E0206: backend output shape mismatch on sample N
What happened
During a conformance harness run, the backend under test returned an output raster for sample N with different dimensions than the reference simulator output. The reference output was [ref_t × ref_n], the backend returned [got_t × got_n].
Why
The backend implementation does not produce the correct output shape for this sample. Possible causes:
- Off-by-one in T unrolling
- The backend silently skips empty input timesteps
- The backend trims trailing zero-spike timesteps
How to fix
- In your backend’s
run_batch implementation, ensure the output tensor has shape [batch, T, N_out] where T and N_out are taken from the artifact — not from the input data.
- The output shape must be fixed by the model, not by the input activity.
Example
