Taxonomy / Precedential adjudication
The norms the AI ought to follow exist in a body of past decided cases rather than in a written rulebook; each new case is settled by analogy to the earlier ones, and its outcome is added to the record as one more holding.
Scroll the diagram sideways to see all of it.
Limitation
The earliest cases generally carry outsized weight, since every later case is decided by reference to them. An early mistake may therefore harden into settled law instead of being corrected. What's more, retrieving a textually similar case is not the same as analogy in the legal sense, where what matters is whether the reasons behind the earlier holding apply to the new facts.
Here the work is constituting the record itself: rather than writing rules, a group renders verdicts on particular situations, and those accumulated verdicts become the standard the system is answerable to. The two versions differ in who decides and how coverage is reached — one runs a deliberative process in which a panel argues cases through to binding precedents (with a variant where a model is prompted to play that role), while the other builds a repository by seeding cases, having experts name the dimensions along which they can vary, generating variations along those dimensions, and putting each to public judgment. Neither decides anything new by itself; both leave behind a case book with an address, which the machinery in the next two strategies consults.
When a new situation arrives, the system pulls the already-decided cases most like it and reaches an answer by working out which features carry across and which do not. The oldest version hand-builds the retrieval and matching procedure over 500 professional-engineering ethics decisions and returns both the past cases and the principles they turned on; the two recent ones train the comparison skill instead — one by reinforcement learning on its own case-augmented reasoning chains, with the written safety rules held in view beside the precedents, the other by bootstrapping analogy chains through retrieval and then optimizing them, with the cases alone, so the model stops leaning on surface resemblance. Those two differ in where authority sits, between statute and precedent or in the cases only, and the literature has one paper on each side. What unites all three is the decision-time move: the standing record is consulted case by case and never compiled away.
This strategy distills the record once and then stops consulting it: past judgments are generalized into a standing artifact that decides new situations without retrieving anything. The artifact takes two forms, which differ in whether a person can read and argue with it — a neural judge trained on a large bank of crowd verdicts about everyday situations, where the judgments end up in opaque weights, or a written constitution recovered from the record and thereafter applied like any authored one. Within the constitution branch the record being read differs too: raters' preferences between two answers in one line of work, a model's own red-teamed failures logged as verdicts in the other.
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.
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.
These two moves 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 is what keeps a body of case law open to revision.
This method anchors judgment in paradigm cases, i.e., clear cases on which agreement is easy, and in the short maxims drawn from them. A hard case is then settled by asking how far it sits from the paradigms, rather than by deduction from a general principle.
This doctrine holds that a court is bound by the decisions of earlier courts on materially similar facts, whether or not it agrees with them. The binding force comes from the earlier decision itself, not from any rule stated in advance.
This approach decides each case on the narrowest ground available and leaves the wider question open. The record then grows by small increments, and later cases stay free to settle what earlier ones deliberately did not.
This claim holds that rights are established by the accumulated decisions of ordinary courts rather than declared in a founding document. The general rule is a summary of what has been decided, not the source the decisions follow from.
This is the casuist's account of which circumstances of a case carry moral weight, e.g., who was involved, what was at stake, and what was known at the time. Cases are sorted by these features, and the sorting is what makes comparison between them possible.
This doctrine permits an agent to act on a defensible opinion even when a more certain one is available, provided the opinion is genuinely supported. It treats moral authority as admitting degrees rather than as either settled or absent.
An Introduction to Legal Reasoning
1949
Legal Reasoning and Political Conflict
1996
One Case at a Time (judicial minimalism)
1999
Introduction to the Study of the Law of the Constitution
1885
Determining the Ratio Decidendi of a Case
1930
The Abuse of Casuistry
1988
Probabilism (Jesuit casuist tradition)
16th-17th c.