Here is what my product does: [describe].<br/><br/>List every decision about a person that my output could influence, including ones I did not design for - hiring, promotion, scheduling, pay, discipline, termination, access, ranking against colleagues.<br/><br/>For each, say whether my output is evidence a human weighs, or the thing that decides. Be strict: a score presented prominently with no competing information is deciding, whatever the documentation claims.<br/><br/>Then flag any where a customer could reasonably use my output as the primary basis for an adverse decision.
Below is how my system logs its scoring or ranking decisions.<br/><br/>Tell me whether I could, six months later, reconstruct for a specific person: which inputs the model received, which were weighted most, what the output was, what version of the model produced it, and what a human did with it afterwards.<br/><br/>For each thing I cannot reconstruct, tell me what I would need to log, and the cheapest place in my pipeline to log it.<br/><br/>Assume I am being asked by someone entitled to an answer.<br/><br/>LOGGING:<br/>[paste schema or description]
Here is how my product currently displays its output to the person making a decision: [describe or paste UI copy].<br/><br/>Rewrite it so the output reads as evidence to be weighed rather than a conclusion to be accepted.<br/><br/>Specifically: surface what drove the score rather than the score alone, show what the model did not see, make disagreeing with it as easy as accepting it, and remove any language implying the system has judged the person.<br/><br/>Then tell me what I lose in usability by doing this, honestly - because if it becomes annoying, people will route around it.
Write the notice a person would receive explaining that an automated system contributed to a decision about them.<br/><br/>It must say in plain language: what data was used, what the system assessed, what it concluded, who reviewed it, and how to contest it.<br/><br/>Constraints: no jargon, no "proprietary algorithm" hand-waving, and nothing that would embarrass us if it were published - because it will be.<br/><br/>Then tell me which parts of my actual system I cannot currently describe this clearly, because that is the real finding.<br/><br/>SYSTEM:<br/>[describe]
My process requires a human to review automated findings before acting. Here is how that works: [describe].<br/><br/>Tell me whether that review is real or ceremonial. Consider: how long a reviewer has per case, how many they handle, whether they see the model output before forming their own view, whether disagreeing creates work for them, and whether anyone tracks how often they override it.<br/><br/>If the override rate is near zero, tell me what that most likely means.<br/><br/>Then give me three changes that would make the review genuinely independent, ordered by how much friction each adds.