How the AI works
This product is built on a large language model. That is worth being specific about, because “AI-powered” is used to mean anything at all, and because the honest version includes things a marketing page would usually leave out.
What actually happens to your documents
We read the text layer out of each PDF. Pages that are scanned images — very common in strata reports — are rendered as images and read visually. All of that goes to Anthropic’s Claude model in two passes:
- Pass one, per document. Extract findings, figures, dates and clauses, each one with the verbatim quote and page it came from.
- Pass two, across documents. Decide what matters most, place findings into sections, spot contradictions between documents, and write the questions worth asking. This pass works from references to what pass one found — it never rewrites a quote, so your evidence cannot be paraphrased on its way into the report.
The check that matters
Before a report is assembled, every quote is matched against the text extracted from your document. A finding whose quote does not appear verbatim is dropped rather than shown to you. This is the guard against the failure people rightly worry about — an AI inventing a plausible-sounding clause that is not in your contract.
It has one hole, and we mark it rather than hide it. A quote read from a scanned page image has no text layer to be checked against. Those citations carry a badge in your report: read from a scan — verify against the original. They are usually right. They are the ones to check first.
Fact and interpretation are kept apart
Every finding has two separate parts: what the document says and what it might mean. The first is anchored to a quote. The second is the model’s reasoning about consequences, and it is exactly the part a professional should weigh in on. We keep them visually separate so you always know which one you are reading.
Where it is weakest
Every report carries these limitations. They are not fine print:
- An inspection reflects one day and accessible areas only.
- Strata records can be incomplete; this decode covers only what was in the documents provided.
- This decode covers only the documents you uploaded — not any document you did not provide.
- No valuation of the property is performed.
Beyond those: the model can misread a table whose layout is unusual, miss something a specialist would notice, and be more confident than it should be. It is at its best on documents that state things plainly and at its worst on documents that are silent — an absence is much harder to spot than a statement, which is why “what these documents did not cover” is a section of every report rather than something we leave you to infer.
What we do not do
- We do not train on your documents. Anthropic does not use them to train its models, and we do not keep a copy — they are deleted when your report is ready.
- We do not have a human review your report before you see it. It is AI output, delivered as AI output.
- We do not decide anything for you. The report has a recommended action on it — that is a summary of what the documents suggest, not a recommendation to bid.
The disclaimer, in full
Explains the reports you already paid for and prepares your questions for your professionals. Not legal, building, or financial advice; not a substitute for a conveyancer or inspector.
More on accuracy and what happens to your files in the FAQ and the privacy page. Or read a complete sample report and judge it yourself — that is the version of this page we’d rather you trusted.
Last updated 20 August 2026.