UPLC implementations disagree on quotientInteger costing

[nielstron]

I noticed that the Python implementation of UPLC and the latest version of aiken (aiken v1.0.21-alpha+4b04517) disagree on the costing of the acceptance tests for quotientInteger from plutus-conformance.

Specifically this program:

(program 1.0.0 [(builtin quotientInteger) (con integer -503783783785265728700234277) (con integer -1777777777)])

Produces the following cost in Python:

{
  "result": "(con integer 283378378503190012)",
  "cpu": 311600,
  "mem": 601
}

And the following in aiken:

{
  "result": "(con integer 283378378503190012)",
  "cpu": 568780,
  "mem": 601
}

The python implementation is using the builtinCostModel.json as a base and updates it parameters based on the latest network config from https://book.world.dev.cardano.org/cm-2023-02-14.html.

Which version is aiken using? Is there a useful external tool to assess the differences?

[MicroProofs]

Not sure if this was answered but I flipped the sign for over the curve to correctly match plutus