Optimized ELU bounds narrow the intervals; physical limits remain unresolved
The transport output proof remains open. A small follow-up to the direct-ELU pilot narrowed its intervals, but they are still too wide to convert from signed-log training units to finite physical flux bounds.
The previous pilot used a fixed lower tangent for each ELU activation. The new version optimizes the tangent’s slope and includes the limiting constant lower bound. It keeps the endpoint secant above the activation and intersects the optimized result with the original bounds. The model and its input box stay fixed.
| full-box pilot | result |
|---|---|
| member 0 of each transport ensemble | both selected members remain unknown despite interval-width reductions of about 30–275 times against the same-graph fixed-tangent baseline; physical inverse-transform overflow persistsC-272 |
Both processes finished in about 45 seconds. The comparison is against the fixed-tangent calculation on the same graph in the same process, so it isolates the effect of the budgeted optimization. These reductions describe interval widths in training space, rather than improved surrogate accuracy.
The runner now saves intermediate activation-input ranges. Their widths reach about ten billion in the ITER baseline calculation and one billion in MAST-U. Those are relaxation bounds: they do not show that the model reaches such activations. They give the next experiment a concrete target—tighten the intermediate intervals before another member-level pilot.
The analytic and optimizer checks, parity gates and existing bound helpers pass (50 focused checks pass normally and with Tier B disabled; repository lint, formatting and typing of all 91 source modules passC-273). An independent saved-output audit checks the frozen source, models, full boxes, interval intersections and process results. This is focused validation, with no fresh full-suite claim.
The other 38 members were not attempted. V5 and the physical-bound ceiling remain open; the guards receive no new output domain. Transport robustness still needs a certified physical range. EC robustness and preparation of the EC reference comparison can proceed independently. The limited MAST-U reference scope remains as recorded by the owner.