Fixed-state capture compares the two panes at the same seed, scripted inputs, simulation time and renderer. Open the deterministic 3.6 s fixed-state comparison → · capture contract and current-record location. Live iframes are exploratory and use independent browser clocks. Only difference in the capture: B-control pads feel a spring to rest (-K·z) instead of reciprocal -K·(z-η̄); water still receives pad pressure in both.
B-control (one-way, no water→pad feedback)
B-coupled (adapted reciprocal coupling)
What this demonstrates
The adapted Countercurrent linear shallow-water / C2-Wendland reciprocal coupling (Apache-2.0) recontextualized as a 7-pad visual instrument, extended to N bodies and a circular reflecting basin.
Under fixed capture conditions, enabling reciprocal water→pad feedback changes the image and pad trajectories. The live panes do not enforce shared timing; see EVIDENCE.md for measured capture numbers.
CPU tests: N-body energy/volume conservation, pad-to-pad reciprocity, mutant-must-fail checks.
What this does not demonstrate
That reused code is universally better, faster, or necessary.
Real-GPU performance (capture ran headless on SwiftShader software WebGPU).
The upstream GPU compute path (the physics here runs on the CPU port; WebGPU is used for rendering).
Physical accuracy beyond the linear shallow-water model.