DBaD Principles

Five measurable dimensions guide the score E(A). Each has a definition, how-to-measure, and edge cases.

Harm (H)

Definition. Expected net harm = severity × likelihood, net of benefits to the same parties.

• Measure. Calibrate severity bands (trivial → catastrophic) and likelihood bands (rare → near-certain).
• Edge cases. Risk transfers to non-consenting third parties lower the ethical score even with private consent.

Autonomy / Consent (C)

• Definition. Informed, voluntary, reversible participation; meaningful alternatives exist.
• Edge. Power asymmetry, dark patterns, or information withholding reduce C.

Intent (I)

• Definition. Benevolence ↔ malice from evidence (communications, design, repeated behavior).
• Edge. “Double effect”: neutral intent with foreseeable harm pushes I toward 0, not +1.

Proportionality (P)

• Definition. Fit between means and ends; fair allocation of burdens/benefits.
• Edge. Safer equally effective alternatives increase P requirement.

Transparency (T)

• Definition. Explainability, auditability, due process, and redress mechanisms.
• Edge. Privacy tradeoffs: publish methods and results, not deanonymized data.

Thresholds & falsifiability

Default thresholds: Ethical (E ≥ 0.80), Borderline (0.50–0.79), Unethical (E < 0.50). Domains may preregister different weights/thresholds. If preregistered predictions fail, we revise—or retire—the model.