Restart gaps in SWMM 5.2.4

Teacher-forced (TF) and free-run (FR) maxima are % of the element's own peak. E is the exposure ratio; tail is the share of the free-run difference outside the windows where the change acts; compensated is the share of active element-windows where |G| < ½|L|; τ is the fitted memory time after the change stops acting. Restart null was exactly 0 in every EXTRAN run and at most 7.2×10⁻⁵ of peak on user2 (17 subcatchments).

About

Teacher-Forced SWMM compares two versions of the EPA SWMM 5.2.4 engine the way language models are trained: by restarting engine B from engine A's exact state and measuring one window at a time, then setting that beside the ordinary free-run comparison. The difference between the two measures is what the network does with a change: carries it, amplifies it, or undoes it.

Version 1.0 · October 25, 2026 · SWMM Enablement

Engine

AI use

Sources

• EPA SWMM 5.2.4 source, tag v5.2.4: github.com/USEPA/Stormwater-Management-Model
• Test models: EXTRAN user's manual examples (OWA_EXTRAN folder) and EPA nrtestsuite models from the 1729 corpus: github.com/SWMMBobSWMM6/1729-SWMM5-Models-2030
• Slot momentum variants: EPA SWMM issue 231 discussion (issues/231), as described in the earlier Slot Lab work
• WASI runtime: @bjorn3/browser_wasi_shim 0.4.2 (MIT/Apache-2.0), bundled into this page
• Teacher forcing: Williams and Zipser (1989); exposure bias: Ranzato et al. (2015)