Did the memory learn a better way to ask itself questions?
We replaced the hard-coded router with a trainable phase writer, an independently trainable query-site correction and a query-conditioned operator-order policy. Then we challenged them with the exact same initial network whose router was frozen. The resulting failure is more informative than a win against an intentionally blind decoder.
01 / Held-out memory retrieval
Every episode is a fresh random key/bit stream. The query names one previously written key. Classify its most recent bit. Results are test accuracy, not pixel reconstruction, language modelling or neural physics. The pilot was exploratory, not preregistered.
02 / The learned computation
learned phase writer→8 × 8 byte phase field
64 B evolving state→query-selected source
two ordered port reads
The query selects a hard launch site using a straight-through softmax approximation. It also learns an AB/BA preference. Paths themselves remain prescribed 12-step grid routes, and the selected source is chosen from a fixed known 128-key vocabulary.
A trainable address table can change its assignments without discovering a better geometry. The real test is whether it outperforms its frozen, identically initialized version. In this pilot, it does not.
03 / What the ratio actually measures
z_AB = exp(i · (source + common))
z_BA = exp(i · (common + port))
ratio = z_AB · conj(z_BA)
= exp(i · (source − port))The port is fixed at zero phase, so the exact ratio reveals the directly addressed source phase. We measured discrepancies below about 1.1 × 10⁻⁶ on actual trained states. The direct-site reader solves the same information problem without two long paths. A successful ratio is therefore not a novel memory-capacity result.
Compute qualification
This PyTorch implementation evaluates 63 candidate paths before selecting one. A hardware implementation could route a single selected path, but that still costs two 12-step transmissions rather than one direct-site access.
04 / Evidence of actual routing changes
Assignments do change during learning. But changed does not mean useful. Here are the three recorded pilot seeds.
05 / Reproduce, challenge, extend
python -m pip install -r requirements-gate1.txt python -m unittest discover -s tests -v python train_gate1.py --smoke --device cpu python train_gate1.py --steps 350 --batch 64 --length 24 \ --eval-cases 512 --seeds 41 97 127 --device auto
Next falsifier: replace lookup-table routing with a geometry inferred from within-episode experience, add actual observation-dependent probe selection, and compare with a complete small causal Transformer on matched computation. Also retain the exact indexed baseline rather than comparing only to weak models. Full protocol and limitations ↗