The grounding ladder
The answer to “how do you know?”, in five lines. It is the reason the rest of this architecture is shaped the way it is, and it is the difference between a system that produces evidence and one that produces plausible sentences.
Risk := a downward path to a Vulnerability AND an upward path toward a top risk Vulnerability := a Fact (grounded below) AND an upward path to a Risk Fact := a downward path to Evidence Evidence := a downward path to a Measure Measure := an observation of the node it measures, grounded on a Twin
Why this is the anti-fabrication argument
A language model asked “are we compliant with Article 26 of the AI Act?” will produce a fluent, structured, confident answer. It will do this whether or not it has any information about your systems, because producing that answer is the task it was given and fluency is what it is good at.
Change the task and the failure mode changes with it. A model asked to attach a finding to a provision hash, and to a fact, and to a piece of evidence, and to a measure taken against a twin of a real system, either finds that path or reports that it cannot. There is no fluent version of a missing edge.
The ladder is the mechanism that turns “sounds right” into “here is the path.”
This is not a claim about model quality and it does not get better with a larger model. It is a claim about what the question makes possible to fake.
The five rules that follow
1. Unevidenced facts are first-class findings, not omissions
The best worked example on this site ships with five of its nine questions unanswered, and the document that produced it calls that “the actual output of the exercise”. A question nobody can answer about a system that is already running is not a gap in the analysis; it is the analysis. The worked example →
2. Absence is output
So it has to be visible. Both of the project's shipped graph tools ghost an unanswered node rather than defaulting it — amber for exposure, green for assurance, ghosted for unanswered — and this site follows the convention:
| Question | State |
|---|---|
| Are inference logs retained? | Yes — 30 days, from the deployment record |
| Has the human-oversight path ever been exercised? | Unanswered |
| Who is the responsible natural person under Article 26(2)? | Unanswered |
Ghosted rows are unanswered. They are rendered rather than omitted, because a table that silently drops what it does not know reads as complete.
3. Computed, not claimed
A number that cannot be recomputed from the graph does not belong in it. This is the rule that keeps a percentage from appearing on a page and then acquiring authority it never had — and it applies to this site's own statistics as much as to a customer's.
4. Regulation as evidence, not checklist
The instrument is what a claim points at, not a list to tick. A checklist compresses a provision into a box, and the compression is where the meaning goes. The example that makes this concrete is AI Act Article 26(5), which a checklist renders as one box and which is in fact a dual obligation: “Suspension without notification is not compliance, and notification without suspension is not either.” One box, two obligations, and the box can be ticked while half of it is unmet.
5. Coverage is a measurable property of the graph itself
How much of an estate is actually grounded is computable — a proportion of nodes with a complete downward path — rather than asserted. It is also the one number this method permits, precisely because it measures the analysis rather than the subject.
What the ladder is not
Three concrete consequences, enforced rather than intended: no tool here outputs a pass, a score, or a percentage of compliance — findings and unanswered questions only. No page says “compliant”, “meets” or “satisfies” without naming the evidence and the measure. And the machine surface carries the same rule for agents: an agent may report which provision a claim points at; it may not report that a requirement is met.
This is not timidity. Everybody else sells verdicts, and verdicts are the thing nobody can defend. It is also the standing sales objection — a buyer of “Standards As A Service” will ask whether they pass. Named as a tension rather than discovered later →
Where this ends and risks.sgit.ai begins
The ladder is shared apparatus, and the two sites own different halves of it. This site owns the provision and the arithmetic — which obligation a finding points at, and whether the finding follows from it. risks.sgit.ai owns the register and the acceptance decision — whether a risk is accepted, by whom, and against what appetite. One canonical copy of any shared example, the other links in. The deconfliction is Q1 and it is open →