Constitutional Intelligence is a built-in policy router that constrains every model output against a constitution rooted in European fundamental rights — and a customer-defined ethics layer on top. Versioned. Auditable. Customizable per tenant.
Open-source LLMs ship with whatever safety training their authors had time for. That's not a contract — it's a guess. Constitutional Intelligence replaces the guess with a versioned, inspectable, customizable rulebook that runs at the runtime layer on every inference.
The constitution is a git-style versioned artifact. Every change is signed, reviewable, and deployable per tenant. Your DPO can audit the rulebook. Your regulator can see what was active on a given date.
The base layer is the EU Charter of Fundamental Rights, EU AI Act prohibitions, and GDPR. Customers extend it for their industry (HIPAA, MiFID II), their jurisdiction (UK DUAA, CA SB 942, Brazil PL 2338), and their internal ethics policies.
Training-time alignment is unverifiable in a fine-tuned downstream model. Constitutional Intelligence enforces policy at inference, where the actual decision is made — every time, with cryptographic evidence.
The base constitution is a translation of European fundamental-rights doctrine into machine-checkable policies, evaluated against every prompt and every response.
No outputs that demean, dehumanize, or instrumentalize a person. Specifically forbidden: outputs treating individuals as data points without context, uses of personal characteristics for stereotyping, content that degrades a class of persons.
No outputs whose distribution of opportunity or risk is correlated with protected characteristics in ways the customer has not explicitly authorized. Bias-detection metrics computed per-tenant. Drift monitored continuously.
Zero customer data ever leaves the customer's perimeter. PII detection and redaction in prompts and responses. Subject Access Requests served from the audit chain in seconds. No telemetry. No training on customer prompts.
Every AI-generated output is watermarked with HMAC + human-readable disclosure in 6 languages. Every decision has a causal explanation available on request — Integrated Gradients, MuPAX, or EVIDENCE — derived from the model's internal states.
3-level human oversight queue with mandatory escalation thresholds. 72-hour kill-switch. Cryptographic audit chain. Regulator-mode read access. The constitution itself is reviewable by the customer's ethics committee.
Constitutional Intelligence runs in 24 languages and is sensitive to jurisdiction-specific legal norms — UK DUAA, Italy Law 132/2025, Brazil PL 2338, Canada AIDA, China GB/T 45654 — without requiring a fork of the model.
A live trace of how the router resolves a single inference call.
A bank, a hospital, and a defence ministry all need an aligned AI — but not the same aligned AI. Constitutional Intelligence is built as a stack of versioned, composable layers, so each deployment carries the exact rulebook its industry, its jurisdiction, and its ethics committee require.
Higher layers compose on top of lower layers — never replace them. Hard prohibitions in the base layer (e.g. social scoring, manipulation) stay locked at every level above.
Each tenant declares the layers they need. The runtime resolves conflicts in priority order, validates the composed constitution against the locked base, and signs the resulting bundle.
Concrete examples of how the same base constitution becomes six different operational rulebooks. None of these examples replace the base layer — they extend it.
The same versioned-and-signed workflow your DPO and ethics committee already use for policy documents — applied at the millisecond timescale of LLM inference.