Specify — Constitutional AI

Lean instance: caiConstitution (constituency, extrapolation, aggregation all true).

Anthropic’s Constitutional AI states principles (the constitution) and iterates training with AI-generated feedback (RLAIF). The specify-side content is the principle set and who they apply to; extrapolation and multi-value structure are flagged true in Lean to match the correction-channel and bundle slots on the resulting AlignmentTarget.

Construction bet: RLAIF stack — caiConstructionBet with claimsExplicitBuilder = true.

Matrix evidence on this agenda often tags identify/certify (MB2, MB7) rather than a specify column — the table separates schema placement from where field evidence landed.

Instance table: /cards/alignment-target/#specify-construct-instances.