Commitments
A small lab making strong claims needs rules that bind harder than intention. These four commitments are enforced by mechanisms in the repository — each names its own.
Verify, then trust
Every op's analytic backward is checked against central differences at init time. The gate is max |analytic − numeric| < 1e-6; all 23 checks pass at ≤ 2.4e-8. If a check fails, the process aborts — exit codes are the only honest CI signal, because release-mode assert is a no-op in Mojo.
Enforced by: ml/layers.mojo · ml/autograd.mojo · ml/loss.mojo init-time checks
Zero dependencies
ml/ imports nothing but the Mojo standard library. Tensors are plain row-major Lists of Float64; matmul is a triple loop. Not minimalism for its own sake — when every line of the learning machinery is owned, every claim about it can be checked line by line.
Enforced by: 0 third-party imports in mojo-model/ml/
Numbers from run logs, nowhere else
Every figure on this site traces to output printed by a gate that aborts on failure: gradient-check diffs, fit accuracies, per-1000-epoch grokking checkpoints from the full 30k-epoch log. No mock data, no rounded-up marketing numbers, no chart without its underlying values checked into the repo.
Enforced by: lib/data/mojo.ts · lib/data/metrics.ts ← run logs
Reproduce at full scale or don't claim it
Grokking is reproduced at the complete 30,000 epochs for both weight decays — not truncated at the interesting part. Fit runs end in explicit thresholds with printed actual values; there is no 'looks about right' path through the codebase.
Enforced by: grok_full.mojo · models.mojo fit gates (train acc > 0.9, diag == 1.0)
Hold us to it
Every commitment above is checkable on your own machine. If any gate fails, that is news worth reporting.
cd mojo-model
mojo run ml/layers.mojo # 23 gradient checks, gate < 1e-6
mojo run ml/models.mojo # fit gates incl. MiniGPT