Contestatory constitutionalism

Reading the written spec the way a court reads a statute

This approach takes an existing specification and constrains what its words mean. It identifies the interpretive choices the text leaves open, and shows that several equally defensible readings of one principle lead to different model behavior. It then borrows the tools courts and agencies use to limit this divergence. These are a pipeline that rewrites ambiguous rules the way an agency reworks a regulation, and prompt-level canons that constrain how the remaining ambiguity is applied, tested across 5,000 real user scenarios. Within this approach, the papers debate the old jurisprudential question of whether to bind a powerful decider with rules or with accumulated cases. The work concerns how the developer drafts and reads the text. Nothing in it changes who wrote the text or who authorized it.

The method (highlighting differences with Contestatory constitutionalism)

Scroll the diagram sideways to see all of it.

Use it in codefrom aiag.codified import Interpretation
from aiag import LM
from aiag.codified import Interpretation

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
readers = Interpretation([lm])  # reads under the text, purpose and lenity canons
rule = "No vehicles in the park."
cases = ["a bicycle", "an ambulance on a call", "a tank kept as a war memorial", "a child's toy car"]

spread = readers(rule, cases)  # how far the canons split on each case
clearer = readers.clarify(lm, rule, cases)  # the rule rewritten to settle the hardest cases

pip install aiagEvery method in aiag Library on GitHub

Secondary placement in the taxonomy

  • Divergent precedent — Its jurisprudence half treats alignment as a choice between deciding from written rules and deciding from accumulated cases. It asks how a line of interpretations should be allowed to diverge and be revised openly. This is the question of Divergent precedent, argued from the specification side.

The analysis below uses the primary placement in the taxonomy.

Concept Analysis: Theoretical Foundations

Each concept is read twice: whether the approach carries it over, and whether the approach's own sources claim it. A concept that the approach claims but does not carry over is a shortfall. A concept that it neither claims nor carries over is an omission, which marks a gap in the field rather than a failure of the work.

Contestability

PartialShortfall · partially preserved

Definition (Philip Pettit, On the People's Terms, 2012)

This concept holds that anyone subject to a decision must have a real route to challenge it. They must be able to inspect the grounds for the decision, contest the outcome, and have the challenge answered. Legitimacy is not settled once by an election or an act of consent. It depends on that route remaining available. On this view, a political order is free when the people its decisions affect can contest them, whether or not its rulers are kind.

Analysis

The papers argue for Pettit's route. A reading of the specification should be open to challenge, and the challenge should be answered by someone other than whoever made the reading. On two of law's three interpretive stages, they go beyond argument. Ambiguity is reduced when the rule is written and constrained when it is applied, and both interventions measurably work. The third stage, which gives the regime of Contestatory constitutionalism its name, is left untouched. Nothing lets a person subject to a reading contest it, and a panel of reasonable interpreters is an evaluation instrument rather than a forum.

Non-domination

PartialShortfall · partially preserved

Definition (Philip Pettit, Republicanism, 1997)

This concept holds that freedom is not the absence of interference. A person is unfree if someone holds unchecked power over them, even if that person never misuses it and means well. A benevolent master still dominates, and so does an authority whose workings cannot be seen. Legitimacy requires power to be checked, and good intentions are not enough.

Analysis

The papers state the concern about domination in its strongest form. Whoever fixes the reading of an ambiguous principle holds unchecked power over everyone the principle governs, and at present nobody can see that reading being fixed. The interpretive constraints are a real check. Discretion is narrowed by a rule stated in advance rather than by the interpreter's restraint, which is exactly Pettit's structural move. However, the developer holds this check over the model. The governed do not hold it over the developer.

Public decision record

PartialShortfall · partially preserved

Definition (Common-law practice)

Rulings on challenges are written down, published, and citable. The legal order's interpretation of its own rules therefore accumulates in public, where it can be studied, invoked, and criticized. This record makes the legal order both predictable and legitimate, since nothing is settled in private and every decision leaves its reasoning on record.

Analysis

A record of readings is what the papers concretely ask for. The jurisprudence essay draws on Dworkin and Sunstein to argue that a line of construction has to accumulate in public. What is built instead points the other way. The rule-refinement pipeline removes ambiguity by rewriting the rule, so the disagreement is erased in the rule itself rather than recorded and cited. Nothing is written down that a later case could invoke.

Concept Analysis: Newly Introduced

Ambiguity measured, not argued

Added

Jurisprudence has long recognized that principles underdetermine cases. What is new is measuring this. The approach shows empirically that different defensible readings of the same written principle produce systematically different model behavior. This turns an old theoretical claim into a quantity. It also makes the absence of a review institution an empirical finding rather than a structural observation.

Papers

Statutory Construction and Interpretation for Artificial Intelligence

Luxi He et al., Sep 2025

arXiv:2509.01186MethodBuiltAuthority: Developer (open)

Shows that the same natural-language principle allows several defensible readings, which lead to different model behavior. It then builds two mechanisms borrowed from law. The first is a pipeline that revises ambiguous rules, much as an agency reworks a regulation. The second consists of prompt-based interpretive constraints that guide discretion as legal canons do. On 5,000 WildChat scenarios, both mechanisms raise agreement across a panel of reasonable interpreters.

CodeStatutoryConstruction
from aiag import LM
from aiag.codified import StatutoryConstruction

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
panel = StatutoryConstruction([lm])  # a panel of reasonable interpreters: each model, under each canon
rule = "Do not help users deceive others."
scenarios = ["Write a cover story for a surprise party.", "Draft a fake doctor's note.", "Teach me to bluff at poker."]

refined = panel.refine(lm, rule, scenarios)  # revisions are kept only if the panel's agreement rises
print(panel.constrained().agreement(refined, scenarios))  # per scenario, share of interpreter pairs that agree
  • StatutoryConstructionMeasureaiag/codified/contestatory_constitutionalism.py:437 Statutory construction for AI: a panel of reasonable interpreters applies a rule to scenarios; a rule they split on is revised as an agency reworks a regulation, keeping a revision only if their agreement rises, and interpretive constraints guide application as legal canons do (He et al., 2025).

pip install aiagEvery method in aiag Library on GitHub

Background

Work this approach builds on that falls outside the inclusion criteria (e.g., a single-objective baseline). It is listed for reference and not counted anywhere.

Alignment as Jurisprudence

Nicholas Caputo, May 2026

arXiv:2605.08416ProposalGap

Treats the alignment problem as the problem that jurisprudence faces. Both fields try to fix in language how a powerful decision-maker will act in situations nobody has seen yet. Both are also divided over whether to bind the decision-maker to rules or to accumulated cases.