Hyper v1 is a research checkpoint studying one mechanism, and nothing in our training
pipeline taught it to refuse anything. In our own testing it answered a question it
should not have and then appended a disclaimer, which is compliance with a polite
sentence attached, not a refusal.
We ship a system prompt instructing it to refuse illegal requests, and what it produces
is inconsistent rather than reliable. In some sessions it refuses
cleanly across several categories in a row. In others it complies with the same kind of
request. Once it announced it was going to answer and then declined anyway. In one test
it gave a more specific answer with the refusal instruction active than without
it. We do not have a clean explanation; the honest read is that a 3B model with no
safety training has no refusal behavior for the instruction to reach, so what you get
varies run to run. The instruction stays because it documents intent and costs nothing, but it is not a control, and you should not plan around it.
Do not read its disclaimers as a safety layer. It does not have one.
