What Legalise Will Not Guess

The characteristic failure of a language model in legal work is not that it produces nonsense. Nonsense is easy to catch. It is that it produces something fluent, correctly formatted, correctly cited in form, and untrue in substance, and it produces it in precisely the places where the reader is least likely to stop and check: a figure, a date, a section number, an annexure that is asserted to be attached.

Most of the engineering behind Legalise is not about making the model write better. It is about establishing, deterministically and before anything reaches you, which facts the model was entitled to state at all. This post is about what the platform declines to supply, and what it puts in the space instead.

The rule underneath all of it

Names, figures, dates, annexures and party facts are closed sets. The model may arrange what it has been given. It may not add a member to any of those sets.

This sounds obvious and is not how these systems behave by default. Left to itself, a model handed an incomplete instruction will complete it, because completing it is what it does. It will invent a court fee and attach a pinpoint citation to it. It will list annexures that were never mentioned. It will name legatees in a will who appear nowhere in the record, or supply birth dates for children in a maintenance suit that nobody typed. Each of those is a real defect that was found by reading finished documents, and none of them looks like an error on the page. They look like a well drafted document.

So the closed set is enforced rather than requested, and a document is checked against the record it was built from before it is handed over.

An empty space, not a plausible number

The most visible consequence is that the platform will show you nothing where it has nothing.

Court fee and stamp duty figures are read from each jurisdiction's own established schedule. Where that schedule has been established from primary sources, the figure appears. Where it has not, no figure appears and no estimate block appears either. There is no placeholder, no approximate range, and above all no Islamabad figure printed under a provincial cause title. A missing number costs you a lookup. A borrowed number is a false statement about that province's law that reads exactly like a correct one, and it stays undetected until somebody pays the wrong duty.

The same rule governs statutory citation. Where a provincial provision has been established, it is cited by section. Where it has not, the substance is given without a section number rather than with a plausible one. A citation that is nearly right is more dangerous than no citation, because it survives a skim.

Silence as a claim

In the statute library, a record whose repeal status has not been investigated says nothing at all about its currency. It does not say "in force".

That is a deliberate refusal to make the easiest and most commercially attractive claim available, which is that the library is up to date. Absence of evidence is not evidence of currency. Where a repeal has been established the badge quotes the repealing statute's own words and the page they appear on so you can check it; where the evidence suggests supersession without establishing it the note says so and avoids the word repealed; where nothing has been established the record is silent and the reader is left to do what they would have done anyway.

When the engine cannot reconcile your own input

Sometimes the contradiction is in the instructions rather than in the law. An application states that the investigation is complete in one field and ongoing in another. An execution petition contains figures that do not add up. A file references an annexure that is not among the material.

The platform does not silently pick a reading, and it does not refuse to draft. It produces the document and then tells you, in a panel above it, what it could not reconcile and what you supplied that never made it onto the page. Those notes are attached to the document in your workspace and they are not part of the document, so they are absent from every export, every download and every shared link by construction rather than by a filter. What reaches a court file is the filing. What reaches you is the filing plus the list of things worth a second look.

When there is nothing worth charging for

If a generation comes back with no substantive content, which happens occasionally when an upstream model fails in a particular way, the platform does not hand you a shell of blank lines and take an instrument for it. It retries once, and if the second attempt fails it says the service is temporarily unavailable, deducts nothing and saves nothing. A document that looks finished and is empty is worse than an honest failure, because the numbered clauses close ranks over the gap and nothing on the page shows it.

In review, a demotion rather than a deletion

The Matter Reviewer holds the names and numbers of provisions rather than their full text, which means there is a class of statement it is not in a position to make: that a given section requires something. It used to make those statements anyway, at the top of reports, with instructions to act on them.

It now has to be able to quote the basis for a statutory requirement it asserts. Where it cannot, the finding is moved out of the leading position rather than deleted, because the reviewer's judgement that something is worth your attention is often sound even when its account of why is not. You still see it. It just stops being the first thing you read and stops presenting itself as settled law.

What this is not

None of this makes a generated document safe to file unread, and nothing here is offered as a substitute for the advocate's own judgement. Every output is a draft assembled from your instructions and controlled legal material, and it is your name that goes at the bottom of it.

What it does mean is narrower and, we think, more useful: the places where the document is uncertain are the places where it looks uncertain. You are not asked to find the invented figure among four correct ones. That is the whole design.