Director level Safety and evaluation, advisory

The work

Designs the tests that say whether a system is good enough to ship.

A standing second opinion. Small commitments, and the ones where being right matters most because nobody is checking your work.

The bar

You have carried the decision at a level where being wrong was expensive, and you can describe a time you were.

Why this one is hard to judge. Writing an eval is easy. Writing one that would actually have caught the failure is not.

What we ask to see

• An evaluation suite you designed, and what it was allowed to fail
• A time an evaluation passed and the system was still not fit to ship
• The evaluation you retired because it had stopped measuring anything

One solid item beats five vague ones. A referee we can actually call beats a link to a landing page.

What gets an application declined

• The only evidence offered is a certification, a course, or a job title
• Every piece of evidence is under an NDA and no sanitised version or referee is offered
• The work described is entirely a demonstration, a proof of concept, or a hackathon
• More than six disciplines are claimed, which we read as a signal that none has been assessed

If we decline you, the reason is written down and sent to you. The software refuses a decline without one.

What it pays

Nothing is published for this discipline in the UK advertised market, so there is no figure to give you and we are not going to invent one.

AI safety does not exist as a title in the UK advertised market. What little there is sits inside research posts at a handful of labs and is not advertised with a figure.