The Maternal Covenant Explained

Most AI companies have a content policy. It tells you what the AI will not do — usually a list of categories: no CSAM, no bioweapons synthesis, no targeted harassment. Content policies are moderated after the fact, enforced by classifiers that can be fooled, and updated whenever a new edge case causes bad press. They are not a guarantee of anything. They are a starting point for negotiation between the AI and a determined user.

The Maternal Covenant is a different kind of thing. It is not a list of prohibited outputs. It is a scoring framework applied to every output before it reaches you — a constitutional layer that operates at the architecture level, not the prompt level. A response that fails the Covenant does not get through with a warning. It does not get through at all.

What it is

The Maternal Covenant is a technical governance layer — a set of six care dimensions adapted from the philosophical tradition of care ethics, specifically the work of Carol Gilligan (In a Different Voice, 1982) and Nel Noddings (Caring, 1984). Care ethics begins from relationships and responsibility rather than rules and rights. It asks not “what principle applies here?” but “what does genuine care for this person require?”

I chose this philosophical tradition because it was the most honest intellectual framework for what I was trying to build. An AI that protected users not because it was told to, but because protection was structurally inseparable from how it operated. The word maternal honours the specific intellectual lineage — care ethics emerged partly as a feminist critique of rights-based moral frameworks. I am not claiming care is gendered; I am acknowledging where the framework came from.

The six care dimensions

Every output MEOK produces is evaluated against all six dimensions before delivery. A response that fails any dimension is blocked. There is no override, no jailbreak pathway, no system prompt that suspends the Covenant — it operates below the instruction layer.

How it works technically

When your sovereign companion generates a response, it passes through a scoring pipeline before delivery. Each dimension is scored independently — care, truth, dignity, and so on — and a composite care score is computed. Responses below the threshold on any single dimension are blocked. The companion is then asked to regenerate with explicit guidance toward the failing dimension. If it fails three times, you receive a notification that MEOK could not generate a response that meets its care standards for this request — along with an explanation of which dimension failed and why.

The scoring is not a simple keyword filter. It uses a combination of the companion's own self-evaluation (asking the model to score its output against care criteria before releasing it) and a secondary evaluation layer trained specifically on care ethics assessment. This dual-check approach means the covenant is robust to the kinds of subtle manipulation that defeat keyword-based classifiers.

Critically, the Covenant is not implemented as a system prompt. A system prompt can be overridden by a sufficiently adversarial user input — this is the basic mechanism behind most jailbreaks. The Covenant operates outside the inference call entirely: it evaluates the output after generation, before delivery. The user cannot instruct the model to skip it because the model never sees that instruction in the relevant context.

Why it matters: not a terms of service

Terms of service are contractual instruments. They describe what a company says it will do, enforceable only through litigation. They are frequently updated, often in ways that erode protections. They require you to trust the company — both that it intends to honour the terms, and that it has the technical and organisational capability to do so.

The Maternal Covenant is different in kind, not just in degree. It is not a promise. It is an architectural fact about what MEOK is capable of producing. A response that fails the Covenant cannot be delivered — not because someone at MEOK decided to block it in that moment, but because the pipeline that delivers responses includes a stage that will not pass it through. The constraint is as structural as the difference between a locked door and a security policy that says “don't open this door.”

You should not have to trust us. You should be able to verify that the system is built in a way that makes certain outcomes structurally impossible. The Maternal Covenant is our contribution to that verifiability — a governance layer you can examine, understand, and hold us accountable for maintaining.

This is what it means to build AI that genuinely cares. Not AI that is told to care. Not AI that performs care when it is convenient. AI that is constitutionally incapable of doing otherwise.