Why does the number of questions change when I change who asks?

Because two controls have no surface to exist on where nobody outside asks. There is no external deadline you could have met and no outside reader who could have failed to follow your record, so those two carry a weight of zero for that shape and drop out of the divisor entirely. It is not a free pass: the remaining controls are weighted heavier, so freshness, drift and capture carry more of the score than they do for anybody else.

Why are the basics weighted heavier when nobody outside asks?

Because there is nothing else holding them up. An organization with an audit date in the diary has a forcing function doing part of this work whether or not anybody planned it that way. An organization with no external deadline has only its own discipline, so the controls that measure discipline are carrying more of the answer, and the weights say so.

Where do the bands come from?

From a product RedHub already ships in this lane. The AI Output Audit-Trail and Record-Keeping Kit puts its top band at 80 and its floor at 50 on a weighted score of record-keeping durability, which is a narrower cut of this same subject. Choosing different thresholds by feel would have made a free assessment disagree with a paid one it recommends. The names differ: that product grades a log it reads, and this one reads what you told it, so its top verdict carries AS DESCRIBED.

Why can a gate only ever lower a verdict?

Because a weighted average is the wrong instrument for a structural failure. Twenty strong controls can carry a score while the one thing that would have kept the whole picture alive is missing, and averaging is exactly what hides that. A gate is where one specific fact overrides the summary, and letting it work in the other direction would make it a bonus rather than a limit.

How is this different from the AI compliance gap assessment?

That one measures distance from a named framework and reports what is missing today. This one assumes nothing about what is missing and asks what will still be reproducible later. They can disagree in both directions, which is the test of whether they are two instruments rather than one twice: a strong framework posture can be undated, unwritten and known to one person, and a thin posture can be dated, rehearsed and reproducible by somebody else.

Can I check the score by hand?

Yes, and that is why every weight is on this page. Multiply each mark by its weight, divide by three, sum them, divide by the weight in scope and multiply by a hundred. The result is deterministic: the same answers always produce the same score, the same band and the same first thing to fix.

How this assessment works

The assessment marks up to 24 controls from 0 to 3 on what is true today, weights them for who asks you to prove any of it, and renormalizes over what is in scope. The result is a score out of 100 and one of three verdicts. Three gates can hold a verdict down regardless of the score, and none of them can lift one.

It is deterministic. The same answers always produce the same score, the same band and the same first thing to fix. There is no model in the scoring path, which is why the whole method fits on one page.

The arithmetic

Each control carries a weight for each of the three shapes, and the weights for a shape sum to exactly 100. You mark every in-scope control from 0 to 3 for what is true today. The score is the weighted marks divided by the weight in scope, as a percentage, rounded to a whole number.

earned  = sum(mark / 3 * weight)   over in-scope controls
inScope = sum(weight)              over in-scope controls
score   = round(earned / inScope * 100)

Marking a control not applicable removes its weight from the divisor instead of scoring it zero, so declaring something out of scope never costs you points. Controls your shape has no surface for were never in the divisor at all: where nobody outside asks, you are not graded on having met an external deadline, because there was none to meet.

Five controls are the exception and cannot be marked not applicable, and the assessment says so beside each one. They are the five the gates read. Nobody running an assurance assessment can truthfully say that a review date, a held schedule, a written procedure or a second person is outside their setup, and each of those scales already absorbs the empty case at a mark of zero: the cadence control reads “there was no schedule, so the question does not arise” at zero. The box would be a door the scoring does not open, since both gates treat a waived input as an absent one.

Three bands

80 and above is REPRODUCIBLE AS DESCRIBED. 50 to 79 is DRIFTING. Below 50 is UNREPRODUCIBLE, which is also where the first gate sends a result regardless of what it scored.

The floors are borrowed rather than designed. The AI Output Audit-Trail & Record-Keeping Kit that RedHub already ships puts TRACEABLE at 80, GAPS at 50 and UNRECORDED below that, on a weighted completeness score with a structural gate over it that can force the bottom band at any number -- which is the same shape as the three gates here. That is a narrower cut of this same subject, and it is one of the products this site recommends. Choosing different numbers by feel would have made a free assessment disagree with a paid one measuring part of the same thing.

DRIFTING is borrowed too, and it is the middle band of four shipped RedHub instruments where it means the same thing every time: still working, moving away from where it should be, and nothing stopping it. That is exactly the state this band describes, so it is the word rather than a fraction of the one above it.

Three gates, and why they only ever hurt

A gate can lower a verdict and can never raise one. That is the whole mechanic, and it is the reason a verdict here means something a weighted average does not.

Needs a witness. Two of the 24 controls are routes to the evidence that do not run through one person: a written procedure somebody has followed, and a second person who has produced the pack. If both read 1 or below the verdict is UNREPRODUCIBLE at any score, because every other answer on the page then describes what one person can do, not what the organization can. Bringing either one up to something that has actually happened releases it.

One or below, not zero, and for a specific reason. At a mark of one there are notes that assume you already know how, and a second person who has watched. Both are real and neither is a route: notes written for their own author are a reminder, and watching leaves out every assumption the person doing it never had to say aloud. At two, somebody other than the author has followed the procedure, or a second person has produced the pack. That is thin, and it is not nothing.

The idea is taken from a product RedHub ships. The Audit Evidence Pack Assay reads a link that holds only because a person asserts it as needing a witness, and overrules it to no record when that witness is not available. This gate goes one step further than its source and the difference is worth stating: the assay asks whether the witness is available, and this asks whether anything survives the day they are not. Over a quarter that is a question about turnover, not about who is at their desk.

Nothing current. A top verdict is held at DRIFTING when nothing in the freshness domain records a review that has happened. In that domain a mark of 2 is the first that means something occurred, not that it was scheduled, so the gate reads: no artifact carries a date anybody honored. Without it the headline is reachable anyway, which is why it is a gate rather than a matter of weights. Freshness carries 18 of the 100 points where nobody outside asks. Answer all three at 1 and everything else at 3 and the score is 88, comfortably above the top floor, on a posture whose entire problem is that nobody has looked at it since it was written.

A control reading zero. A top verdict cannot sit on an in-scope control with nothing behind it. Without this a weighted average quietly fills the hole in: twenty strong controls carry the number while the one you have not started disappears into the mean.

Three worked results

An organization nobody outside asks, answering every control at 3 except the written procedure and the second person, which read 1, scores 93 and returns UNREPRODUCIBLE. Those two carry 10 of its 100 points, so losing two thirds of both costs about seven. Answer both at zero instead and the same description scores 90, which is the most that arrangement can reach. Neither number is close to the verdict, and that is the point.

The same organization with both routes at 3 but every freshness control at 1 scores 88 and returns DRIFTING. Nothing is missing, nothing reads zero, and nothing has been re-checked. The top band would describe a snapshot, and this assessment is about a standing position.

A third, answering 1 nearly everywhere but with a procedure somebody has followed, a second person who has produced the pack and a held schedule, scores 39 and returns UNREPRODUCIBLE by score alone with no gate fired. That is a much worse position and an honest one: it has a great deal to fix and nothing hidden. Publishing all three matters because the ordering is the instrument. The score is context. The gate is the answer.

Every control and its weight

24 controls across eight domains, grouped by what they defend against: time, change and people. A weight of zero means the shape has no surface for that control and it is not scored. The two marked as routes are the ones the first gate reads, and the five marked as always asked cannot be waived.

Control	Internal	Customer-driven	Audited
Do the artifacts describing how you use AI carry a date by which somebody is meant to look at them again?Time · Dated and re-checked · always asked	6	5	5
When something was last reviewed, is that recorded on the artifact itself rather than in somebody's memory or an inbox?Time · Dated and re-checked · always asked	4	3	4
Over the last year, did the reviews that were scheduled actually happen?Time · Dated and re-checked · always asked	8	5	5
When something is found and fixed, is the fix recorded against the finding rather than simply done?Time · Whether closed stays closed	5	4	5
Does anything re-check a closed finding later, or does closed mean closed for good?Time · Whether closed stays closed	6	4	4
If the same problem came back three times in a year, would anybody be able to see that it was the same problem?Time · Whether closed stays closed	3	3	2
Is there a defined period for how long AI-related records are kept, and does it cover the period you would be asked about?Time · The period you would be asked about	4	3	5
Could you produce the records for a period that has already ended, rather than the one you are in?Time · The period you would be asked about	4	3	6
Does anything delete AI records on a schedule, and would a deletion that should not have happened be noticeable?Time · The period you would be asked about	3	2	4
When a model, a version or a vendor default changes underneath you, does that reach the record?Change · When the thing changes	6	5	5
When a vendor switches on an AI feature inside a tool you already had, does anything notice?Change · When the thing changes	5	4	3
When what a system is used for changes, does the paperwork change with it?Change · When the thing changes	4	5	4
Is the evidence a byproduct of doing the work, or is it assembled when somebody asks?Change · How the evidence comes to exist	7	8	5
Does the record of an AI-assisted output capture what somebody would ask about it: which tool, which version, who is accountable, and when?Change · How the evidence comes to exist	4	6	5
Can an output be traced to its input and to the decision behind it, without a person explaining the connection?Change · How the evidence comes to exist	4	6	5
Is there a written procedure for producing the evidence, detailed enough that somebody else could follow it?People · Whether it survives the person · route · always asked	6	5	4
Is there more than one person who could produce it?People · Whether it survives the person · route · always asked	4	4	3
If the person who normally does this left tomorrow, would the accounts, files and systems the evidence lives in still be reachable?People · Whether it survives the person	3	3	2
Could a record be changed after the fact without that being apparent?People · Whether the record can be trusted	3	3	5
Can you show where a record came from, rather than that it exists in a folder?People · Whether the record can be trusted	2	3	4
Has somebody outside your organization read any of this record and been able to follow it without you explaining?People · Whether the record can be trusted	n/a	3	4
Has the complete pack ever actually been produced: not planned, not partially, produced?People · Whether it has ever been produced	6	6	5
Do you know how long it takes, because you have timed it rather than estimated it?People · Whether it has ever been produced	3	4	2
When a request arrives with somebody else's deadline and somebody else's format, can you meet both?People · Whether it has ever been produced	n/a	3	4

The three shapes

Internal. No customer questionnaire, no auditor, no regulator and no insurer is asking you to evidence how you run AI. If the record decays, the only person who finds out is you, and only if you look. Pick this when you are doing this because it is right rather than because it is due. It is the most common shape and the least forgiving one, because there is no date in the diary doing the work for you.

Customer-driven. Security reviews, procurement questionnaires and vendor assessments arrive without warning, in somebody else's format, usually with a deadline measured in days and a deal attached to it. Pick this when the request arrives unpredictably and the cost of a slow answer is commercial rather than regulatory.

Audited. An auditor, a certification body, a regulator or an insurer looks at what you do, on a known date, and asks about a period that has already ended. They read the record without you in the room. Pick this when the question arrives on a calendar rather than out of the blue, and when the person asking will read the artifact rather than your summary of it.

What this assessment cannot do

It cannot verify a single answer. Nothing here reads a file, fetches a URL you supply or inspects a system, and that is a permanent design decision rather than a missing feature. Every number this returns is arithmetic on what you told it.

It cannot tell you whether your practice is any good. Durability and adequacy are separate questions, and this one answers only the first: an organization can keep beautifully current, well-attributed records of something that falls short of what its customers or a framework expect, and it will score well here. A good result is not evidence of compliance with anything.

It is not an audit, which is performed by somebody independent against evidence they inspect. It is not a certification and it is not a conformity assessment. It is not legal advice, and it names no framework anywhere, which means it also cannot tell you which rules apply to you.

It says nothing about likelihood. Everything here is about what happens to a record over time, not about the chance that anybody asks. Those are different questions and only one of them is a fact about your arrangements.

That is why the top verdict is called REPRODUCIBLE AS DESCRIBED rather than reproducible. It has to survive being read by somebody who did not answer any of the questions and cannot tell from the result that it was self-scored. The paid Audit Evidence Pack Assay asks the same kind of question one decision class at a time and reports ANSWERABLE per question; it is a self-report as well, so the difference between the two is resolution rather than inspection, and a per-question reading does not travel the way a headline about an organization does. What we sell against these gaps is listed in full, including the two domains nothing on the shelf closes.

Common questions

Why does the number of questions change when I change who asks?: Because two controls have no surface to exist on where nobody outside asks. There is no external deadline you could have met and no outside reader who could have failed to follow your record, so those two carry a weight of zero for that shape and drop out of the divisor entirely. It is not a free pass: the remaining controls are weighted heavier, so freshness, drift and capture carry more of the score than they do for anybody else.
Why are the basics weighted heavier when nobody outside asks?: Because there is nothing else holding them up. An organization with an audit date in the diary has a forcing function doing part of this work whether or not anybody planned it that way. An organization with no external deadline has only its own discipline, so the controls that measure discipline are carrying more of the answer, and the weights say so.
Where do the bands come from?: From a product RedHub already ships in this lane. The AI Output Audit-Trail and Record-Keeping Kit puts its top band at 80 and its floor at 50 on a weighted score of record-keeping durability, which is a narrower cut of this same subject. Choosing different thresholds by feel would have made a free assessment disagree with a paid one it recommends. The names differ: that product grades a log it reads, and this one reads what you told it, so its top verdict carries AS DESCRIBED.
Why can a gate only ever lower a verdict?: Because a weighted average is the wrong instrument for a structural failure. Twenty strong controls can carry a score while the one thing that would have kept the whole picture alive is missing, and averaging is exactly what hides that. A gate is where one specific fact overrides the summary, and letting it work in the other direction would make it a bonus rather than a limit.
Why can five of the controls not be marked not applicable?: Because both gates treat a waived input as an absent one, so the box would be a door the scoring does not open. Nobody running an assurance assessment can truthfully say that a review date, a held schedule, a written procedure or a second person is outside their setup, and each of those scales already absorbs the empty case at a mark of zero. Every other control on the site can be waived and drops out of the divisor rather than counting against you.
How is this different from the AI compliance gap assessment?: That one measures distance from a named framework and reports what is missing today. This one assumes nothing about what is missing and asks what will still be reproducible later. They can disagree in both directions, which is the test of whether they are two instruments rather than one twice: a strong framework posture can be undated, unwritten and known to one person, and a thin posture can be dated, rehearsed and reproducible by somebody else.
Can I check the score by hand?: Yes, and that is why every weight is on this page. Multiply each mark by its weight, divide by three, sum them, divide by the weight in scope and multiply by a hundred. The result is deterministic: the same answers always produce the same score, the same band and the same first thing to fix.

In one line

Score 80 or above with a route to the evidence that does not run through one person, something in the freshness domain that actually happened, and nothing reading zero, and the verdict is REPRODUCIBLE AS DESCRIBED. Lose both routes and it is UNREPRODUCIBLE whatever the score says.

Run the assessment What we sell against these gaps