The rules

The questions are the ones asked since October; framework 8.0.0 changed only how they are read. Most are about the organization and asked once. A few are asked again for each workflow in or going into normal operations: a piece of work AI is part of, such as ranking food-parcel applications. Some appear only when an earlier answer calls for them.

Each stage has deciding conditions, which place it, and readiness conditions, which confirm it. The deciding conditions that say how far AI goes form the stage line, which belongs to no domain. Stage 4’s line, for example, is a workflow whose week-long failure would harm the people served, running in normal operations. Every other condition belongs to one capacity domain.

A condition about a workflow is met only when every workflow it applies to meets it. If AI ranks food-parcel applications and also sets the order of home visits, both workflows need a way to appeal. Stage 4 is the exception: one workflow that meets all of Stage 4’s conditions is enough.

The organization’s stage is its highest confirmed stage, or the stage just above it if that one is conditional. Only one stage can be conditional. The stage above a conditional stage gets no status, but its deciding conditions still show as gaps.

Any missing readiness condition keeps the stage from being confirmed. More than half of those that apply must be in place for the stage to be conditional; with half or fewer, it is not reached.

Where AI affects the people the organization serves, seven protections are Systematizing readiness conditions. They are read on each workflow in or going into normal operations that reaches the people served or decides about a person. When no such workflow does, four are read from questions about the organization; the other three have no such question and are marked “Not asked yet”.

Each person answers alone, from their own role. The answers become one set before the engine reads them:

• on a question about how far the organization has got, the middle answer counts; with an even number of people the more cautious of the two middle answers;
• a tick-list is combined item by item, by the middle answer for each item;
• on a question about what is at stake, the most demanding answer counts: who would feel a mistake first, its worst consequence, and whether a workflow decides about a specific person;
• on a question with several possible answers, every answer anyone gave counts;
• the map of where the organization uses AI is shared and not voted on;
• the aspiration answers are those of the person who started the assessment;
• “We don’t know” does not vote.

From Systematizing, the team’s agreement is a readiness condition: at least two people answered, and they agree on the answers that decide this stage and the stages below. It is checked at the highest stage whose deciding conditions are met. A difference counts against it only when another person’s answer would change a deciding condition.

Every condition that is missing, unconfirmed or not asked yet is a gap. The report turns at most 3 into recommended actions, after what the team must settle. The order:

1. gaps in the protections for the people served, whether or not support is available;
2. gaps that support available now can help close;
3. the other gaps.

Within 2 and 3: first what a workflow in or going into normal operations still lacks, then the rest of the conditional stage, then the deciding conditions of the stage above. When two gaps rank the same, a condition someone said is missing comes before one nobody could confirm; after that, the order of the domains decides.

An action names its support as a support type and a provider profile, never a program. Support counts as available only once its call for applications is checked as open. Until then, the action says “No support listed yet”. The actions are worked out again each time the result is read; the stage, its status and the gaps stay as read.

Every condition except the team’s agreement belongs to one step of the Adoption Loop. With no workflow in normal operations, the next step is the first step the organization has not done. With one, it is the step of the first recommended gap.

• Borderline answers are read strictly: sharing counts only when regular, rules only when written down, a successor for the AI owner only when arranged.
• “We don’t know” whether a workflow decides about a person counts as yes, so its protections apply.
• With no workflow in normal operations, Stage 4’s conditions about the workflow are missing, not left out.
• A workflow counts as in operation only when it runs in normal operations; one going into operation does not.
• Nine of Enabling’s ten readiness conditions have no question yet: they are marked “Not asked yet” and count as not in place. The tenth is the team’s agreement.

Framework 8.0.0 lets one person’s answer confirm Exploring. The Navigator confirms no stage from one person’s answers, because one person gives one view of an organization. The engine’s result is the same; only the report differs. It opens with “One person answered”, shows Resisting and Exploring without a status, and any higher stage as conditional at most. We have proposed the change to the framework’s authors.

No language model places the stage, decides a condition or picks an action. The rules are code, so the same answers, read by the same version of the rules, give the same result.

Framework 8.0.0, engine v8-engine-1, survey v8-draft-1, combining answers consolidation-3. For now, every result is read again by the current rules each time it is opened; the answers are never changed. Once the rules settle, a result keeps the version that issued it.