Strict physical residuals and continuous transport
We needed the ITER transport trajectories to solve a consistent discrete equation before their time-step comparison could support uncertainty analysis. The primer explains why the neural transport model operates behind a guard and a physics fallback.
Earlier trajectories could finish with finite outputs and a zero solver flag while retaining substantial physical residuals. Continuing the same iteration, adding damping and learning a damping choice did not establish a remedy. A discrete operator audit also found that stabilization changed the equation being iterated. A small linearized update was insufficient evidence of a physical root.
We changed the experimental integration in three places. A strict Newton solve uses the unstabilized native residual and retains the previous iterate if its line search fails. A continuous transition blends the admissible surrogate with fallback near guard boundaries and smooths the particle diffusion/convection switch. An invalid proposed time step restores the complete previous state and auxiliary data before trying a smaller step (the new research baseline uses a continuous guarded closure, Newton trials accepted only after residual decrease, and adaptive rollback of the complete state and auxiliary dataC-304).
The continuous transition changes the transport closure. Its passing result belongs to this new baseline. It does not make the old hard-switch equations converged, and the historical failed runs stay failed. Existing production defaults are unchanged; the validated acceptance protocol is carried by frozen research adapters.
The final adapters check channelwise physical residuals, the exact native residual, global balances, positivity and finiteness before accepting a step. An explicit crash stops this study. The residual gates pass across the complete ensemble (maximum accepted physical-channel L1 is 1.76 × 10−9 against a 10−8 gate, and native mean L1 is 1.01 × 10−14 against 10−10; no crash occursC-307). Failed trials remain visible: 372,141 accepted proposals and 2,659 rejected proposals are retained in 23,100 raw chunks; rejected-step rollback and retry schedules pass reconstructionC-308.
An independent NumPy auditor reconstructs the physical defects from retained cell terms and checks the actual retry schedule, raw profile continuity and reporting endpoints. It shares the declared discrete equations with the solver; it is not a second independent plasma model. Complete auxiliary rollback is additionally checked in the source and synthetic controls.
The accompanying full-ensemble result reports both adjacent maximum-step comparisons. M5 now has a passing numerical baseline for this closure. Physical-model review and any later uncertainty study remain separate work.
Technical details → continuous closure and rejected-step rollback.