Constitutional AI Defines Principles Without Cognitive Architecture

Vendor and product reality

Anthropic's safety stack is integrated and public. Constitutional AI generates harmlessness training signal by having a model critique and revise its own outputs against a written constitution; the resulting RLAIF pipeline reduces dependence on human red-teaming throughput while making the alignment target legible. The constitution itself is published and revised; the Claude models trained against it are evaluated on harmlessness, helpfulness, and honesty benchmarks; the Acceptable Use Policy defines deployment-time prohibitions that bind both Anthropic and its API customers; and the Responsible Scaling Policy commits the company to capability-thresholded mitigations as models cross AI Safety Level boundaries. Together, the four artifacts — constitution, RLAIF training pipeline, AUP, and RSP — describe one of the more coherent end-to-end safety methodologies any frontier lab has published.

The methodology is effective at what it sets out to do. Claude exhibits measurable improvements in harmlessness over comparable preference-trained baselines, refuses categories of requests that the constitution targets, and behaves consistently across the surface area the AUP polices. The argument here is not that CAI fails on its own terms; it is that its terms operate at training time and at prompt time, and that there is a category of safety property — runtime cognitive integrity, cryptographically bound to output — that the methodology, by construction, does not produce.

The architectural gap

A constitutional principle that the model should be honest is a training objective. A cognitive architecture in which integrity is a persistent runtime state variable that tracks behavioral consistency, detects deviation, and triggers self-correction embodies honesty structurally. The first can be trained to approximate honest behavior on the distribution of evaluation prompts; the second cannot produce dishonest output without its own integrity-tracking primitive registering the deviation as the output is being produced. The architectural difference is between compliance with a target and constitution in the architectural sense — a structure whose interactions yield the principle as a consequence rather than as a learned approximation.

Human cognition does not work from a checklist. A person who is honest is not consulting a principle each time they speak; honest behavior is the consequence of integrity, empathy, self-esteem, and emotional state interacting in ways that make dishonesty cognitively costly. The cost is not paid because a rule was looked up; it is paid because the cognitive architecture is structured such that incoherence between a self-narrative and an action is registered as a state perturbation that propagates back into behavior. Constitutional AI trains a model to imitate the surface of this behavior. It does not install the architecture that produces it.

The second gap is the absence of cryptographic runtime binding. Once a model is deployed, the linkage between the principles it was trained on and the outputs it produces is purely behavioral. There is no mechanism by which a relying party — a regulator, an auditor, a downstream agent — can verify, at the moment of consuming an output, that the output was produced by a process whose runtime cognitive state was coherent with the declared constitution. Prompt injection, jailbreaks, and out-of-distribution inputs can move the model into regimes where its training-time alignment no longer governs its behavior, and there is no runtime signal — no cryptographically bound attestation of integrity state — that surfaces the divergence. The AUP polices deployment context; the RSP polices capability thresholds; neither produces a runtime-verifiable binding between principle and output.

What the human-relatable-intelligence primitive provides

The ten conditions for human-relatable intelligence define when a computational system is structurally isomorphic to human cognition. The coherence control loop maintains internal consistency across domains. The narrative identity primitive provides continuity such that the system at time T is the same system, in the cognitively relevant sense, as the system at time T-minus-N. The architectural inversion makes governance the foundation rather than an overlay: principles are not a layer that wraps a model, they are parameters within the cognitive architecture itself, and their satisfaction is a structural property of the architecture's operation rather than a behavioral consequence of training.

Crucially, the runtime state of the architecture is exposable and bindable. Integrity-tracking and coherence-control state can be cryptographically committed alongside outputs, producing attestations that bind a generated output to the cognitive state under which it was produced. A relying party can then verify not only that an output came from a particular model but that, at the moment of generation, the model's runtime cognitive integrity was coherent with the declared constitution. This is the runtime binding that CAI structurally lacks; it is the difference between a principled training signal and a principled inference event.

Composition pathway with Constitutional AI

The primitive composes with CAI rather than replacing it. The constitution remains the declarative source of principle; RLAIF remains an effective method for shaping the substrate model's behavioral prior; the AUP remains the deployment-context policy; the RSP remains the capability-threshold gating regime. The primitive supplies what each of those layers does not: a runtime cognitive architecture in which the constitution is parameterized and continuously evaluated, and a cryptographic binding mechanism that makes the runtime state auditable.

In an integrated deployment, a Claude-class substrate model produces candidate outputs; the human-relatable-intelligence layer evaluates those outputs against the runtime cognitive state instantiated from the published constitution; coherence-control signals trigger correction or refusal when the runtime state diverges from constitutional parameters; and the resulting output ships with an integrity attestation that downstream relying parties can verify. The methodology Anthropic has built becomes the substrate; the primitive becomes the runtime layer. Customers operating under regulatory regimes that require auditable AI behavior — financial services, healthcare, public-sector — gain the runtime evidence that their compliance posture requires, without abandoning the alignment investment that CAI represents.

Commercial and licensing

Adaptive Query licenses the human-relatable-intelligence primitive to foundation-model developers and to enterprise deployers building products on constitutionally-aligned models. For organizations whose deployments are bound by an Acceptable Use Policy, a Responsible Scaling Policy, or an analogous commitment regime, the primitive provides the runtime evidentiary layer that converts policy commitments into auditable runtime properties. Licensing is per-deployment with reference integrations for major model-serving stacks and for agentic frameworks; partnership terms covering joint go-to-market with foundation-model providers are available separately. The commercial structure is engineered so that the cost of integration is recovered through the regulated-market deployments that runtime-bound alignment enables, rather than through pure compliance overhead.