Eight writable walls make 256 possible records. Instead of measuring every record to train the listener, we select a limited, causally plausible set of wall combinations and reconstruct the others from their interactions. A second strategy uses the same number of calibration states but chooses by interaction order.
The path estimates search shortest tours among six representative boundary cells per sector, from a source to all required patches and then the fixed receiver. A 40-step safety margin comes from earlier K=8 observations; it is not a rigorous support bound. The physical wave simulator is independent of this tour calculation.
| Calibration interaction order | Geometry-first | Degree-first |
|---|---|---|
| Waiting for the simulation. | ||
Three independent Gaussian-noise trials per unseen record. For comparisons at zero noise, the noiseless exact counts are also shown. The control uses interaction order first, then geometric arrival within each order. All calibration states are restricted by the same causal eligibility cutoff.
The actual waveform is used only here, to evaluate the two predictions made from their restricted training sets.
The plotted source is the first chosen physical source, with 100 samples from the selected window. Decoding uses all the sources selected above.
Success: geometry picks a smaller calibrating family with equal or better held-out accuracy, or identifies a physically grounded rule predicting which interactions must be modeled before hearing them. Failure: geometry loses to degree-first at equal calibration and acquisition budget. Both outcomes teach us about this cavity's readable memory.
There is no claim of novel wave physics, learning wall plasticity, polynomial-time general memory, or transfer of OpenAI's 2026 Euclidean local-smoothing theorem. Calibration is exhaustive over the subset of records selected, and the full 256-state oracle is an expensive diagnostic used to evaluate failure honestly. The page reports a single deterministic noise seed, not a general statistical superiority result.
Other necessary experiments: lower-memory geometry selection as K grows, independent geometries and grid sizes, instrument-energy matching, uncertainty calibration, and a local physical write rule. The previous four-bit live reader remains available as a separate experiment.