Two offerings in four months.
One model that checks
its own work.
Anthropic in May, OpenAI in September. What these two announcements have in common lies in their architecture, and the consequence shows in use.
On 17 September, OpenAI announced Astra for Law, a configuration of its most recent model for legal work, backed by an index covering United States law across more than 230 million addresses. Four months earlier, Anthropic had released Claude for Legal, a suite of twelve practice-area extensions and more than twenty connectors to the software the market runs on.
Both offerings are serious and solve a real problem: grounding a model on a verified corpus changes the nature of the exercise. The reference cited exists, it says what it is made to say, the practitioner can check it.
The two offerings do not stand in the same place on this point. The Astra for Law index covers American law and nothing else: outside that jurisdiction, grounding on verified sources does not apply to the law the practitioner works in. Claude for Legal proceeds differently: access to sources runs through open connectors, several of which cover other legal systems.
What they have in common lies elsewhere, and it is structural. Claude for Legal runs on Anthropic's models, Astra for Law on OpenAI's. When one of these offerings checks an analysis or flags a weakness in a line of reasoning, it is the same family of models that drafts and that controls. The control then shares the biases of the drafting, its blind spots and its characteristic errors.
No maker has any reason to do otherwise: it would be building against its own product.
Reading this month
Claude for Legal and Astra for Law
What these two announcements reveal about their architecture, and what they do not do.
“Our data does not leave the firm”
Why the objection is sound, and why a ban does not answer it.
How many subscriptions per matter?
The count is instructive, and the real cost is not the subscriptions.
The full corpus : max-legal.ai/blog
Model, layer, application
The model produces the text. The layer decides what it is asked, on which sources, and what is done with its answer. The application is what you see. The same model gives very different results depending on the layer that governs it, and it is that layer which determines what belongs to you.
That is exactly what the two announcements put at stake: which of these three levels are you buying.
The full definition →Worth noting
Built into the system, or added afterwards
On 25 September, Ireland's data protection authority published a report on five years of artificial intelligence supervision: some 180 products and services examined at Apple, Google, Meta, Microsoft, OpenAI and others. One finding recurs: data protection has to be treated as part of the system being built, not as paperwork attached to it afterwards.
The distinction matters for anyone adopting a tool, not only for anyone releasing one. A control described in an annex is read once; a control written into the architecture runs on every request. At MAX, the verification against the official sources of law and the review by a second model that did not draft are not commitments: they are stages of the output.
Something missing? MAX can now ask for more information while a request is being processed.
When information essential to building a deliverable does not appear in what you sent, MAX comes back to you to get it. You reply in the same email thread, and it carries on with what you have provided.
It may be a whole document: you have a company's articles of association drawn up and you mention a shareholders' agreement without attaching it. MAX calls for it, then resumes on that basis.
Not yet subscribed and would like to
Receive the next one
Once a month, at the start of the month.
Four minutes to read.