Problem
The Sleep-EDF prompt suggests using TCNs or dilated convolutions to encode sleep-stage transition regularity. However, the model receives only one independent 30-second epoch at a time with shape (Batch, 1, 3000). It cannot observe previous or following epochs and therefore cannot model transitions such as Wake to N1 to N2.
Trigger scenario
An agent follows the hint and adds temporal modules intended to model stage transitions, but those modules can only process temporal structure within a single epoch.
Impact
- The prompt misrepresents what the model interface can express.
- Agents may add unnecessary model complexity.
- Reports may incorrectly claim that the model uses cross-epoch transition priors.
- Different agents may interpret "transition" inconsistently.
Expected fix
Remove or rewrite the stage-transition hint in the current task. Clarify that the model can only learn waveform and spectral-temporal structure within one epoch.
If transition modeling is desired in the future, introduce a new task version with a sequential input contract such as (Batch, Sequence, 1, 3000).
Problem
The Sleep-EDF prompt suggests using TCNs or dilated convolutions to encode sleep-stage transition regularity. However, the model receives only one independent 30-second epoch at a time with shape
(Batch, 1, 3000). It cannot observe previous or following epochs and therefore cannot model transitions such as Wake to N1 to N2.Trigger scenario
An agent follows the hint and adds temporal modules intended to model stage transitions, but those modules can only process temporal structure within a single epoch.
Impact
Expected fix
Remove or rewrite the stage-transition hint in the current task. Clarify that the model can only learn waveform and spectral-temporal structure within one epoch.
If transition modeling is desired in the future, introduce a new task version with a sequential input contract such as
(Batch, Sequence, 1, 3000).