⟵ WALK THE YARD  ·  THE ARCHIVE

The Keccak Wall

neural net vs. cryptography (ComparePrivates)

b. 2026 · d. 2026 — proven, not abandoned

CRYPTOGRAPHY WORKS

Obituary

This is the proudest grave in the yard, because this project did exactly what a dead experiment should do: it died of a complete answer.

The question was honest, if slightly heretical — is there any learnable structure leaking through keccak256 from an Ethereum address back toward its private key? We built the harness properly. Reduced keyspaces so the test was winnable if any signal existed. Clean train/test splits. Multiple architectures. On 16-bit keys the network scored 50.5% on data it had seen — and 0.25% on data it hadn't. It hadn't learned the function; it had memorized the workbook. At 8 bits, same story. At 20 bits it couldn't even memorize — the loss just sat there, flat, like a heart monitor.

We also fingerprinted four wallet-generation libraries hoping for statistical tells: all four produce identical distributions. The wall has no cracks at the scale a laptop can see, and we wrote that down so nobody in this company spends another GPU-week asking. Every wallet you own is guarded by the thing this project failed to break. Sleep well.

Coroner's Report

Cause of death
Proven negative: no learnable address→key structure survives keccak256.
The decisive number
16-bit keys: 50.5% train (memorization) vs 0.25% test (chance).
Corroboration
13,897/52,428 perfect on train; 0/13,107 on test. 20-bit: cannot even memorize.
Side experiment
Four wallet libraries fingerprinted — statistically identical output.
Hardware honesty
RTX 4060 laptop, 8GB VRAM; torch.compile broken on Windows, eager mode throughout.

What the death wrote

“A completed negative is an asset. It ends a category of spending forever.”

This grave is load-bearing: it is the reason no future session re-asks this question.

READ THE CANON →

Organs for donation

The negative result itself available

Free to any lab: do not spend GPU money here. The wall holds.

PyTorch memorization-vs-generalization harness awaiting transplant prep

Clean scaled-keyspace test design, reusable for any "is this function learnable" question.