vt-v1-74dfbd2-46380052-47646604

Written result

Hypothesis. A variational transcriber - v9 whose 18 x 18 stem grid passes through a per-cell Gaussian latent (c = 8 channels) held at a KL budget of C = 2,000 nats per icon - keeps v9's persistent spatial record of what is still undrawn and so reconstructs held-out icons from its posterior mean close to v9, while its latent is smooth at the posterior scale. Pass, on the first 339 validation icons, IEEE float32 (TF32 off) greedy decoding, 72 px pixel error, bootstrap intervals, all comparisons paired, requires A1 and A3; A2 is a collapse check. (A1) reconstruction from mu is non-inferior to v9 re-decoded in float32 on the same icons - the upper 95% bound of (VT - v9) is at most +0.015 - and beats the nearest training render or its mirror (every model trains with mirrors), the 95% interval on the reduction excluding zero. (A2) held-out KL within [0.8C, 1.2C] = [1,600, 2,400] nats per icon with beta above its floor of 1e-4; the controller enforces this band, so it detects collapse and is not evidence of anything else. (A3) decoding one posterior sample instead of mu costs at most 0.010 pixel error, paired mean. Reported without a criterion: the z = 0 decode, N(0, I) samples as the prior-hole control (ink coverage, distinct, rendered), 8 lerp strips of posterior means (a canvas dissolve), graph-decoder latency. Decision: A1 and A3 pass - fit the flow prior (arm 3, lfp-v1). A1 fails - run once more at c = 16 under a new run identity; a second failure stops the canvas line.

Criteria

Result (339 validation icons, IEEE float32 greedy decoding (TF32 off: matmul and cuDNN); pixel metrics at 72 px)

Resources

Sheets: reconstructions.png (reference, VT from mu, VT from a posterior sample, v9, nearest training render or mirror), interpolations.png (lerp of posterior means, a canvas dissolve), prior-samples.png (N(0, I), the prior-hole control).