DEAD · 2026-07-17

The symptom site is not the treatment site

DESIGN  grow a small convnet where it measurably conflates inputs
RUNS    v1: 5 seeds · v2 confirmation: 10 fresh seeds, preregistered
COST    three nights, one RTX 4070 SUPER

The idea has real intuitive pull. A network that is too small for its task must, somewhere inside, be mapping inputs it should distinguish onto the same internal representation. Those collisions are measurable. So: train a seed network, find the layer where irrecoverable collisions concentrate, grow new capacity exactly there, initialised specifically to separate the colliding pairs, and repeat. Capacity delivered like medicine, to the site of the injury, instead of scattered uniformly the way a wider architecture scatters it.

This is the story of that idea surviving one experiment in a way that should already have worried us, and then dying completely in the confirmation run. It is the cleanest illustration in this ledger of why fresh seeds exist.

Act one: the encouraging pilot

Five seeds, a ~6.5k-parameter seed network on CIFAR-10, three growth arms (random growth, gradient-guided growth in the GradMax style, and collision-guided growth) plus a dense-from-scratch control at matched final size.

armaccuracyparams
seed only (floor)67.75 ± 0.446,474
grow random71.31 ± 0.77~11,136
grow gradient-guided71.37 ± 0.52~12,785
grow collision-guided72.23 ± 0.539,606
dense from scratch73.22 ± 0.4511,286

v1, mean ± std over 5 seeds. Collision-guided growth was the best growth arm with 14–25% fewer parameters.

The preregistered gate required paired t > 2.24 against both baselines. Against gradient-guided: t = 3.76, all five seeds, pass. Against random growth: t = 2.19. Fail, by 0.05 of a t-statistic. We recorded it as a fail, because the alternative, rounding a preregistered threshold in your own favor, is how ledgers rot. But the ablations looked promising: decomposing placement from initialisation showed the collision-trained filter itself carrying +0.60 points at t = 3.19, the largest and most reliable effect among the growth arms. A near-miss with a mechanistically clean story. Any reasonable person runs the confirmation.

Act two: the confirmation

GATE · WRITTEN BEFORE THE RUN Ten fresh seeds, never used in v1 — reusing the original five would be significance chasing. Gate 1: the filter effect confirms at paired t > 2.262. Gate 2: collision-guided growth with its own placement signal matches dense-from-scratch at identical architecture and parameter count.
comparisonv1 (5 seeds)v2 (10 fresh)
filter effect+0.60pp, t=3.19+0.34pp, t=1.61 · FAIL
growth vs dense, matched arch−1.91pp, t=6.84, 10/10 · FAIL

Both preregistered gates failed. The pilot effect did not confirm; growth lost to scratch training in every single seed.

Both gates failed, and the second failed in the worst possible way for the idea: at identical architecture and parameter count, a network trained from scratch beat the grown network by 1.9 points in 10 of 10 seeds. Growth wasn't merely failing to beat a fair baseline. The growth process itself was leaving accuracy behind.

The placement signal turned out to be worse than useless. Given a free choice, the collision metric chose the deepest layer in 163 of 180 growth events, and following that choice cost 3.2 points (t = −16) compared to just forcing all growth into layer 0. Collisions really do concentrate deep, the measurement was correct, and acting on it was harmful: deep growth builds architectures that are bad even when you train them from scratch. The signal was pointing at where the error is visible, which is not where capacity helps.

VERDICT Collision-guided growth is dead, and the frame it came from is damaged beyond this implementation: constructive growth never matched dense-from-scratch in any tested configuration, v1 or v2. Two lessons, both transferable. Where irrecoverable error is observed is not where capacity helps — the symptom site and the treatment site are different places, in networks as in medicine. And pilot effects shrink: t = 3.19 on five seeds became t = 1.61 on ten fresh ones. Never claim from a single small sweep; the confirmation run is not bureaucracy, it is the experiment.

What this bought us

Respect for a 0.05 margin. If we had rounded t = 2.19 up to a pass in act one, the ledger would today contain a false positive with a plausible mechanism attached, which is the most dangerous kind, and the confirmation run that killed it might never have been scheduled. The preregistered threshold did its job twice: once by refusing the claim, and once by forcing the run that settled it.