The summary.
REMOVING PROTECTED ATTRIBUTES DOES NOT MAKE A MODEL FAIR
Other features correlate with them. Measure performance disaggregated by group, and choose a fairness definition deliberately — several cannot hold at once.
Never claim a model is fair without stating what that means.
MODELS ARE REUSED FOR PURPOSES THEY WERE NEVER EVALUATED FOR
Which is why documenting intended use, and what the model must not be used for, matters more than documenting architecture.
Never deploy a model nobody can describe the limits of.
OVERSIGHT REQUIRES INFORMATION, TIME AND AUTHORITY TO DISAGREE
Without all three it is rubber-stamping. A very low disagreement rate means either an excellent model or oversight that is not happening.
High appeal overturn rates mean the model should not be deployed as it is.
BUILD THE THINNEST END-TO-END PATH FIRST, WITH A TRIVIAL MODEL
The path is the real risk, not the model. Months of modelling before anyone has seen a prediction in place is how projects die.
AGREE WHAT PERFORMANCE IS GOOD ENOUGH BEFORE BUILDING
Afterwards, the answer becomes whatever was achieved.
