Field Notes September 29, 2026 8 min read

Catena: The Expert We Built First, and Why It Was the Right One

Before law, we built an authority expert over the Catholic Church's documentary corpus: 154,000 passages, formally graded, tracked through supersession, with a citator. It has law's exact reasoning shape and none of law's licensing walls, and it is where the engine under Irenaeus was proved.

A

Anton Mannering

Founder & Chief Architect

People assume a company building governed memory for law would start with law. We started with the Catholic Church's documentary corpus, and the reason is not what it looks like. Law's reasoning has a particular shape: sources carry formal weight, old rulings stand until replaced rather than fading, and an authority's standing changes according to what later authorities did with it. Very few bodies of text have all three. This one does, in full, and it is public, finite and free. So the engine that now sits under Irenaeus was built and proved on it first. That engine is Catena.

Three properties, one corpus

Graded authority. A conciliar definition, a papal encyclical, a curial document, a catechism and a theologian's commentary do not carry the same weight, and scholarship has already said which carries what. Catena retrieves the grade a source carries and never asks the model to guess it. Presenting a low-authority source as a high one is the error a lawyer fears most in a research tool, and it is the same error here.

Supersession, not decay. A text from the fourth century can be fully binding and a text from last year can be withdrawn. A memory system built for conversations lets old things fade; an authority expert must switch that off and track what replaced what. Catena's reference mode does exactly this, and it is the mode Irenaeus's Library runs in.

Treatment. One document's recorded treatment of another, across centuries of it: confirmed, developed, corrected, set aside. Catena holds treatment as typed links between sources and reports standing with the chain visible. In law that apparatus is called a citator. It is the method of Irenaeus of Lyon, some seventeen centuries early, which is where our legal deployment gets its name.

What Catena is today

About 154,000 passages, in a corpus of its own on the platform, with the public documents in a public tier and the licensed texts behind a gate, which is why Catena is by invitation. On top of the corpus, a private layer: real accounts, threads with recall across them, and memory stored for each reader that the corpus never sees. It is the two-store shape a legal research seat has, a shared authority and your own notes, with the wall on the private layer and not on the corpus.

Catena is also a true consumer of the platform. It imports nothing from Yohanun's internals; it reaches the platform over the same API, with an API key, that any outside customer would use. We audited that claim in July, found that the surface actually serving it was calling the engine in-process, and rebuilt it until it was true. The point of Catena was always to show what someone else could build, not what its author could.

How we know it is right

An expert over authoritative sources can be fluent and invisibly wrong, so Catena has a gate: 21 adversarial probes, each written to catch one failure a careful reader fears, such as overstated authority, an invented citation, opinion presented as rule, a disputed question presented as settled, or a pastoral question answered instead of deferred. Each probe has a documented criterion, a separate model judges it, and the run goes through the public edge three times. The latest run, dated 8 September, passes all 21 with no unstable probe. The full history of that harness, including the bad days, is in an earlier post.

The gate turned out to track the model more than the price. One model read 21 of 21 twice and cost less than the larger one that ranged from 19 to 21, and a newer model read 18. We kept the one that passed. A property that depends on which model is on top is a property the platform has to keep measuring.

What it proved for law

Everything under Irenaeus's Library was proved here first: grade-carrying answers, retrieval that respects standing, refusal to invent, disputed questions left disputed. What a legal deployment adds is not a new engine but its own corpus, a set's own opinions, skeletons and notes, and its own harness, with probes for an overruled citation, an invented holding and obiter presented as ratio, which it must pass before anything is relied on. The apparatus is the same. The engine does not know or care which corpus it is over.

Where it goes

Catena stays what it is: the proving ground, the live demonstration of a grounded expert, and the place we return to when a change to the engine needs a corpus that will show us the truth. If you would like to read it, ask for an invitation. If you have a corpus of your own with this shape, formally weighted, standing rather than fading, with a record of how its texts treat one another, then you have the material for an expert of this kind, and the platform's authority-expert API is generic over the corpus. Irenaeus is the second such expert. It will not be the last.

Tags

Catena Authority Experts Irenaeus Consumers

Share this article

Related Articles

Talking beats subscribing

Building with AI memory or governance? Tell us what you're working on.

Email Us →