CASMO (Confidence-Adjusted Signal-to-noise Momentum Optimizer) is a drop-in replacement for torch.optim.AdamW that scales each coordinate's update by the signal-to-noise ratio of its gradient.

from casmo import CASMO

optimizer = CASMO(model.parameters(), lr=1e-3, weight_decay=0.01)

Adam adapts to gradient magnitude and cannot tell a strong consistent signal from large random noise. CASMO tracks the gradient's first moment m (signal) and its centered variance s (noise), and forms a per-coordinate signal fraction AGAR = m̂²/(m̂² + ŝ) ∈ [0,1]. That fraction drives two multipliers: trust, an absolute per-tensor pace that shrinks when gradients are noise-dominated, and focus, a relative per-coordinate weight that shifts the step toward reliable directions. The robustness dial sets how strongly the absolute axis applies — 0 behaves like AdamW, 1 is maximally noise robust.

CASMO(params, lr=1e-3, robustness=1.0)   # noisy labels, long-tail, continual learning
CASMO(params, lr=1e-3)                    # default 0.5 -- general purpose
CASMO(params, lr=1e-3, robustness=0.0)   # clean data, DP-SGD, maximum speed

5 seeds, mean ± stdev, clean test labels. Reproduce with python research/validate_redesign.py.

The last row is a real limitation, not an omission: label noise and isotropic gradient noise want opposite policies. Use low robustness for DP-SGD-style workloads.