Write four persistent changes into the boundary of a reflecting ellipse. Let the wave die. Then choose where and when to send a fresh ping. The question is not which echo is strongest, but which ping best distinguishes the records.
Display field = visual aid only. The reader does not get a full-field FFT, virtual directional camera, or access to the hidden wall. The actual sensor is the point at the opposite focus.
Three physically different source positions; two listening windows of exactly 100 samples. Each observation costs one ping. Windows use equal sample counts, not equal elapsed listening time. The receiver stays at the opposite focus.
A seeded noisy recognition test. All methods get the same number of one-cell pings and the same dictionary of all sixteen clean echoes. This tests probe selection, not learning new wall combinations or reducing calibration cost.
Methods: adaptive posterior-weighted separation, fixed focus/late, cycling late sources, and uniformly random source/window.
Known mathematics: finite-difference damped scalar wave propagation on an ellipse, discrete Neumann reflections with binary Robin boundary springs, Bayesian Gaussian template discrimination, and adaptive measurement selection. For each fixed wall the wave evolution is linear; joint wall changes yield non-additive echoes through repeated scattering.
Measured locally by this page: which of six real point-receiver interrogation actions separates sixteen pre-calibrated wall records under a chosen noise and shot budget. Adaptive choice uses the current posterior's expected pairwise waveform separation — a heuristic, not an optimal-information theorem.
Not demonstrated: physical hardware, spontaneous memory writing, generalization to unseen wall combinations, an exponential calibration saving, or a consequence of the October 2026 OpenAI Euclidean local-smoothing theorem. The paper inspired questions about coherent observation, but its free-wave proof does not transfer to this bounded cavity.