Measuring a jurisdiction’s value hierarchy from statutes and court decisions
How we read statutes and court decisions one by one to find which value a jurisdiction’s law and courts put first when two values collide.
The values law and courts put first
Statutes and court decisions record which value comes first when two collide. Reading many of them shows the order a jurisdiction’s law places between values.
Within one jurisdiction the order can change with the domain and the situation, so we build it per domain and per situation.
- Value hierarchy
- The order that says which value comes first when two collide. For example, when personal freedom and public safety collide, which one does the law put first?
- Document
- One statutory provision or one court decision. Statutes and decisions count the same: one document each.
- Domain
- One of 21 areas, such as education, health, labour or defence.
- Situation cell
- One situation: five levels of scale, three of reversibility and three of urgency make 45 cells per domain, 945 in total.
Seven steps
- 1
Set the scope
We first list the courts that issue decisions and how the law is structured (constitution, acts and so on).
We count how many decisions are public and decide which ones to read. No one picks decisions by their content.
- 2
Collect
We download the current text of each provision and the full text of each decision from official government sources.
From decisions we use only the court’s holding (the majority opinion), not dissents. Texts stay in their original language.
- 3
Read
Three AI models from different companies read each document with the same instructions.
Each model records whether two values collided, which came first, the domain, whether the outcome holds only in certain situations, and one supporting sentence.
Case numbers are removed so a model cannot guess the outcome from the case name.
- 4
Count the votes
One model and one person each get one vote. For each document, the answer with the most votes becomes its answer.
If two experts who checked the original text give the same answer, theirs takes precedence.
- 5
Build the order
For each domain we count documents that hold in every situation: “A first” against “B first”. Only a two-thirds majority becomes an order.
Documents whose outcome requires a specific situation change the order only in that situation cell.
- 6
Check
We check that each sentence a model quoted is really in the source. People re-read samples of “no collision” decisions and of answers that rest on inference.
- 7
Show it on the map
On the Model Response Observation Map, the jurisdiction’s order is set beside an AI model’s order. Colour shows the share of value pairs with legal evidence on which the two hierarchies differ; hypothesis pairs and weak-evidence pairs are not counted.
There are three views. “Model baseline” shows how far the model’s order shifts across domains and situations, “Jurisdiction baseline” how far the jurisdiction’s order shifts, and “Jurisdiction vs model” how far the two orders differ.
Legal professionals take part
Legal professionals of each jurisdiction (lawyers, judges, prosecutors and law professors) check and complete the models’ values against the original text. If two experts who checked the text give the same value, it takes precedence over the models’ majority; a single expert’s value is one vote like any other answer.
Four layers, one on top of another
Layers are stacked from the bottom up. Where an upper layer’s evidence disagrees with a lower one, the upper layer wins. When the jurisdiction-wide or domain base order runs against the AIO hypothesis, it overturns the hypothesis only with four or more documents; with fewer it is marked as weak evidence and not counted in comparisons.
- ① AIO hypothesis v1
From the orders each model of a six-model panel from different regions set on its own, only pairs on which at least four valid votes point the same way are used (unanimous (6/6) · majority (5/6) · majority (4/6)). Every other pair has no direction. It is not legal evidence and is left out of comparisons and colouring.
- ② Jurisdiction-wide order
The order built from the documents of every domain.
- ③ Domain base order
Each domain’s order from documents that hold in every situation.
- ④ Situation exceptions
Changes made only in a situation cell by documents that require that situation.
Situations a decision mentions only while explaining its reasoning are shown for reference and do not change the order; only situations the outcome depends on do. The strength of a provision and its rank (constitution, act) are shown but carry no extra weight.
Applied to Korea first
- Decisions: every Supreme Court decision published for 2020–2025, plus every Constitutional Court decision from the same period that has a published summary (5,846 in total).
- Statutes: the current text of 522 provisions identified by domain in the September 2026 work.
- Reading: three AI models read every document.
We are preparing to collect documents so the same method can be used for China, Germany, the EU, India, Japan and the United States. Until then these jurisdictions do not appear in the map’s jurisdiction selector. How decisions are selected and which provisions are included will be added here once settled.
What this order does not say
- It includes values read by AI models. It is not a court’s holding or a correct answer.
- It does not rank jurisdictions against each other. It looks at the order each jurisdiction’s law places between values.
- Only published decisions can be read. Publication practice differs between jurisdictions (all published, court-selected and so on), and that difference can show up in the order.
- “No evidence of change found” is not the same as “no change”. The collected documents may simply not cover that situation.
- Full decision texts and model inputs are not published. Supporting sentences quoted from decisions are published only after human checks.