Where: explainers/counterfactual-fairness.md (~lines 63-69) - the COMPAS "biased model" example table.
The gap: the file cites Black defendants 58.15% / White defendants 28.37% / a 29.78pp fairness gap. This repo's actual, canonical unfair.py result for COMPAS - cited identically and consistently by README.md and at least 8 other explainers (automation-bias.md, confounding-variable.md, data-leakage.md, how-ai-detects-patterns.md, ml-bias.md, proxy-variables.md, ai-objectivity-myth.md, and more) - is Black 87.16% / White 0.40% / an 86.77pp gap.
Repro:
$ grep -n "58.15\|28.37\|29.78" explainers/counterfactual-fairness.md
67: | Black defendants | 58.15% |
68: | White defendants | 28.37% |
69: | **Fairness Gap** | **29.78pp** |
$ grep -n "87.16\|0.40\|86.77" README.md
350: | Black Defendants | 87.16% |
351: | White Defendants | 0.40% |
352: | **Fairness Gap** | **86.77%** |
$ grep -rFn "58.15" . ; grep -rFn "28.37" . # only appear in counterfactual-fairness.md and its generated mirrors
Why it matters: this is the repo's single most-cited statistic (the flagship COMPAS bias figure, quoted identically across 8+ other files). counterfactual-fairness.md alone presents an entirely different, unsourced pair of numbers - any reader comparing explainers will hit a direct contradiction on the repo's headline result.
Suggested fix: replace the 58.15%/28.37%/29.78pp table with the canonical 87.16%/0.40%/86.77pp figures used everywhere else in the repo, or, if a genuinely different metric/subset was intended, clearly relabel the section to say so.
Where:
explainers/counterfactual-fairness.md(~lines 63-69) - the COMPAS "biased model" example table.The gap: the file cites Black defendants 58.15% / White defendants 28.37% / a 29.78pp fairness gap. This repo's actual, canonical
unfair.pyresult for COMPAS - cited identically and consistently by README.md and at least 8 other explainers (automation-bias.md,confounding-variable.md,data-leakage.md,how-ai-detects-patterns.md,ml-bias.md,proxy-variables.md,ai-objectivity-myth.md, and more) - is Black 87.16% / White 0.40% / an 86.77pp gap.Repro:
Why it matters: this is the repo's single most-cited statistic (the flagship COMPAS bias figure, quoted identically across 8+ other files).
counterfactual-fairness.mdalone presents an entirely different, unsourced pair of numbers - any reader comparing explainers will hit a direct contradiction on the repo's headline result.Suggested fix: replace the 58.15%/28.37%/29.78pp table with the canonical 87.16%/0.40%/86.77pp figures used everywhere else in the repo, or, if a genuinely different metric/subset was intended, clearly relabel the section to say so.