Add explainers/fairness-accuracy-tradeoff.md.
Mitigation Strategies covers the pre-/in-/post-processing ladder and its effect on the fairness gap, but never mentions accuracy at all (confirmed: the word "accuracy" doesn't appear in that file). What it leaves out is the other half of the same story: those same mitigation strategies also change model performance, and paper/results-frozen/results_performance.csv already has the numbers to show it, currently unused by any explainer.
Suggested structure (match mitigation-strategies.md): one-sentence definition -> why it matters (a fairness fix is rarely free) -> core concept (the Pareto frontier - you can't always improve one without costing the other) -> concrete example tied to a repo audit, using real frozen numbers from results_performance.csv alongside results_fairness.csv -> runnable code that plots or tabulates accuracy vs. gap across the five strategies -> limitations (the trade-off isn't fixed or universal; a real fix can occasionally improve both) -> related concepts/projects/further reading.
A verified real anchor: COMPAS logistic_regression accuracy drops from baseline to post_processing while the race demographic_parity_diff closes from 85.5% down to 2.3% (with a brief overshoot past zero at in_processing, -1.4%, not statistically significant) - both pulled directly from the frozen CSVs, the exact kind of concrete before/after this explainer should quote precisely (verify the exact accuracy numbers from results_performance.csv yourself rather than trusting this issue's paraphrase).
Related: mitigation-strategies, accuracy-equality, bias-variance-tradeoff.
See CONTRIBUTING.md: add explainers/<slug>.md plus a one-line entry in assets/explainers-data.json; the build script generates the page.
Freeze-safe: explainers are fully open for contribution during the paper freeze. If you quote any Fair Code benchmark result, use the frozen numbers in paper/results-frozen/ (never re-run your own) - see CLAUDE.md.
Add
explainers/fairness-accuracy-tradeoff.md.Mitigation Strategies covers the pre-/in-/post-processing ladder and its effect on the fairness gap, but never mentions accuracy at all (confirmed: the word "accuracy" doesn't appear in that file). What it leaves out is the other half of the same story: those same mitigation strategies also change model performance, and
paper/results-frozen/results_performance.csvalready has the numbers to show it, currently unused by any explainer.Suggested structure (match mitigation-strategies.md): one-sentence definition -> why it matters (a fairness fix is rarely free) -> core concept (the Pareto frontier - you can't always improve one without costing the other) -> concrete example tied to a repo audit, using real frozen numbers from
results_performance.csvalongsideresults_fairness.csv-> runnable code that plots or tabulates accuracy vs. gap across the five strategies -> limitations (the trade-off isn't fixed or universal; a real fix can occasionally improve both) -> related concepts/projects/further reading.A verified real anchor: COMPAS logistic_regression accuracy drops from baseline to post_processing while the race
demographic_parity_diffcloses from 85.5% down to 2.3% (with a brief overshoot past zero atin_processing, -1.4%, not statistically significant) - both pulled directly from the frozen CSVs, the exact kind of concrete before/after this explainer should quote precisely (verify the exact accuracy numbers fromresults_performance.csvyourself rather than trusting this issue's paraphrase).Related: mitigation-strategies, accuracy-equality, bias-variance-tradeoff.
See CONTRIBUTING.md: add
explainers/<slug>.mdplus a one-line entry inassets/explainers-data.json; the build script generates the page.