The AI Delivery Contract

AI value is a lifecycle property,
not a model property.

Taking AI into production is not about picking a clever model. Value is created or destroyed at every stage, from data to switch-off, and governance is the thread that runs through all of it. The model is the easy, replaceable part. The contract around it is the product.

No code required. Click around. Everything on this page runs in your browser.

The lifecycle, as a contract between teams

A good architecture states what exists, what does not, and who is responsible. The whole journey is four stages, with a governance "spine" running across all of them. Click any stage to see what has to be true before work can move on.

GOVERNANCE SPINE  ·  one register of every AI system · evidence for every claim · a human who can stop it · one accountable owner
each step is a gate: work cannot pass until the condition is met and the evidence exists

The big idea: enforce the rules like code

Most "AI governance" is a document nobody checks. This project ships a small program that checks the rules automatically, the same way a test blocks a broken build. The headline rule is fairness. Below is the exact check the project runs, made visual. Flip the switch and watch what happens.

Who gets advanced, by group

The four automatic checks

What you are looking at. The dashed line is the fairness bar (a long-standing legal rule of thumb: the lower group must reach at least 80% of the higher group). When a change quietly makes the assistant favour one group, Group B drops below the line, the fairness check fails, and the change is refused before it can ship. The rule was decided in advance and is enforced by the machine, not by good intentions.

One contract, different weight for different systems

The same contract applies whether AI advises a human or acts on its own, but the risk lands in different places. Switch between the two and see where the attention goes.

What goes wrong, and what catches it

A reference earns trust by naming the failures and the control that owns each one. None of these are model problems. They are lifecycle problems.

What goes wrongWhat catches itStage

Dig deeper

This page is the doorway. The repository has the full reference, the reusable templates, two worked examples, and the runnable check itself.

Run the fairness check yourself: clone the repo and run make gate, then ADC_INJECT_BIAS=1 make gate to watch it fail.