An independent CPU audit of Apple MLX 0.32.2. At float16 x=1000, arcsinh and arccosh return finite values but zero JVP/VJP, although the correct slopes are about 0.001. Intermediate squaring overflows.

A normalized real-derivative patch passes 518 native comparisons across float16, bfloat16, float32 and float64: 188 mismatches before, zero after. Second and third derivatives are checked for float32/64 only. This is one shared defect affecting two functions, not 188 separate findings.

Full report: https://www.gero.uz/research/articles/mlx-inverse-hyperbolic-overflow.html
Code, patch, results and source pins: https://github.com/kadyrbekovhamit-cyber/gero-numerical-observatory/tree/main/audits/2026-09-10-mlx-inverse-hyperbolic-overflow/
Pinned upstream source: https://github.com/ml-explore/mlx/blob/81ba1c6a0e50a9268b931579c2d4f1158b9aab5a/mlx/primitives.cpp

Scope: isolated native translation-unit rebuild before an existing CPU archive, not a clean current-main build. GPU, speed, compile/vmap, complete nonfinite behavior and full-model effects were not tested. A bounded search found no exact MLX duplicate; absolute priority and upstream acceptance are not claimed.

Narration uses a synthetic voice (macOS Samantha; fictional narrator Alex Vector). Original evidence diagrams. AI-assisted research, code, writing and production; editorial review by Technology Product / Xamit Kadirbekov. No sponsor or paid promotion.

#MLX #Autodiff #NumericalComputing