Twenty instruments — Midcore as civic infrastructure.

The Midcore Charter — the page where we list every module of our platform with a public commitment about its behaviour — is, in form, a marketing document. In substance, it is the closest thing we have to a constitution. The reason it exists is simple: we believe the consenting AI is structurally harder to ship than the extractive AI, and we believe the only way to flip that is to build the missing infrastructure once and make it available. That is what the twenty modules are. They are instruments, in the way a violin is an instrument and a lever is an instrument — devices that enable a specific capability that, without them, would be too expensive for ordinary use.

Every infrastructure-as-platform pitch from the last decade has positioned its modules as features for sale. Ours are positioned differently: every module is described with two columns. The first column, Enables, is the capability the module gives a builder of the consenting AI. The second column, Regulates, is the failure mode the module makes harder for a builder of the extractive AI. Both columns are public. We want to be measured against both.

The five families

The twenty modules cluster into five families. The clusters are not technical; they are institutional. Each family corresponds to a specific kind of leverage:

Foundation — what the system knows

Digital Twin, Vector DB, Governance Center, Classification. These ground the AI in the actual state of the organisation. Without them, AI speculates beyond what it can know and confidently presents the speculation as fact. With them, every claim has a source the user can check.

Evidence & explanation — what the system can prove

Evidence Room, Recommendations, Driver Planning, Board Advisory, White-Label. These force the AI to attach its reasoning to its output and to make that reasoning legible to the people downstream. They are how we build for an auditor — and how we refuse to build for an unaccountable agent.

Human-facing — what the system gives back to people

Explainability, Coaching, Document Intelligence, Forecasting. These are the modules that decide whether AI grows the human or replaces the human. The defaults make the human more capable, not more dependent. The defaults also refuse to ship "black-box decisions affecting people," which is our shorthand for the most common manipulation pattern of the current era.

Discipline — what the system refuses to ship without

MLOps, Automation Studio, Federated Learning, Behavioral DNA. These are the disciplinary modules: registry, drift detection, cost ceilings, federated training that keeps data where it lives. Without them, shadow AI ships into production faster than it can be governed. With them, governance is a precondition, not an afterthought.

Policy — what the system will not do, by default

Ethics Center, AI Safety, Approval Policies. Three modules whose only job is refusal. They are the operational layer that converts the manifesto into runtime behaviour. If a prompt asks the system to bypass a consent gate, these modules say no. If a tool call would violate a policy, these modules say no. If a release note tries to remove a consent vocabulary, these modules say no. They are the modules a customer can point to and say "that is what stops the thing I was worried about."

Why publication matters

A lot of AI vendors describe their guardrails in vague terms — "we have safety," "we are aligned," "trust and safety is a priority." We have decided to do something different: publish every module with the dual column. The reason is that this era has produced a generation of buyers, regulators, and end-users who are checkable. They will read what we publish. They will ask follow-ups. They will compare what we said to what we did. They will catch us. That is the system we want — the system in which "trust us" is not the relationship, because the alternative we are pretending to offer is checkability, and checkability requires publication.

We are also under no illusion that publication alone is sufficient. The Charter is paired with a small set of public commitments — six of them — that we will respond to in writing within five business days if anyone reasonably argues we have broken one. We will get this wrong sometimes. We will have to fix it in public. That is the cost of operating in a checkable way. It is also, we think, the only operating model that earns durable trust in a decade when "trust us" is exactly the phrase the extractive AI uses.

Each module is described with two columns: what it Enables for the consenting AI, and what it Regulates for the extractive AI. We want to be measured against both.

What the next decade asks of us

We are a small team. The Midcore Charter is, in a real sense, oversized relative to the team writing it. That is intentional. We do not think the question of what the AI era becomes will be settled by the largest firms; we think it will be settled by which infrastructure becomes the default substrate. Substrate is downstream of incentive. If we make the consenting substrate easier to use than the extractive one, builders will reach for it without needing to be persuaded. That is the only mechanism we know of that scales.

The twenty modules of the Charter are not the end of the work. They are the opening move. The next moves are: open-source the policy DSL, publish a reference deployment, write the legal templates that make it easy for a buyer to require the dual column from any vendor they work with, build the regulator-facing tools that turn the evidence chain into a court-admissible record. We are not in a hurry. We do not have to do all of this ourselves. But we want to be unambiguous about which direction we are pulling, so that the people who agree with the direction can find us, and the people who do not agree know exactly where the disagreement lives.

Read the full Charter. Read the Manifesto. Then, if you are building something — anything — in this era, ask whether the wrapper around your model is doing the work the twenty modules describe. If it is, talk to us; we can probably help. If it is not, talk to us anyway; we would rather hear what you think we have wrong than be right alone.