Bounded MLP development probe — 2026-09-30

Purpose: choose small, real changes to the two original MLP procedures for
Yaroslav's separate submissions. This is development evidence, not an
official sandboxed score and not a claim that every random run will pass.

The first probe tested one candidate per difficulty on one requested A100-80GB
Modal container. Both original sources and both candidates were evaluated
on the same six public MNIST development draws, seeds 9401–9406, produced
by the unchanged scorer's mnist.draw API. Each method received one untimed
Fashion-MNIST warm-up, seed 9400. The methods initialize all trainable and
optimizer state on every call. No weights, release transforms, or data from
this probe are embedded in either submission.

D1: 400 -> 200 training steps, TF32 matmuls enabled, fused AdamW on CUDA.
Architecture, dropout, noise, label smoothing, batch size, peak learning
rate, and one-cycle schedule shape are unchanged. The shorter schedule is
computed over 200 steps. The fixed torch.manual_seed(0) remains inside the
training function, so each call trains from scratch.

D2 first candidate: ensemble members 16 -> 8. Width, steps, batch size, learning rate,
regularization, AdamW, EMA, and prediction averaging are unchanged.
Changing K also changes how many random values the existing generator
consumes; this is a new eight-member ensemble, not cached predictions or a
selection of members based on these test labels. The per-call generator is
freshly seeded and all parameters, optimizer state, and EMA state are fresh.

                 Mean time (ms)    Mean MNIST error
D1 original          520.610             3.4000%
D1 candidate         235.413             3.8300%
D2 original          915.952             3.3283%
D2 candidate         519.548             3.3383%

Thus the D1 candidate took about 54.8% less time with increased error still
below its 5.40% band. D2 took about 43.3% less time with 0.01 percentage
points more error, below its 3.40% band on this sample. D2's remaining
accuracy margin is narrow. An official fresh-draw run must independently
qualify it. The last D2 candidate call took 635.751 ms, versus roughly
495–498 ms for its other five calls; that value is retained in the mean.

Timing used synchronized wall time within one unsandboxed process, with
one CPU torch thread. TF32 was explicitly disabled for the original D1
and enabled for the other methods so module imports could not contaminate
that comparison. This probe does not include sandbox transport overhead,
the official independent clock floor, randomized dataset ordering, the
hold-out accuracy checks, or measured energy. Official results are separate.

All 24 GPU calls completed and printed their per-draw results. Returning
the final dictionary then failed locally because torch.__version__ is a
TorchVersion object and the controller's environment has no torch package.
dev-mlp-probe.json was reconstructed verbatim from the complete measured
rows in dev-mlp-probe.log; its recovery_note records this. The controller
now converts the version to str, but no extra GPU run was launched.

The first GPU app was ap-8JhbOrWkC3oQT8YvG7pmut.

Second and final D2 candidate

Because the first D2 candidate's mean error was only 0.0617 percentage
points below the band, a second bounded probe evaluated K=8 with 500 steps
and EMA decay 1-4/500 = 0.992. All other settings remain unchanged. It was
paired with the original K=16, 400-step source on the same six public seeds,
again with one Fashion-MNIST warm-up per method. There were no other variants.

                      Mean time (ms)    Mean MNIST error
D2 original               922.961             3.3233%
D2 final candidate        782.466             3.2567%

The final D2 source is 15.2% faster than the original in this second paired
probe, with 0.0667 percentage points less error. Its margin under the 3.40%
band is 0.1433 points, which remains subject to fresh-draw variation. The
final chosen sources are mlp_d1.py and mlp_d2.py. The discarded 400-step
candidate is preserved as evidence/dev-mlp-d2-k8-s400.py. The first controller
now reads that preserved file so its earlier experiment can be reproduced.

dev-mlp-probe-500.json records the second probe, including final D2 source
hash, actual device name, torch version, app id, and all twelve measured
rows. This controller completed without the earlier deserialization error.
The original D2's small accuracy difference between the two probes is
consistent with nondeterministic GPU operations; deterministic algorithms
were not forced. Timings should be compared within each paired probe.

The unchanged scorer's source checks pass for both final standalone
candidates, with no review flags. Source SHA-256 hashes are in
dev-mlp-probe.json (D1) and dev-mlp-probe-500.json (final D2).