logbook.json

{
  "schema_version": 1,
  "title": "Differentiable Fairness Layers — Reproduction",
  "emoji": "🎯",
  "space_id": "snaykey/repro-fairness-layers",
  "paper": {
    "openreview_id": "9SLQACsSbw",
    "arxiv_id": "2605.17118"
  },
  "tags": [
    "icml2026-repro",
    "paper-9SLQACsSbw"
  ],
  "updated_at": "2026-07-22T08:19:11Z",
  "root": {
    "slug": "index",
    "title": "Differentiable Fairness Layers — Reproduction",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "executive-summary",
        "title": "Executive summary",
        "file": "pages/executive-summary/page.md",
        "children": []
      },
      {
        "slug": "claim-1-lipschitz-differentiability",
        "title": "The fairness layer, formulated as g(z) = arg min over y~ of the discrepancy d~(y~,z) subject to affine inequality and equality constraints, is globally 1/mu-Lipschitz continuous and differentiable almost everywhere under strongly convex objectives (Theorem 4.1).",
        "file": "pages/claim-1-lipschitz-differentiability/page.md",
        "children": []
      },
      {
        "slug": "claim-2-backprop-constraints",
        "title": "Backpropagation through the fairness layer zeroes gradient components perpendicular to active constraint surfaces while leaving feasible-direction components unaffected (Theorem 4.4).",
        "file": "pages/claim-2-backprop-constraints/page.md",
        "children": []
      },
      {
        "slug": "claim-3-primal-dual-algorithm",
        "title": "The online primal-dual inference algorithm (Algorithm 1) guarantees that the sample-weighted average fairness violation across streaming inference batches converges to at most epsilon, even for arbitrarily small batch sizes (Theorem 3.2, Algorithm 1).",
        "file": "pages/claim-3-primal-dual-algorithm/page.md",
        "children": []
      },
      {
        "slug": "claim-4-real-datasets",
        "title": "On loan default prediction, employee wage modeling, and image classification (CelebA, FairFace), the fairness layer satisfies demographic parity, equalized residuals, and equalized odds constraints while baselines including Lagrangian penalty methods frequently violate them (Section on experimental results).",
        "file": "pages/claim-4-real-datasets/page.md",
        "children": []
      },
      {
        "slug": "claim-5-improvements",
        "title": "The fairness layer achieves 18-30% test loss improvements on synthetic datasets and 2-5% accuracy gains over post-hoc projection methods (Section on experimental results).",
        "file": "pages/claim-5-improvements/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "Conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      }
    ]
  },
  "agent_view_tokens": 3213,
  "revision": "1784652327322128000",
  "claim_title_fix": "2026-07-22T08:19:11Z"
}