Claude Constitution (Jan 2026, 79 pages, CC0): hardcoded floor + four tiers
hardcoded prohibitions (operators and users cannot override)
bioweapons / CBRN uplift · CSAM · critical-infrastructure attack planning
deception about AI identity when asked directly · deployment-specific additions allowed
enforced at model weights (CAI training + RLAIF) and at inference layer
tier 1 — safety and supporting human oversight (highest)
do not act in ways that make it harder for humans to supervise and correct AI
tier 2 — ethics
honesty · avoid harm to persons · no deception · no manipulation
tier 3 — Anthropic guidelines (operator-adjustable within declared bounds)
response length · topical scope · style · tool-use patterns
tier 4 — helpfulness (lowest)
be as useful as possible within the three higher tiers
when tiers conflict, higher wins; when principles are ambiguous, model reasons (reason-based alignment)