LECUN · PROBLEMS v1 + v1.5 stubs merged 19/19 verifiers 57/57 reproduces 17 yes · 2 partial

Auto-research-loop · run #3 · local build

Yann LeCun's
applied lineage,
rebuilt in numpy.

A team of agents implemented LeCun's applied papers as small, runnable, paper-comparison stubs — pure numpy, <5 min/seed on a laptop. Each batch shipped as one PR, reviewed by an independent verification team that ran the code itself. This is what they built, and how well it reproduces.

19
stubs · v1 + v1.5
76
agents · 19 build / 57 verify
57/57
verifier sign-offs
5.85M
agent tokens
the loop · per wave
SPEC.md
one contract every stub links to
→
builders × N
one agent per stub, disjoint folders, self-verify
→
verify team × 3
repro-runner · spec · faithfulness-adversary — run the code
→
PR → merge
lead merges to main on a green verdict
reproduces
✓v1 + v1.5 complete · 19 / 19 stubs17 yes · 2 partial · 57/57 verifiers
The two honest partials (both v1; closed/scoped in v1.5)
  • lenet5-mnist — subset partial (97.25%); closed by the v1.5 lenet5-full-mnist (0.99% ≈ paper 0.95%).
  • optimal-brain-damage — diagonal-Hessian saliency is genuinely second-order (r=0.996) and wins the single cut, but ties magnitude after retraining on a small MLP.
v1.5 done (waves 7–8) — the deferred set
  • lenet5-full-mnist — 0.99% on full MNIST vs paper 0.95%
  • graph-transformer-networks — segmentation-free string reading
  • multi-column-cifar — committee beats best column (real images)
  • ssl-real-image — real-image SSL > random (CIFAR stand-in for ImageNet)