Model validator
The one entitled to say no before it goes live
A model validator examines a system before it decides anything at all and can refuse to let it go live, like a vehicle inspector whose verdict governs the right to drive.
- Floor 4The Organisation
- P2Writes the application
- Established titleIt has been in adverts long enough to mean something.
What they answer for
The deliverableA reasoned opinion that holds, clearing or refusing a release. If they leave, the decision chain loses its stopping point.
The profile
Five activities, scored 0 to 3- ResearchProducing knowledge that does not yet exist.
- BuildShipping a system that runs, deploys and breaks.
- OperateKeeping it in production: cost, incidents, drift, on-call.
- VerifyMeasuring, testing, attacking. Producing a verdict that holds.
- LeadDeciding, persuading, driving adoption, answering to others.
What the work is
Where the trade comes from
The trade was not born with AI. Banks and insurers have validated their risk models for decades, under a written doctrine and under a regulator’s eye. The arrival of models whose individual decisions nobody can explain did not create the function; it made it far harder.
Why it reads code
This is the off-diagonal case of floor 4: you answer for an organisational decision, and yet you read code, replay test sets, recompute measurements. An opinion that merely re-read a piece of documentation would not hold against the person who wrote it, nor against a regulator.
The trap
The trap is real authority. The trade only exists if the opinion can stop something. Where validation is an administrative step in a schedule already fixed, the post consists of signing off what would ship anyway, and that is the worst place to be standing when the incident arrives.
A week in the role
- Re-reading a model’s documentation, and above all what it does not say.
- Recomputing measurements yourself, because a number you have not recalculated opposes nobody.
- Writing an opinion, with its reservations, in language that holds in front of a third party.
- Talking to the people who built it, who are not expecting to be slowed down.
Ways in
- From quantitative risk, in banking or insurance, which is the historical route.
- From data science, accepting that the deliverable becomes an opinion and not a model.
- What the role requires and adverts describe badly: a taste for holding a position under pressure.
Reading an advert
2 signs- The advert does not say who receives the opinion, or what it can stop.
- With no power to block, this is not validation. Ask for the last time an opinion pushed a release back.
- The post reports into the team that builds the models.
- Independence is the condition of the trade. That reporting line is not fatal, but it has to be offset somewhere, and you need to know how.