Validator-green is necessary and not sufficient. A filing can pass every Arelle plugin in the standard pipeline and still carry a tagging defect that will quietly mislead every downstream consumer for the rest of its public life. Audit-grade review is the discipline of going beyond green status to confirm that the rendered text, the canonical value, the concept choice, the context, and the unit all agree on what the fact actually means. You are not auditing the validator; you are auditing the meaning.
Treat every fact as a unit of meaning with five layers. The rendered text is what a human reader sees. The canonical value is what a machine consumer parses. The concept choice is what the taxonomy says the fact is. The context is when, who, and under what dimensional combination the fact applies. The unit is the measurement system. Disagreement among any two of these is a tagging defect, regardless of validator output. The most common silent failure is canonical-versus-rendered drift: a "1,200" rendered with scale="3" and decimals="0" reads to a consumer as 1,200 but to a human as 1,200,000.
Sample, do not exhaustively check. A representative sample of three to five facts spread across at least two extended link roles, plus one block tag, plus two contexts (one instant, one duration), plus two units (one currency, one non-currency) catches the overwhelming majority of cross-cutting defects in a few minutes. Audit identity gets applied on every sampled fact: canonical equals transform of rendered text times ten to the power of scale times sign. If the identity fails, the fact is wrong.
When an AI agent produces a filing, ask three things in addition to the file itself. First, the validator output (which plugin, which version, which codes raised). Second, an audit-identity check on at least three sample facts the agent picks. Third, a one-paragraph self-summary describing what was generated, what was deliberately not generated, and which validator profiles were run. An agent that cannot summarise its own output cannot be trusted to have produced correct output.
Document every finding with code, file, line, and proposed fix. The finding becomes the prompt, the prompt becomes the fix, the fix lands in a diff, the diff lands in the filing. Every step is traceable. That trace is the audit trail you will need when the regulator asks how the filing was prepared.
It is worth closing the course where it started, on the picture of the graph. Everything you have learned to check is, in the end, a check on one knowledge graph: the five-layer fact is a node, the context is how that node connects to an entity and a period, the calc and presentation arcs are the edges, and the DTS is the schema the whole graph obeys. When you prompt an AI agent to author or repair a filing, you and the agent are reasoning over the same graph from opposite ends. The agent's strength is that it can traverse the entire structure at once, every concept, every arc, every label, far faster than you can scroll. Its characteristic weakness is that it will state an assertion the graph does not actually support, confidently and in fluent XML, the digital equivalent of a number that foots on the page but lies to the parser. Audit-grade review is exactly the discipline of checking the agent's assertions against the graph rather than against its tone. You sample nodes, you follow edges, you confirm the acyclic arithmetic settles, and you trust the filing only when the graph and the claims agree. A validator tells you the graph is well-formed; you are the one who tells whether it is true.
Audit-grade review treats every fact as a unit of meaning. The rendered text, the canonical value, the concept choice, the context, and the unit must all agree.
A defect that passes a validator is still a defect. Wrong concept, wrong sign, wrong scale, wrong namespace, wrong identifier scheme are all silent failures.
Sample contexts and units. Sample blocks. Sample facts across at least two ELRs. The fast read is the first line of defence; the sample is the second.
When an agent produces an iXBRL filing, ask for the validator output, the audit-identity check on at least three sample facts, and a one-paragraph self-summary of what was generated and what was not. Trust no agent that cannot summarise its own output.
Document findings with code, file, line, and proposed fix. The finding becomes the prompt; the prompt becomes the fix; the fix lands in the diff; the diff lands in the filing. Every step is traceable.
3 questions. Druk op 1 tot 4 om te kiezen, Enter om te bevestigen.
Door deze quiz af te ronden, markeer je deze sectie als voltooid.
Tip: kies met 1 tot en met 4 en bevestig met Enter.