Why every number in a getax report carries a citation
A tax position is only as defensible as its weakest sentence. Here is the chain each figure in a getax report has to pass through before it can appear in one, and what it means that a claim is never a fact until a council rules.
Carol Ireland · 4 min read
Ask a partner what worries them about a cross-border position and they rarely say the analysis. They say the sentence they cannot trace. A rate that was right two years ago. A threshold copied from a summary that copied it from another summary. A treaty article quoted from memory. The position reads well, the client acts on it, and eighteen months later somebody has to find out where the number came from.
getax was built around that sentence. The rule inside the system is simple to state and expensive to keep: nothing appears in a report unless it can be walked back, hop by hop, to the official text it rests on. This post describes the walk.
Six hops from a government page to a sentence
Every fact in the knowledge base is the end of a chain, and every link in the chain is stored.
- Source. A tax authority, a consolidated statute, a treaty text. Not a commentary, not an accounting firm’s summary, not an encyclopaedia. Those can point us at a source; they are never the source.
- Document version. The page or file as fetched on a given day, with its hash. When the authority edits the page, that is a new version; the old one is kept, so a fact that relied on it still resolves.
- Verbatim evidence. The exact passage, in the original language, unaltered. It is never rewritten and never translated. If the Portuguese says one thing and an English gloss says another, the Portuguese is what the fact rests on.
- Normalised claim. A single English proposition extracted from that passage: this rate, this threshold, this condition. Extraction proposes. It does not decide.
- Council deliberation. A panel of independent model reviewers reads the claim against the evidence, one of them assigned to argue it is wrong, and an arbiter writes a rationale. Every opinion is kept as a record.
- Fact version. Only now does a claim become a fact, stamped with the models that judged it, the code that ran, and the time. A fact is never edited; a correction is a new version that supersedes the old one, and the old one stays.
A claim is not a fact until a council rules
The distinction sounds bureaucratic. It is the whole point. Language models are good at reading a page and producing a plausible sentence about it. Plausible is not the standard a firm signs its name to. So the system separates the two jobs: extraction may propose as many claims as it likes, and none of them can reach a report until reviewers who did not write the claim have checked it against the passage it cites.
The council cannot be talked around by consensus either. What a fact is allowed to be called depends on the evidence tier, not on how confident the reviewers feel. A rate or threshold can be marked validated only on tier-one evidence: the authority, the statute, the treaty. Evidence from a lower tier can at best leave a fact corroborated, however many reviewers agree, and the arbiter is bound by the same table.
Absence is information
A knowledge base that only records what exists misleads by omission. Several jurisdictions have no net wealth tax and no inheritance tax, and a client planning a move needs to know that as surely as they need a rate. So the system records documented absences as facts in their own right, anchored to the authority’s own catalogue of taxes, and marks them as absences rather than letting an empty category look like a gap nobody got to.
The same honesty runs the other way. Where a source is behind a wall we cannot reach, the coverage view shows the gap and the reason. A country is never quietly marked complete.
What the reader sees
None of this machinery is visible in a report, and it should not be. What the reader sees is a numbered citation after each material sentence, an evidence card that opens to the verbatim passage with its retrieval date, and a sources index at the end that lists every official document the position relies on. The reviewer at the firm can check any sentence in under a minute. The client’s file, years later, still shows where each number came from.
A position your firm can defend is one where the weakest sentence still has a source.
The sample report shows the result on a fictional engagement. The security page shows what the system does with a client’s data along the way. We publish both because the same rule applies to us: a claim about the product is not a fact until someone can check it.