The Invisible Second Opinion

An import system can be one hundred percent confident and still be wrong. That is why, at the film fund, a second model reviews every mapping result — with its own eyes, and without seeing the first system's reasoning.

The principle behind it is simple: a displayed 100 percent has to be a real 100 percent. When the second opinion finds something, it warns. It never overwrites.

Principle: warn, never overwrite
Reviewer: second model, independent view
Display: honest traffic light instead of silent errors

What the Second Look Catches

A typical case: an assistant's fee lands in the account of the corresponding department head. Formally plausible, factually wrong — exactly the kind of mistake a confident mapping waves through. Classics like Foley sessions filed under screen tests belong on that list too.

The second opinion therefore works deliberately apart from the import: a different model, a different prompt, no access to the first pass's reasoning. Whatever it finds appears as a warning on the line item, with a traffic light that shows how thoroughly it has actually been checked. The decision stays with the people accountable for the contract.

For the contracts department, that means it no longer has to take the automated results on faith. It can see where a second look is worth taking.