; its limitation for this research question is that a scalar residual does not preserve which physical field or objective is failing. This helped define mechanism ablations: sampling only, local loss only, regional but no variable awareness, unguarded control, and full VARA.
Progress Update
The repository now had both a post-training safeguard and a direct adaptive- sampling comparator. The project was transitioning from "Does VARA work?" to "Which part works, under what compute definition, and at what cost?"