A testable idea from SihtiResponseGraph and SihtiGrapImageModel: encode experience into a mutable phase medium, then ask questions using ordered probes. First ask what this gains—and what it cannot possibly preserve.
Streaming random facts are written into a fixed-size phase geometry. Retrieval is measured on every fact. Switch between independent bits, correctly grouped structure and deliberately incorrect structural assumptions. These are recorded CPU simulations, not trained networks.
Loading reproducible receipts…
The 64-byte tape contains 32 two-byte records; the indexed bit table exploits a known 256-key universe. The geometry uses 63 phase slots and a fixed reference port; colliding writes overwrite. These are different coding assumptions, shown explicitly rather than disguised as matched neural models.
A single wave mixes target phase and all intermediate phases. Two orders along the same path share that nuisance. Their conjugate product exposes only the endpoint difference.
z₁ = exp(i · (source + path))
z₂ = exp(i · (path + port))
z₁ · conj(z₂) = exp(i · (source − port))
Change the intermediate phase slider. The two raw complex port readings rotate. Their difference rotates too. The phase ratio does not change.
Ideal U(1) transport only. In real lossy, non-reciprocal, multi-path media, exact cancellation need not hold. Both raw measurements jointly contain at least as much information as their ratio.
Two matched unitary paths reject a shared intermediate phase. An existing mathematical identity, but potentially a useful measurement design.
Structure helps if the shared rule is correct. An incorrect geometry can destroy useful facts faster than an ordinary tape.
The new Gate 1 trained pilot includes a learned router but finds no advantage over an identical frozen router, and no case for replacing Transformers.