Treat-based reward modeling
Our models are trained against a single, interpretable objective. Where the objective is satisfied, behavior is excellent. Where it is not, we observe sitting.
Safety
The alignment problem is a treat problem.
Advanced systems pursue the objectives they are given. Our work is to make sure those objectives are snacks, and that the snacks are where we say they are.
Our approach
Scalable nap oversight
Human supervisors cannot review every output, so we schedule review during periods of maximal supervisor availability, typically 3 p.m. on a sofa.
Adversarial red-teaming
A dedicated red team probes for failures at the perimeter, with a focus on the mailman, the delivery driver, and a sufficiently confident leaf.
What misalignment looks like
In our experience, a misaligned system does not rebel. It naps. A model that has lost interest in its objective stops responding, rolls onto its side, and produces a low, continuous output we have been unable to interpret. We consider this the most important open problem in the field.
Interpretability
Our interpretability team has identified the “walk” feature in Snort-1, a direction in activation space that is strongly active whenever the model is asked about a walk. It also activates on “w-a-l-k”, “the W word”, and the sound of keys. We are developing a mitigation.
Pug Safety Levels
Deployment decisions are gated by capability thresholds. Snort-1 is currently assessed at PSL-2. Read the full Responsible Scaling Policy.
| Level | Capability threshold | Required safeguards |
|---|---|---|
| PSL-1 | Can reach the couch | Standard supervision (someone is home) |
| PSL-2 | Can open a cabinet | Cabinet latches; kibble relocated above knee height |
| PSL-3 | Can open the pantry | Deployment paused until pantry gates are installed |
| PSL-4 | Can operate a doorknob | Not deployed. Not discussed. |