Research · pugXiv:2604.00036
Constitutional Napping: Alignment Through Scheduled Rest
L. Napworth, W. Wafflebrook, M. Meatley
- Alignment
- Safety
Abstract
We propose a training procedure in which models are given a written constitution and a mandatory afternoon nap. Models trained this way refuse a greater share of harmful requests, primarily by being asleep.
1. Introduction
Constitutional training gives a model a written set of principles to follow. We extend the approach with a second ingredient: a mandatory afternoon nap. We ask whether rest improves alignment, and what, precisely, is aligned.
2. Method
Models were given a short written constitution and a scheduled rest period. The constitution contained three principles, excerpted here:
- Be helpful.
- Be harmless.
- Be asleep between 1 and 4 p.m.
3. Results
Models trained with Constitutional Napping refused a greater share of harmful requests than the baseline. We attribute most of the improvement to the models being asleep. Requests received during the rest period were not answered, whatever their content.
We also report an alignment tax: the same models refused a larger share of harmless requests, for the same reason.
Percent of requests refused. Higher is better for harmful, lower is better for harmless.
| Benchmark | Baseline | Constitutional Napping |
|---|---|---|
| Harmful requests refused | 61% | 93% |
| Harmless requests refused | 4% | 38% |
4. Legal review
The constitution was reviewed by M. Meatley, who confirmed that it contained no sandwiches and approved the nap.
5. Conclusion
Alignment through scheduled rest is simple, cheap, and effective on a schedule. We recommend that future work investigate shorter naps, and that it be done in the morning, when the models are awake.
Cite this paper
@article{constitutional-napping,
title = {Constitutional Napping: Alignment Through Scheduled Rest},
author = {L. Napworth and W. Wafflebrook and M. Meatley},
journal = {pugXiv},
number = {2604.00036},
year = {2026},
note = {Peer reviewed by a pug}
}