Skip to main content

Troubleshooting

Common issues, what they mean, and how to fix them.

NotImplementedError for a derivative order​

NotImplementedError: order 3 is not implemented for activation 'relu'

Cause. The activation does not have a closed-form tower at that order. relu, silu, and gelu support only op="identity" and op="grad".

Fix. Switch to a smooth (Riccati) activation — tanh, sigmoid, softplus, gaussian — for op="laplacian" or any order n ≥ 2. See the stability matrix.

ValueError for a negative order​

ValueError: derivative order must be >= 0

Cause. A fastpath was asked for n < 0. This is always an error by contract.

Fix. Check the call site computing the order.

Backends disagree at the last bit​

Cause. You are almost certainly comparing in float32. Parity is a float64 guarantee; in float32 both backends round to 32 bits but intermediate orderings can differ in the last bit.

Fix. Re-run the comparison in float64. A genuine float64 mismatch is a bug — report it. See cross-backend parity.

ImportError after install​

ModuleNotFoundError: No module named 'omnibias.torch'

Cause. Backends are separate distributions. Installing omnibias-core does not install a backend.

Fix. Install the backend you need: pip install omnibias-torch (or omnibias-jax). See installation.

Keras backend not found​

Cause. KERAS_BACKEND is unset or inconsistent with what is installed.

Fix. Set it explicitly before import, e.g. KERAS_BACKEND=jax.

Certificate verification fails​

verify_certificate_digest(cert) -> False

Cause. The certificate's content changed since it was sealed (even a whitespace edit), so the recomputed hash no longer matches.

Fix. Re-fetch the original certificate; never hand-edit a sealed file. If you need a new bound, regenerate the enclosure and make a fresh certificate. See provenance & certificates.

theorem_prover_verified is false​

Cause. No Lean toolchain was present, or the obligation did not pass.

Fix. This is honest behavior, not a crash. Install the Lean toolchain to attempt a kernel check; a false flag must never be reported as proven. See the formal loop.

A PINN residual is unstable at high order​

Fix. Confirm you are using a smooth activation, prefer the closed-form poly-Laplacian over nested autodiff, and consider a structural cage for hard constraints. See high-order PDEs.

Still stuck?​

  • Check the FAQ.
  • Skim the glossary for an unfamiliar term.
  • File a minimal reproduction (float64, exact versions, seed).