risks.sgit.ai / examples / Article 26(5)

Article 26(5): one provision, one agent, one graph

The complete instance. A single regulatory provision, a single deployed agent, and the whole ladder run end to end — from a fact about log retention up to a board-level exposure, and back down to the five questions nobody could answer. It carries its own node and edge inventory, and its most defensible finding is a subtraction.

The inventory

TypeCountNotes
Reality1The running system
Twin1The deployed agent
Fact8One deliberately unevidencedF7: the suspension procedure has never been exercised
Evidence7One absent, which is why there are eight facts and seven evidence nodes
Provision5Annex III 5(b), Articles 26(5), 26(6), 14, 27
Vulnerability3Each derived from a fact and a provision — the bridge point between the two graphs
Risk5Four in a chain, one meta — R5: no acceptable level has been defined
Stakeholder4ML platform lead, head of lending operations, DPO, CFO
Decision3Plus one deliberately absentD3
Question9Five unanswered“those five are the actual output of the exercise”
Project2Plus one unfunded

The finding that carries the whole example

Thirty days of log retention observed. At least six months required by Article 26(6). — “arithmetic, not judgement, which makes it the most defensible finding in the graph.”

Everything else in a risk assessment can be argued with. A severity is an opinion, a likelihood is an estimate, an impact is a model. This one is a subtraction, and it survives every conversation an executive can have about it: there is no rating to negotiate down and no methodology to dispute. It is also, deliberately, an external anchor — the six months is not the organisation's number, so no business unit can decide it is inconvenient. External anchors defeat systematic downgrading →

Three deliberate absences

The most instructive thing about this graph is what is missing from it on purpose. Each absence is a modelling decision rather than an oversight, and each demonstrates a different concept:

What is absentWhat it demonstrates
F7's evidence — the suspension procedure has never been exercised, and there is no record either wayNot knowing is a fact. F7 is recorded as unevidenced, which makes it countable and assignable rather than a blank
D3 — a decision that should exist and does notUnaccepted equals critical. The risk it should have covered surfaces on the CFO's register as an unowned critical item, with nobody having escalated it
R5's threshold — no acceptable level has been defined anywhere in the organisationArticle 9(5) mandates the judgement and never defines the word. The absence is itself a rateable meta-risk
“R3 appears on the chief financial officer's register as an unowned critical item, and it got there without anybody escalating it deliberately. That is the mechanism working: not doing something is a measurable action.”

The decisions, and their intervals

DecisionIntervalReading
D11 monthThe default rung — assemble and fund
D21 monthSame
D3noneDeliberately absent — this is the one that rolls up
D43 monthsThe upper end of “funded project”, approaching “waiting to see”

Five unanswered questions as the output

The claim that inverts what a risk assessment is normally judged by. The graph carries nine Question nodes; four have an answers edge and five do not — and the source says of those five that they are the actual output of the exercise.

An assessment that produced no unanswered questions almost certainly did not look hard enough.

It follows the quality gate on the same page: a question is not a risk. A sentence that cannot carry a named acceptor and an interval belongs in a different node type — and once it does, the unanswered ones become the work list rather than the residue.

Four new edge types proposed here

The example proposes edges the earlier ontologies did not have, which is a good sign about the method — a worked instance pushing back on the schema rather than fitting into it:

EdgeWhat it connects
governed_byA fact or system to the provision that governs it — the bridge between the risk graph and the regulation graph
in_scope_whenA provision to the condition that brings it into scope, so applicability is a path rather than an assertion
answersEvidence or a decision to the Question it resolves — and its absence is what makes a question unanswered
re_ratesA later decision to the rating it supersedes, which is what makes the level ledger queryable

Provenance

Source
briefs/08/02/vault-as-substrate/v0.33.55__arch-brief__…end-to-end-worked-example-article-26-5-creditworthiness-agent-fact-to-board.md
Repository
SGraph-AI__App__Send @ v0.33.55
First written
2 August 2026
Scenario status
Preserved from the source: the organisation is invented, and every invented element is marked. Stakeholder titles (ML platform lead, head of lending operations, DPO, CFO) are generic roles, not people
Provisions
Cited from the Regulation Graph vault — the EU AI Act parsed from official Formex XML and SHA-256 hash-verified. Legal points are factual and are not legal advice
Licence
CC BY 4.0 at source and here

For an agent

Article 26(5) — the complete worked instance. Inventory: Reality 1 · Twin 1 · Fact 8 (one deliberately unevidenced — F7, the suspension procedure has never been exercised) · Evidence 7 · Provision 5 (Annex III 5(b), Articles 26(5), 26(6), 14, 27) · Vulnerability 3 (each derived from a fact and a provision) · Risk 5 (four in a chain, one meta — R5, no acceptable level defined) · Stakeholder 4 · Decision 3, plus one deliberately absent (D3) · Question 9, five unanswered · Project 2 plus one unfunded. Intervals: D1 = 1 month, D2 = 1 month, D3 = none, D4 = 3 months. The load-bearing finding: 30 days of log retention observed against at least 6 months required by Article 26(6) — “arithmetic, not judgement, which makes it the most defensible finding in the graph.” It is an external anchor, so no business unit can downgrade it. Three deliberate absences each demonstrate a concept: F7's missing evidence (not knowing is a fact), D3's missing decision (the risk surfaces on the CFO's register as an unowned critical item with nobody escalating it), R5's missing threshold (Article 9(5) mandates the judgement without defining the word). The five unanswered questions are “the actual output of the exercise.” Four new edge types proposed: governed_by, in_scope_when, answers, re_rates. The organisation is invented and every invented element is marked.