← Parallax Metrology · Credit & provenance · Existing computational explanation; toy quantities are not validated production-model influence or legal allocations.

Parallax Metrology

Diffusion Microscopetechnical exploration

corpus 16 images tensor 24×24×3 model kernel attention

11 linked experiments.
One corpus, one noise seed, one inference trace.

Orientation · claim boundary

Trace four different quantities

exact ≠ estimated ≠ logged ≠ allocatedmove through the chapters

Ready. Select a chapter or press →.seed 6042 · deterministic

Interactive visual explanation of a finite-mixture diffusion denoiser, patch attention model, training run, reverse sampler, and source participation trace.

Reading rule

Boundary

Controls

Synthetic corpus manifest

Every training image is generated locally from deterministic geometry and color equations. The browser turns each recipe into a real 24×24×3 tensor and checksums its canonical little-endian Float32 bytes. No external image or network request is used.

Replayable toy inference receipt

This receipt records source-memory participation and the declared aggregation policy. It is evidence about this instrumented toy mechanism, not a legal or causal finding.

How to read the microscope

Arrow keys move between chapters. Space runs or pauses the active experiment. R resets it. M opens the corpus. E opens the receipt. Click pixels, patches, bars, and memory cells to inspect their numbers.

Four quantities stay separate throughout: exact posterior responsibility in the analytic finite mixture; learned or estimated behavior in the toy model; logged source-memory participation during inference; and a contractual allocation policy applied afterwards.