15 scenarios, about 11 minutes, five dimensions scored the same way for everyone. What follows is the reasoning behind each of those choices, and an account of what has been established about this instrument and what has not.
Fixed, and the same for everyone. They are not weighted: a profile is five separate readings, and the index is the plain mean, so no dimension quietly counts for more than another. Each carries 3 of the 15 items.
A dimension score is the mean of the 3 answers given for it, so it lands on thirds rather than whole numbers. The band is the word attached to that number. The line between bands is not how much AI someone uses, it is how predictable the result is and whether they know where it stops.
You have not yet worked out where this applies to you, so the risk it carries is unmanaged rather than accepted.
You are using AI here and finding out what it does. Results are real but inconsistent, and you cannot yet predict which you will get.
You use this reliably for real work and know where it stops. This is the level at which AI is genuinely doing something for you.
You direct this deliberately, check it where checking matters, and have built it into how work gets done rather than doing it by hand each time.
The step that matters most is Exploring to Applied. Below it results are real but inconsistent, and a person cannot predict which they will get. At and above it the work is dependable and its limits are known.
Every person answers the same 15 scenarios, and every answer is scored the same way. Using the same fixed form removes variation caused by people receiving different questions. A difference between scores may still reflect changed judgment, familiarity with the items, feedback, or other factors.
A conversational assessment cannot make that guarantee. If the questions adapt to the person, two people have not taken the same measurement, and the gap between their scores contains the effect of whatever the model chose to ask. That may still be useful as coaching. It is not a measurement you can aggregate, compare between departments, or track round over round, which is the whole of what this instrument is for.
The cost is real and worth stating. A fixed form cannot follow up on an interesting answer, and 15 items cannot cover everything inside a dimension. Comparability is bought with that depth.
Option order is permuted per run. The items are fixed; where an option sits on screen is not, so position carries no meaning and a group comparing notes afterwards cannot reconstruct the key from each other’s answers.
Verification asks what someone does with output before they act on it, and it is scored in both directions. Accepting a confident-sounding answer without checking the part that carries the risk scores low. So does checking everything.
The highest-scoring answer demonstrates judgment, which is frequently not the most thorough or the most cautious option. Checking or controlling more than the situation needs is scored below matching the effort to it.
Over-reliance shows up as a preference for the answer that trusts the tool where trust is not warranted. Its mirror, checking exhaustively, looks careful and is treated as a failure of the same judgment, because a person who verifies everything has not worked out which claim is load-bearing and will not sustain the habit. What scores highest is triage: identifying the one claim that would matter most if it were false, and checking that.
This is also the dimension a model provider is least able to grade about its own product, since the thing being measured is how much a person should rely on it.
If any dimension comes out at Unaware, the headline band is held at Exploring however high the mean is. The number is never altered. Only the word next to it changes.
Four dimensions at Fluent, one at Unaware. The mean is 3.4, which would round to Applied. The headline reads Exploring.
A mean averages five things and hides its worst component. Someone strong on four dimensions and Unaware on Verification would otherwise be labelled Applied, reading as broadly capable while being unable to tell whether the output in front of them is true. That is the exact failure this instrument exists to surface, so it is not averaged away.
The rule is deliberately narrow. It triggers only at Unaware, and it holds at Exploring rather than dropping to the floor, because the strength elsewhere is real and the profile still shows it. It applies to whichever dimension is lowest, not to Verification alone; where two are tied at the bottom, Verification is the one named, so a report cannot call Verification the weakest dimension while blaming the cap on something else.
The evidence pack records what was measured and identifies the areas of the frameworks below that the measurement is relevant to. Each is named with the qualifier that travels with it.
This document records measurement results and identifies the framework areas that those results are relevant to. It does not assess or assert compliance with any law, regulation, bulletin or framework, it does not state that any of them require this or any product, and it is not legal advice. Determining what any framework requires of this organization is a matter for its own counsel.
About 15 hours of live, hands-on training a year for each person in a training group, built on their own work and measured every quarter. Each quarter, a training group of up to 30 people works on its two weakest dimensions below Applied: two 90-minute hands-on workshops a few weeks apart, then a 45-minute practice clinic on real tasks from their own jobs. A re-score at 90 days picks the next focus.
The focus is read from the data, not chosen by us: the two dimensions furthest below Applied, one when only one is below, and a stated tie for leaders to resolve when the second pick is unclear. When no dimension is below Applied, training still continues each quarter on the team’s two lowest dimensions, at the Applied level, working toward Fluent. Leaders choose who is in each group; the dashboard never picks people. Results are shown only for groups of five or more.
A re-score shows whether answers changed, not what caused the change. The same fifteen items are asked again, so familiarity and the training push in the same direction, and no claim is made that the training moves a score. The results are not for hiring, promotion, pay or termination decisions.
Every organization account includes an optional work task after the assessment, unless the organization has asked for it to be switched off: the taker reviews an AI-drafted memo with a supplied AI tool, explains the decision to the person who signs, and says how they would do it next time. It does not affect your score. Your assessment is complete before it starts, and nothing in it changes your result or any figure your organization sees. Its purpose is research: to learn whether people whose scores differ also work differently. It is not a validated measure of performance, and it is not presented as one.
The instrument is built and in production. Answer-position and length bias have been identified and corrected. On content validity, precisely: reviewers have agreed to key the items independently; the review pack is held until the current item revision is final, so that they key the version that ships rather than one about to change; and no keyings have been returned yet. Until they are returned and their disagreements reconciled, every level in the answer key is still the author’s own. We publish who the reviewers are when their keyings are complete, and only with their consent — agreeing to review an instrument is not the same as agreeing to be listed as endorsing one. Calibration data publishes once one hundred organic results have been recorded, and not before. We would rather show you the timeline than borrow someone else’s study.