Who judges the shore?
A passive wave changes its bed. An agent can instead try a small material change, measure what happened at the harbour, and undo the change if it was worse. Run 200 candidate changes for each of eight fixed beds, and see how strongly the proposal's shape matters.
Benchmark controls
Fixed seeds: 1001, 1009, 1013, 1019, 1021, 1031, 1033, 1039. Both targets · five proposal families · 200 trials each. Click Run.
Objective = receiver power |u|²: quiet means lower power (negative dB); loud means higher power (positive dB). Each candidate is actually evaluated using the changed wave operator. Improvement is monotonic under the acceptance rule by construction; that is an agent property, not a physical-learning law.
Median change in receiver power across eight beds
Quiet harbour: lower is better ↓
Loud harbour: higher is better ↑
Objective, reversibility and computation costs
| Material proposal | Quiet median | Loud median | Accepted Q / L | Receiver reads + proposal solves (per task, 8 beds) |
|---|---|---|---|---|
| Run the benchmark to see measured results. | ||||
Receiver reads are 201 × 8 per task/shape. Storm proposals also require 200 × 8 separate wave solves to construct their shapes; others do not. These counts exclude display rendering and the externally supplied acceptance/reversion logic. A conservation rule is imposed by subtracting each proposal's mean before the trial; this does not represent natural mass flux.
What this tests—and what it cannot tell us
The test directly addresses Claude's new G8 result: trial learning can work with loss because it evaluates performance after changing the medium, instead of asking a dissipative backward wave to carry an adjoint gradient. But the system gets a privileged agent, a receiver objective, and perfect restoration of rejected material changes. The original Memory Drift lab separately compares a privileged, bed-frozen stress reset against an active teacher-free rest that itself moves the bed.
Read the protocol in SITE_G9_PROTOCOL.md. The eight seeds and candidate families are frozen for reproducibility, and all results are synthetic browser calculations. They must not be described as a direct replication of ComputingBeachClaude or as evidence of autonomous sediment optimization.