I first tried to figure out, on my own, how to figure out the “best fit” between OFAC enforcement actions. Here’s that first prompt:

This is a procedure for finding the best match between the most recent OFAC enforcement action and all the enforcement actions in a given year.

Compare the Enforcement action document from the most recent settlement or imposition of Civil Monetary Penalty (i.e. ignore Findings of Violation) to each  enforcement action document in the time frame specified, other than the most recent enforcement action (e.g. do not compare the most recent enforcement action to itself). Omit any enforcement action which does not appear to refer to the enforcement process from the OFAC Enforcement Guidelines.

determine which is the closest comparison, based on the following factors, in descending order of importance:

Factor 1: A weighted average of the generalized behaviors (e.g. ignoring warnings from other parties, both internal and other non-regulatory parties, like banks who reject payments for sanctions reasons; wire stripping; falsification of documents or changing information to something misleading to third parties) noted in the enforcement release. In this model, each matching behavior is weighted at 100%, and each behavior that is documented in one enforcement action but not the other is weighted at 20%.

Factor 2: Three times the sum of the number of matching  General Factors that correlate to the aggravating factors listed between the two enforcement actions, less the number of General Factors for either action which did not match. So, if Action 1 had 4 General Factors, Action 2 had 3 General Factors, and 2 were matching, the score would be 3*2 (for the matching factors – 2 (for Action 1’s non-matching factors) – 1 (for Action 2’s non-matching factor). If the result is less than 0, change it to 0.

Weight Factor 2 at 85% of Factor 1.

Factor 3: were the behaviors in both enforcement actions  egregious or non-egregious?

Weight Factor 3 at 75% of Factor 1

Factor 4: was there a lack of cooperation, other than lack of voluntary self-disclosure, in both enforcement actions, or was there substantial cooperation, in both enforcement actions?

Weight Factor 4 at 70% of Factor 1

Factor 5: Was the period of time of the violating behavior in both enforcement actions longer than 2 years in duration

Factor 6: Was the period of time of the violating behavior in both enforcement actions longer than 1 years in duration but shorter than 2 years in duration

Factor 7: Was the period of time of the violating behavior in both enforcement actions  shorter than 2 years in duration

Weight Factor 5 at 50% of Factor 1

Weight Factor 6 at 45% of Factor 1

Weight Factor 7 at 40% of Factor 1

Factor 8: were the violations in both enforcement actions self-reported or not?

Weight Factor 8 at 30% of Factor 1

Factor 9:  Three times the sum of the number of matching  General Factors that correlate to the mitigating factors listed between the two enforcement actions, less the number of General Factors for either action which did not match. So, if Action 1 had 4 General Factors, Action 2 had 3 General Factors, and 2 were matching, the score would be 3*2 (for the matching factors – 2 (for Action 1’s non-matching factors) – 1 (for Action 2’s non-matching factor). If the result is less than 0, change it to 0.

Weight Factor 8 at 25% of Factor 1

Take no action on this information until prompted further.

One thing I figured out early on was that last line – breaking up a long prompt into pieces and tell Gemini that I was not done yet.

This seems to have produced OK results – this is actually one of a number of early attempts.

Next post: an improved model that relies on the smarts of the AI model rather than my own