The receiver has no catalogue of possible pasts. It learns to choose three noisy reads, retains their answers, then predicts new questions about the original memory. This page runs the actual trained weights.
This past is shown for explanation. It is absent from the receiver inputs. The initial uniform sum is zero for every history.
Twelve leaky traces retain the past. The receiver receives one initial noisy sum and three further noisy scalar answers. The gate commands come from its own state.
These dense signed questions were not the training question family. The gate selector never receives them. Change the question: the acquired readings and receiver state stay the same.
Original retained state Predicted original state. The original values are scoring targets, never receiver inputs.
The receiver predicts twelve coordinates; new linear questions read this prediction. Three scalar answers cannot generally reconstruct an arbitrary twelve-dimensional state. The history distribution and trace geometry make this a structured problem.
512 shared episodes per condition · three independent trained models per approach · three new reads · no test-time learning. Error bars below resample episodes conditional on these three trained models.
| Receiver / control | Answer NRMSE ↓ | Full objective ↓ | State change |
|---|
NRMSE is relative to predicting zero: 1 is that baseline. The full objective penalizes future-answer error and original-state displacement. The query-count penalty is constant because every approach gets exactly three new readings.
Frozen protocol · Outcome ledger and limits · Receipt · All selected weights and training traces
Training used an exact differentiable simulator and supervised future-answer labels. The sender is a stationary linear trace memory; the receiver is a small tanh network. No biological spike waveform, learned physical law or language-model replacement is established.
The matched recurrent receiver has 941 parameters against adaptive's 944, about 5% fewer neural MACs and three more persistent values. Adaptive commands twelve coefficients per read, whereas a fixed schedule can be shared beforehand. This is a prediction result under comparable resources, not a communication or energy win.
The adaptive head also changes the receiver's nonlinear computation through its own gate command. This comparison measures the implemented interface as a whole; it does not isolate every benefit of physical gate placement from that additional computation.
The read changes memory without mathematically erasing a dimension at η = 0.25. Erasure is a separate control. The original ecg.json feedback accident, AnttisBrain2's moons and BrainAsInverseModeler supply the question's history; their equations remain distinct.