Tracing Justinian
One legal question followed through Justinian's published architecture, layer by layer: where each of the ten pre-flight decisions is made, and what each one is for.
Anthropic's research on tracing the thoughts of a large language model follows one prompt through a model to see which steps produce the answer. This experiment borrows the idea at another level. We do not show Justinian's internals and make no claim about them. We follow one question through the architecture we already publish, and show where each of the ten pre-flight decisions is made.
Read Anthropic's researchAn illustration, not a recorded run. The layers and the ten decisions are the ones on our Macro Architecture diagram. The notes under "In this example" explain what each decision is for, applied to an invented question. They are not output captured from the product.
Follow the Question
Our client, a Dubai company, wants to end a distributor's exclusivity early. The distribution agreement is governed by UAE law and drafted in English. What notice do we owe, and what are the risks?
Invented for this page. The parties are fictional.
Step 1 of 9
User Query
Decision 04
Entities
Parties, dates, obligations, amounts and governing law, pulled out as structured data.
In this example
The parties (our client, the distributor), the instrument (the distribution agreement) and the governing law (UAE law) are pulled out of the sentence as structured data before anything else reads it.
What This Trace Does Not Show
- How the model reasons inside. Anthropic's method looks at a model's internal features. We have not measured Justinian that way, so nothing here describes its internals.
- Timing, cost or the models used. The architecture names what is decided, never how it is served.
- A legal answer. The trace stops where the answer would begin.