Six Months Later, Legalise Drafts, Reviews and Researches on an Upgraded Architecture

In October 2026, Generate, the Matter Reviewer and Research Lab moved onto an upgraded architecture with more thorough checks. Nothing changes in how you use them. Open Generate as usual and your next draft uses it; the Reviewer and Research Lab work the same way.

Two measures, in a fixed order

We hold the tools to two measures, and the order between them is deliberate.

The first is wrong citations: a document that cites a law that does not apply, or a section that says something other than what the document relies on it for. The second is precision: the appropriate laws recalled and applied where they are needed.

The architecture is built to reduce wrong citations first and to raise precision second. A missing citation is a gap the advocate sees and fills. A wrong one reads correctly and can be filed. Citing something wrong is far worse than not citing it at all.

Generate

Generate is now better at finding the clauses specific to your jurisdiction. Across our benchmark of 250 tests:

• On most basic drafts, precision rose from ~72% to ~97%.
• On some difficult drafts, it rose from ~66% to ~87%.
• On basic to intermediate drafts, wrong citations came to zero.

What remains lies mostly in edge cases within highly complicated matters, such as constitutional writ petitions, rather than in everyday instruments such as agreements and NDAs. Those matters are also the only place a wrong citation still appeared.

Generate keeps its template method. Filling a structured template has repeatedly produced higher quality than free-form drafting, so it stays. Free-form drafting is still being refined.

The Matter Reviewer

The Reviewer runs on the same architecture and is likewise better at finding the clauses specific to your jurisdiction. It is significantly better at finding errors and issues in legal documents. The report reads as it did before.

Research Lab

Research Lab carries the most thorough checks of the three, which is also why a report takes longer to arrive. Across the same testing it improved by ~36% on average, with a ~70% lower error rate on harder questions than before. Improvement here means precision in citing the right law for the question asked.

The benchmark

The figures above come from a benchmark of 250 tests. We aim to take it past 1,000 and to make it open to the profession.