Haimesian Self‑Purifying Algorithm (HSPA)

Pattern

1. Confess (Detect): continuous fairness audits (DBR, DLI, EG).
2. Make Amends (Repair): reweighting, counterfactual augmentation, fairness constraints.
3. Explain (Witness): publish “Harm & Repair Notes” to a provenance ledger.
4. Resolve (Guard): carry forward non‑recurrence constraints across retrains.

Key Metrics

• DBR: Disparate Burden Ratio → 1.0
• DLI: Dignity Loss Index (false denials, high‑stakes weighted)
• EG: Exposure Gap (harmful exemplars without corrective context)
• XS: Explanation Sufficiency
• MM: Mercy Margin (avoid over‑correction)

Governance

• Diverse rotating panel, quorum for policy changes
• Open provenance hashes (datasets, metrics, notes)
• Kill‑switch outside the model team

while training:
    y_hat = model(x)
    loss = task_loss(y_hat, y)

    metrics = fairness_audit(y_hat, y, groups)
    if metrics.exceeds():
        weights = reweight(x, y, groups, metrics)
        x_cf, y_cf = counterfactual_augment(x, y, groups, metrics)
        constraints = update_fairness_constraints(metrics)

        loss += fairness_penalty(y_hat, y, groups, constraints)
        x, y = concat((x, y), (x_cf, y_cf))
        write_provenance(harm_repair_note(metrics, constraints))

    step(loss)