Five compact cases
Apple MLX: one-element strided slice
A plain slice has shape (1,3), while vmap produces (4,2,3); reverse-mode gradients are also scattered to the wrong positions.
Samsung nntrainer: five unary derivatives
Real calcDerivative calls for sqrt, negative, sine, cosine and tangent fail ordinary chain-rule oracles when the incoming derivative differs from one. A minimal in-place correction makes all five focused regression tests pass.
Google JAX: log2 of exact powers of two
Reproduced 46/254 float32 and 105/254 bfloat16 mismatches. The observation is firm; the boundary between API defect and accuracy/performance tradeoff remains under discussion.
Original report by gonnet →PyTorch: eager versus Inductor
For unrepresentable float32 clamp bounds, eager raises while Inductor silently returns the input. The split reproduces for clamp, clamp_min and clamp_max.
Original report by laolvfan →TensorFlow: MaxPool channel symmetry
Three identical channels produce [NaN, NaN, 3] when packed and [3,3,3] when processed separately. The proposed upstream fix is still open.
The GERO method
State a necessary invariant. Construct the smallest counterexample. Run the actual library code. Preserve the version, command and raw output. Then distinguish a confirmed defect from a documentation mismatch or a numerical boundary.
This is the product direction of GERO: transform necessary claims into explicit contracts, route them to heterogeneous checks, search for counterexamples and return bounded verdicts with replayable evidence.
Read the full canonical article →