Article · AI & compliance
Calibration per use case
A threshold that works on one model and one task is not a number you can take with you. That is the mistake that most often turns a gate into decoration.
AI & compliance · method · about 3 min
Once you have found a boundary that works — the model has to defer here, it may run on its own there —
the temptation is to use it again on the next task. That is understandable. It is also where most
gates stop working without anyone noticing.
Why the number cannot be moved
The signal a gate reads is a distribution, not a physical constant. The distribution moves
when something underneath it moves:
- The model. Two models on the same task have different distributions. A threshold that
sits in the middle of one can sit far out in the tail of the other.
- The platform. The same model served in two places is not necessarily the same model in practice
— quantisation and serving setup change the signal you read, without anything in your code
changing.
- The task. Short answers and long answers give different distributions. Multiple choice and
free text give different distributions. If you profile in one format and measure in another,
you are not comparing the same thing.
A threshold copied from another setup looks like it works — right up to the
day it was supposed to catch something.
What calibration involves
- Run it on your own task with the model and the platform you will actually use. Not an
equivalent one.
- Run deep enough to find the ceiling. Stop too early and it looks as though the system
has stopped improving, because you stopped looking.
- Analyse before you design the gate. The distribution tells you where the boundary belongs. Set
the boundary first and look afterwards, and you have picked a number and then gone looking for a
justification.
- Measure in the same format you evaluate in. Otherwise you measure one thing and decide on
another.
- Calibrate again when something changes — model, version, vendor or task.
The cheap version
It does not have to be big. A couple of dozen representative tasks, run with the settings you
will actually use, already tell you whether there is enough spread at all to set a
meaningful boundary.
If there is not — if everything sits at one end — then the answer is not a different threshold. The
answer is that the gate cannot carry that decision on that task, and it is better to know beforehand.
Why this belongs to compliance
Article 15 asks for the level of accuracy and the measurement methods to be declared. A number measured on another
model, another platform or another task is not your level of accuracy. Calibration is
what makes the difference between a number you have and a number that applies to you.
See also the audit pack and the open measurement protocol.