Can a question steer a memory probe?
The two 64-byte writers are frozen. A question launches a unitary point wave into the learned phase medium. Twelve phase-and-shift steps guide it to one of two fixed detector ports. How many scalar measurements are needed to answer?
Results by measurement budget
Change the number of complex detector readings per query. The plot averages two independently trained, frozen writers; individual seed accuracy appears below. Measurement paths were committed geometrically, without label optimization.
Loading the numerical receipts…
Per-seed held-out accuracy
| Reader | Seed 41 | Seed 97 | Mean |
|---|---|---|---|
| Loading… | |||
The bars encode accuracy on a 50–100% linear scale for visual separation; exact numbers and errors are in the published receipts.
The gate result
The question-directed path sees the relevant region. Fixed paths see all regions regardless of the question, and fixed passive ports see only two untransported sites. At four measurements, targeted wave reads average 93.58%, versus 65.31% for question-independent routes. However, four direct local phase samples average 99.02%.
The complex differences of the two phase/shift orders perform relatively poorly. The raw paired order measurements can be much more informative; particularly, with two shots and frozen seed 97, paired orders reach 98.73%. This effect is not matched by seed 41, and reading phase directly still wins.
New: common-path phase cancellation
The exact endpoint ratio is real, but it is not universally a better reader.
For a matched 12-step path, measuring phase→shift and shift→phase yields
z1×conj(z2) = exp(i×(phase(source) − phase(port))).
Intermediate contributions cancel algebraically. The new v3b frozen-checkpoint
test shows 98.58% with two readings on seed 97, but only
59.23% on seed 41. Its identical source-site direct control
scores 99.95% and 62.74%, respectively; the chosen source matters.
The raw paired response scores 98.73% and 78.37%: shared-path phase
can also contain useful information rather than only nuisance.
Explore the new controlled ratio experiment →
What is actually measured
| Reader | What it receives |
|---|---|
| Active phase→shift | K target-selected point waves measured at two ports after 12 steps |
| Active shift→phase | Same paths, reverse local operator order |
| Both orders | K/2 path pairs, two port measurements per path |
| Order difference | Difference of those paired responses, never more information than both |
| Fixed routes | K question-independent wave paths, same fixed ports and depth |
| Direct local phase | K addressable phase sites at the target. A stronger direct-access interface |
| Direct local amplitude | The same phase sites as invertible amplitude values; no waves |
| Passive ports | Repeated direct phase values at the two fixed ports; intentionally weak |
A single active shot traverses 12 phase cells and returns one complex detector value. A direct phase shot samples one phase cell and reads it directly. Thus the experiment matches reported measurement count, not underlying phase-cell accesses or compute costs. No feedback-based route adaptation, PDE wavefront, physical speed map, or motion/occlusion history is used.
Source evidence and frozen constraints
Actual JSON: seed 41 · Actual JSON: seed 97 · Predeclared success rule · Benchmark source · Unitary route implementation
Both writers retained the v2 checkpoints; the four answers per image were evaluated using identical train/validation/test scenes across checkpoints. Only small reader heads were trained. The route measurements were cross-checked numerically against explicit twelve-step unitary wave evolution. No v3 results were used to revise the writer weights.
SihtiGrapImageModel — an experimental inquiry into what a retained visual state can answer under restricted observation.