Deciding new cases against retrieved precedents

Definition (Edward H. Levi, An Introduction to Legal Reasoning, 1949)

This procedure decides a new case by comparison with already decided ones. The judge states which features of the earlier case are relevant and which are not, and that selection is generally where the reasoning happens.

Analysis

The approach carries over the comparison part of Levi's procedure, and CARO trains this comparison rather than approximating it. There, the model generates an analogy chain from the new case to retrieved ones and is optimized on that chain. Jin et al. (2026) puts decided cases in the context next to the written rules, so the comparison runs in readable text. SIROCCO does hand-built structural matching against 500 NSPE decisions. What is missing is the step that Levi says the reasoning consists of, namely stating which features of the earlier case are relevant and which are not. CARO's divergence note records that this selection stays inside the weights. Jin et al.'s (2026) chains invoke cases without naming what was discarded. SIROCCO's relevance is a computed similarity rather than a stated ground. The first part is carried over, but not the part the reasoning rests on.

Definition (Arthur L. Goodhart, 1930)

This distinction separates the part of a past decision that binds later cases (the holding) from the remarks made along the way (the dicta), which do not. Without a way to tell the two apart, a body of precedent cannot indicate which of its own decisions are actually binding.

Analysis

Two of the three papers give the model whole cases as text with an outcome attached, so nothing marks which part of a past decision governs and which was said in passing. SIROCCO is the exception, and it is why the concept is partially preserved rather than omitted. Its case base indexes each NSPE decision by the code provisions the Board applied to the facts. This locates a decision's operative ground instead of treating the opinion as one undivided text. However, locating the applied provision does not meet Goodhart's test. Nothing separates a provision necessary to the result from one that was merely discussed, and no statement of the material facts is extracted. Since no output of any of these systems constrains a later one, the distinction has nothing to apply to.

Definition (Common-law practice)

These two techniques let a later court revise the record. Distinguishing confines an earlier precedent to its own facts, and overruling sets it aside altogether. Both are done openly and with reasons stated, which keeps a body of case law open to revision.

Analysis

In all three systems, the case base is a fixed input, and no decision that any of them reaches is written back into it. Jin et al. (2026) supplies it at training and inference, CARO bootstraps it once and then optimizes over it, and SIROCCO hand-encodes it from Board opinions. As SIROCCO's divergence note puts it, the system reads a body of precedent it can never join. Because the record never grows through use, there is no later precedent of the system's own to confine to its facts or to set aside. Neither paper offers a mechanism for stating reasons that would narrow or retire a case already in the base.

Definition (Albert R. Jonsen & Stephen Toulmin, The Abuse of Casuistry, 1988)

This way of reasoning bases judgment on paradigm cases (i.e., clear cases on which agreement is easy) and on the short maxims drawn from them. A hard case is then settled by asking how far it is from the paradigms, rather than by deduction from a general principle.

Analysis

All three systems reason the way a casuist does, starting from cases. Jin et al.'s (2026) main result is that demonstrating safety through decided cases works better than spelling the code out at length, which degrades helpfulness. CARO reasons outward from clear retrieved cases to the ambiguous one. SIROCCO gives abstract code provisions content only through the cases in which they were applied. What is absent is the structure that makes casuistry work. No case is designated a paradigm, and nothing orders the case base from central to marginal. A hard input therefore sits near whatever the retriever happened to return, rather than at a measured distance from an agreed clear case. The maxim is missing too. The trained systems draw nothing short and portable out of their cases, and SIROCCO's principles come ready-made from an external professional code instead of being drawn from the paradigms.

The taxonomy treats the Codified and Case-accumulated locus categories as alternatives. This approach supplies both at once and lets the model reason across them. Real legal systems work this way, but neither locus category describes it. In such systems, the codified text and the decided cases are inputs to one act of reasoning rather than competing sources of authority.

Levi describes analogy as something a judge does and can be held to. Making it a training objective optimizes the model toward the form of analogy. This produces the form reliably but provides none of the accountability. A chain of analogy that reads correctly is rewarded whether or not the feature selection behind it would survive scrutiny.