The Twelve Bridges

A legal argument can cite only real cases, characterize each accurately, and still reach a conclusion that no court has held and that nothing in the cited authorities requires. The defect is not located in the citations. It is located in the word that joins them: therefore.

The Institute's working drafts take as their test case a synthesis of this kind. Brown v. Board of Education, Gideon v. Wainwright, and Roe v. Wade all establish that government power over individuals is presumptively limited; therefore, regulations affecting equality, access to justice, or personal autonomy are presumptively invalid. Each case exists, and each characterization is broadly defensible. The conclusion nevertheless rests on nothing. The three decisions arise under different provisions (the Equal Protection Clause, the Sixth Amendment, and the Due Process Clause), and no court has joined them into the principle the synthesis announces.

The drafts call this an inferential hallucination, to distinguish it from the factual kind. A fabricated citation can be caught by looking it up. An inferential hallucination cannot, because every lookup succeeds. The validity of a synthesis is a property of the relation among authorities, not of any authority taken alone, and no citation database records that relation.

The question, then, is not whether a conclusion is attractive. It is whether the conclusion follows from the authorities by a route the law recognizes.

The claim the framework makes

The framework's central proposition is that legal reasoning is constrained, not creative. A therefore must rest on a recognized doctrinal mechanism; where it does not, the conclusion may be argued but may not be stated as law. The framework calls each mechanism a bridge and identifies twelve. It also identifies eight faults, recurring patterns of invalid synthesis that imitate one bridge or another. The bridges operate as a whitelist, the faults as a blacklist, and both are to be applied at once.

The metaphor earns its place, though only to a point. A bridge is a structure whose adequacy can be tested independently of the traveler's wish to cross it. Each bridge in the framework is likewise defined by conditions that are or are not satisfied, whatever the advocate hopes to reach on the far side.

Two distinctions on which the framework depends

Nearly every condition in the framework is expressed in terms of two distinctions.

The first is between holding and dicta. Only the holding, the rule necessary to the decision, binds a later court. Statements made along the way, however quotable, do not. The most basic bridge requires that cited language be holding, and one of the faults consists precisely in treating dicta as though it were.

The second is between rule and rationale. A rule states what the law requires on given facts; a rationale states why the court adopted it. They travel differently. A rule reaches new facts that satisfy its elements, and nothing more is required. A rationale can reach facts the rule does not cover, but only if the reasoning in fact motivated the decision, applies to the new facts for the same reason, and, where it is a policy, was stated by the court itself. Much of the framework elaborates when a rationale may be carried beyond its rule.

A third distinction runs beneath both: between what a court has done and what an advocate proposes that a court do. Several bridges turn on whether a court, rather than the person making the argument, performed the synthesis. That condition separates a statement of law from a proposal for it.

The twelve bridges

The framework groups the bridges into four categories by the kind of connection each supplies. The numbers record the order in which the bridges were identified, not their categories. The names are the framework's own.

Authority-based bridges

In each authority-based bridge, the connection has already been made by an institution with power to make it. The advocate's task is identification, not construction.

Direct Holding Application (Bridge 1) requires that the cited language be holding rather than dicta, that the new facts satisfy each of its elements, and that it not have been overruled. The drafts' example is Miranda v. Arizona: where a suspect was interrogated in custody without the required warnings, the holding applies of its own force.

Established Multi-Case Doctrine (Bridge 2) applies where a court, and not the advocate or the machine, has synthesized several decisions into one test. The drafts illustrate it with the undue burden standard that Planned Parenthood v. Casey formulated from Roe and its progeny. The example now illustrates structure rather than current law, since the Institute's burden-shifting materials record that Dobbs v. Jackson Women's Health Organization overruled both Roe and Casey. Its point survives: one cites the synthesizing decision for the synthesis, and does not assemble a synthesis and attribute it to the cases.

Controlling Superior Court Adoption (Bridge 5) applies where a higher court explicitly adopts a principle from a lower court or another jurisdiction; the adopting opinion is the authority to cite. The drafts offer New York Times Co. v. Sullivan, which formulated its actual malice test by borrowing a rule a number of state courts had adopted.

Statutory Codification (Bridge 6) applies where a legislature enacts a judge-made rule, changing its source, weight, and modifiability. The drafts point to Title VII's disparate impact provision, 42 U.S.C. § 2000e-2(k)(1), which states in statutory text a burden structure that had been judge-made. The condition is correspondence: the text must actually state the rule for which it is cited.

Reasoning-based bridges

In the reasoning-based bridges the advocate constructs the connection. What makes the construction lawful is not ingenuity but the satisfaction of conditions the law has already fixed.

Recognized Analogical Extension (Bridge 3) carries a prior case's reasoning, not its holding, to new facts. It requires functional similarity: the principle that animated the earlier decision must apply for the same reason, and the limits that bounded it must remain unchanged or be accounted for. The drafts' paradigm is J.E.B. v. Alabama ex rel. T.B., which extended the rule of Batson v. Kentucky against race-based peremptory strikes to strikes based on gender. They also use Katz v. United States, and its principle that the Fourth Amendment "protects people, not places," to show a principle's reach and its boundary: United States v. Miller declined to extend it to bank records conveyed to a third party. Recognition, the drafts add, is a matter of degree, ranging from extensions the source court itself endorsed down to novel proposals no authority has accepted. Only the upper levels support a statement of law.

A Fortiori Reasoning (Bridge 7) holds that a result reached where the case was weaker follows where it is stronger. The drafts trace the form to the Talmudic kal va-chomer and the Roman pair a maiore ad minus and a minore ad maius. It requires an ordering grounded in authority rather than asserted, the same constitutional or statutory text governing both cases, the new case standing higher on that ordering, and no countervailing authority. The drafts catalog its failures: a false hierarchy; a dimension mismatch, where the ordering invoked for the precedent is not the one invoked for the new case; the ceiling problem, where the argument climbs past the top of one hierarchy into another domain; a missing justification; and competing hierarchies. In Kyllo v. United States, on the drafts' reading, majority and dissent disagreed less about the facts than about what the hierarchy measured: activity inside the home, or the invasiveness of the technology.

Negative Inference, or Expressio Unius (Bridge 8), applies where a court or statute expressly limits a rule to enumerated circumstances. The enumeration must be exhaustive ("only," "shall include," "means"), not illustrative ("such as"), and the omitted item must have been within the scope of consideration. The bridge separates a gap, which the court did not address, from an exclusion, which it implicitly made.

Structural Constitutional Inference (Bridge 9) derives a conclusion from federalism, separation of powers, or supremacy rather than from a particular text, as McCulloch v. Maryland did. Courts must have recognized the structural relationship explicitly. An appeal to structure no court has made is a policy argument in constitutional vocabulary.

Evolution-based bridges

Intra-Doctrinal Evolution (Bridge 4) extends a principle developed step by step within one line of cases, such as the privacy line from Griswold v. Connecticut through Eisenstadt v. Baird to Roe. The constraint is the silo: every case must belong to the same provision, statutory scheme, or common-law doctrine. Evolution within a silo is a bridge. Evolution across silos is the first fault.

Overruling Chain Completion (Bridge 10) addresses the case overruled in part. The later decision must expressly overrule some portion and preserve others, and the proposition cited must fall within what was preserved.

Policy-based bridges

Identified Policy Rationale Extension (Bridge 11) permits a rationale the court explicitly stated to reach new circumstances implicating the same concern. A purpose the court never stated cannot enlarge what it decided.

Stare Decisis Weight Factors (Bridge 12) is of a different logical kind. It derives no new conclusion; it argues that an existing rule is secure because of reaffirmation, reliance, and integration into the law, the marks of what has been called super-precedent. The drafts are candid about its limits. Dobbs overruled Roe, often described as approaching that status, and Loper Bright Enterprises v. Raimondo overruled Chevron notwithstanding forty years of reliance. Weight factors describe what makes a precedent costly to abandon. They do not establish that a court will decline the cost.

Eight ways a therefore fails

Each fault resembles a bridge from a distance and lacks the condition that would make it one.

Thematic clustering is the paradigmatic inferential hallucination, and the Brown, Gideon, and Roe synthesis is its textbook form. Temporal reversal attributes to an earlier case a characterization that appeared only later; precedents do not retroactively acquire the meanings later courts assign them. Test conflation carries a test into a domain governed by a different framework. The drafts' example, Chevron deference applied to criminal statutes where the rule of lenity governs, has since acquired a second lesson, because Chevron no longer states the law. An implicit overruling claim treats tension between two decisions as though a court had resolved it. On the framework's position, tension is a reason to ask a court to resolve it, not a finding that it has.

The drafts explain why these faults are characteristic of machine synthesis. Dissents are often more quotable than majorities. Models trained on text discussing cases in their modern interpretation project that interpretation backward. And retrieval by textual resemblance measures thematic similarity, not functional similarity: it finds cases that discuss the same subject, not cases that stand for the same rule.

Valid, rhetorical, and the space between

On the framework's account, an inference is valid when it crosses at least one bridge with every condition satisfied and triggers no fault revealing the crossing to be illusory. It is rhetorical when its conclusion depends on a connection no recognized mechanism supplies.

The framework is careful not to treat rhetorical as worthless. It distinguishes a synthesis that is not law from one that is bad law. A conclusion that crosses no bridge may still be a legitimate argument about what the law should be; it must be presented as argument. The proposed validation report keeps the distinction in three ratings. Legal Authority runs from BINDING through PERSUASIVE to ARGUMENTATIVE. Synthesis Quality runs from SUPPORTED through QUALIFIED to UNSUPPORTED. Advocacy Viability asks separately whether the proposition is worth arguing.

Applied to the test case, no bridge is confirmed and the first fault is triggered: different provisions, no shared text, no synthesizing decision. The synthesis is ARGUMENTATIVE and UNSUPPORTED. The recommendation is to relocate the idea, not abandon it: narrow it to one doctrine, search for a synthesizing decision, or present it openly as a novel argument.

Two features keep the framework honest. It evaluates structure, not substance; an inference that crosses a valid bridge may still fail on the merits, and the drafts compare the check to type-checking, which confirms that operations are well formed without confirming the answer. Structural validity is necessary, not sufficient. And the checks run in parallel, because they are independent. A synthesis may cross no bridge and commit no fault, in which case it is novel argument rather than error, or it may appear to cross a bridge while committing a fault, as when a quotation that would satisfy Bridge 1 proves to come from a dissent.

Where the framework is exposed

Three points invite scrutiny.

The first is completeness. The claim that twelve bridges exhaust lawful synthesis rests on Lamond's observation that reasoning from precedent proceeds by rule application, analogical reasoning, or balancing of reasons. The framework assigns Bridges 1, 2, 5, 6, 8, and 10 to the first mode; 3, 4, 7, 9, and 11 to the second; and 12 to the third, and argues that any new bridge would specialize an existing mode. That is an argument from classification, only as strong as the classification. It is not a proof, and the framework does not claim one.

The second is overlap. J.E.B. appears as the paradigm of Recognized Analogical Extension in one draft and as a case study in A Fortiori Reasoning in another. Nothing forbids an inference from crossing two bridges, but a system reporting which bridge was crossed should be able to report more than one, each independently satisfied.

The third is scope and measurement. The framework was built for Anglo-American common law, and the drafts concede that civil-law systems may need a supplementary taxonomy. They also concede that no benchmark of expert-annotated valid and invalid syntheses yet exists. The conditions are stated precisely enough to be checked. Whether a machine can check them reliably, given the difficulty of separating holding from dicta in an opinion's text, remains an empirical question.

The question to ask of every therefore

Whenever a brief or a machine writes therefore, ask which bridge the conclusion crosses, whether every condition is satisfied, and whether any fault shows the crossing to be illusory. If no bridge is crossed, the conclusion belongs in the argument section, labeled as argument, and not in the statement of governing law.

This constrains description, not advocacy. The difficulty with an inferential hallucination is not that it argues for something new. It is that it describes something new as though a court had already decided it. The apparent authority resides in the citations; the conclusion's support resides nowhere.

The formal statement is at The Twelve Bridges, within the program on Legal Knowledge Engineering.