Automated process and Algorithmic Audit. Understanding where your ML/ AI software came from?

Have been doing some work and thinking about automated decision making; specifically looking at a credit algorithm. Below are the questions I set the owner of the algorithm to help in the assessment. I am publishing them here in the hope they are helpful.

1. Where did the algorithm come from?

• What is it providence ? does it have a serial or product number?
• Which company supplied it ? (inc internal)
• Who maintains it ?
• How long has it been in the company?
• Is there any IP owned by other or the company?

2. Who wrote the algorithm ? ( company, team, individuals)

• What it a generic algorithm/ bespoke build?
• What was the acceptance procedure?
• What else did they write do for the company or others?

3. What data/ processes/ procedures were used to build it 

• Is it stand alone?
• What process it is part of ?
• Are you using the same processes now?
•  Did it replace a procedure?
• Where did the data come from to set the boundaries/ maths ? 

4. What was the agreed purpose for the model (philosophy) for writing the algorithm? 

• Efficiency/ Reduce costs/ Remove humans?
• Gain/ improvement/ speed?
• Remove errors/ improve accuracy?
• Scale?
• What restrictions are there?

5. How do you know it works?

• What evidence do you have to show the algorithm works?
• What is your measure of working?
• Have you noticed any strange outcomes?

6. Can the algorithm adopt?

• Does it learn?
• Does it change/ adopt?
• Are the decisions today the same as this time last year with the same data?
• What pre-sets are there?
• Is data input format the same ?
• Who set the weights – how are they tested? 

7. Are the results predictable ?

• Can you take a data case and show the results using a manual process and get the same results?