When memory goes stale.
A wave changes its medium. Its old idea of “normal stress” may no longer be true. Run twenty teaching episodes, watch an internal baseline drift, and compare a frozen memory with a rapidly adapting one.
Run a wave-memory experiment
Choose the settings, then run.
Each arm begins from exactly the same initial bed and target. Evaluation applies the same maximum per-node material step to every nonzero update; this is an imposed control, not spontaneous physical learning.
What the receiver actually observes
Receiver loss, relative to epoch zero ↓
Update alignment with exact descent ↑
Material profiles at the selected epoch
All arms begin with the same rough material.
| Writer | Final loss / initial | First gradient cos | Epoch-10 gradient cos |
|---|---|---|---|
| Run the experiment to populate the measured controls. | |||
The “two-exposure subtraction” control obtains the cross term by subtracting two separately calculated field energies in software. The privileged reset arm exactly replaces θ after an extra forward-only solve, without moving material. The new active-rest arm allows teacher-free exposure to move material, then adapts θ toward the new bed's forward energy at T. This requires a separately scheduled teacher/rest gate and at least two additional wave measurements per epoch. These branches do not receive identical material-update budgets and must not be mistaken for autonomous sand learning.
Claude's objection, checked beyond a single seed. Runs eight fixed rough beds (1001, 1009, 1013, 1019, 1021, 1031, 1033, 1039), for both quiet and loud targets, with the wave loss, teacher strength, step and memory timescale currently selected above.
No batch has run yet. The rest branch can move material under the teacher-free wave; a privileged reset cannot.
| Writer | Quiet power dB ↓ | Loud power dB ↑ | Median relative task loss · Q / L |
|---|---|---|---|
| Run the eight-bed audit to see measured medians. | |||
The loud task's complex-amplitude target and loudness in dB are different observables. The active-rest arm gets an additional wave exposure, another material update, and an externally scheduled teacher-off period. This is not matched-resource evidence of passive plasticity. For learning with a separate global judge, see the Agent Trials lab.
What the numerical instrument actually solves
A complex frequency-domain wave operator couples neighboring sites: A(h,ω,γ)u=F, with a tridiagonal spatial stencil and local material-dependent coefficient. The source is node 3, the receiver node 44. The teacher is injected at the receiver with the complex output error as its amplitude. Damping changes both amplitude and phase. Every model solve uses the actual current material state.
Each patch stores θᵢ, its old ordinary field energy. A single teaching exposure gives rᵢ = |uᵢ + τᵢ|² − θᵢ. The local G8 edge relay exchanges neighbor differences to make a zero-net-material update; it does not access a global mean. The control named “exact gradient” uses the analytic wave sensitivity and is not an autonomous material rule.
Claude's G8 warning: A recalibration that changes θ but forbids bed motion has an external privilege. The active-rest branch allows a second material update during a teacher-free exposure, then a T-rate baseline update. It needs an externally scheduled teacher-off gate, and may cost extra wave solves and additional material displacement. Compare the two rather than treating the privileged result as a physical prediction.
Read the plots as falsification tools: a positive cosine at epoch one does not imply the next nineteen steps remain aligned. Tracking the teacher can absorb the desired deviation, while never updating θ can cause staleness. Here the material is an effective numerical parameter, not bed depth constrained by an Exner sediment law. The synthetic loss/damping sweep is illustrative and does not reproduce the 2D cavity's measured Q.
For the held-out Python experiments, see G6, G7, and G8. For Claude's independent 20-epoch quiet/loud experiments, see ComputingBeachClaude.