The four checks, in order of cheapness
- Is the quoted sentence in the document? Search the file for it. This costs seconds and disposes of the failure people most fear. A finding whose quote is not in the source is wrong, and nothing after it needs examining.
- Does the rule say what the finding says it says? Open the cited statute or regulation and read the section. Summaries drift; sections do not. If the rule is narrower than the claim, treat the finding as a lead rather than a conclusion.
- Does the date appear in the clause it is attached to? Extracted dates should always be readable next to the obligation they belong to in the original. A deadline that cannot be located in its own clause should not be diarised, let alone relied on.
- Is this the kind of claim that can be checked at all? Severity rankings and money estimates are interpretations of the document. They are arguable by design, and treating them as facts is the mistake that a well-formatted report invites.
Which parts are fact and which are opinion
| Part of the output | Verifiable? |
|---|---|
| The quoted clause | Yes — it is in the file or it is not. |
| The rule cited | Yes — read the section. |
| The extracted date | Yes — against the clause it came from. |
| The severity | Partly. It is a ranking, and rankings are arguable. |
| The money figure | Partly. Check the inputs first, then the arithmetic. |
| The recommended action | No. That is advice-shaped, and it is where your judgment belongs. |
Where a document’s type or a figure cannot be established from its text, no confidence number is displayed at all. An empty space is honest; a plausible percentage produced from nothing is the single most expensive thing an analysis can hand a professional, because it survives review.
What a tool should refuse to tell you
Read what a product declines to claim — it is faster than reading what it promises. Main AI publishes no testimonials, no customer count, no logo wall and no usage statistic, for the straightforward reason that there are none it could source. It claims no compliance certification it has not earned. It does not guarantee a legal outcome, and it is not a law firm.
Those refusals are enforced by a check that runs on every deploy and refuses to publish a page containing an attributed quote, a star rating, a logo-wall component or a population-scale count. That is a weaker guarantee than an audit and a much stronger one than a promise, and stating which of those it is happens to be the same discipline this whole page is about.
The failure mode this method is aimed at
It is not hallucination in the dramatic sense. It is a finding that is broadly right, formatted confidently, and wrong in the one detail that matters — a date that belongs to a different clause, a statute that governs commercial rather than residential tenancies, a severity that reads as urgent because the phrasing is strong. None of those look wrong. All of them fail the four checks above in under a minute.
Start with a document where you already know the answer
Run something you have read closely and see whether the flags match what you found. Calibration is worth more than any claim about accuracy, including ours.
Read a document →