The Case You Never Pulled

Every legal AI instrument now on the market grades the contents of a draft. Motions are lost on the complement.

That sentence is not rhetoric. It is a statement about inputs. A citation checker receives a document and evaluates the citations inside it: does each one exist, does it say what the document says, is it still good law. Those are properties of the document's contents. An omission is not a property of the contents. It is a property of the relation between the document and a specification of what the document should have contained. A checker that receives only the document cannot compute that relation, because absence has no representation in its input.

This paper defines four classes of omission, explains why dispositive motions make them tractable, and specifies a protocol a lawyer can run before filing, with or without AI.

Short answer. You know your research is complete enough to file when every element of every claim at issue has five things recorded against it: the governing standard, controlling authority, record support (or a pointed absence), a check for adverse controlling authority, and a check for courts that have already rejected your argument. Each check must end either in a citation or in a logged search that found nothing. A blank cell means a question nobody asked.

1. Definitions

Let D be the draft. Let S be the specification the draft must satisfy: the claims at issue, their elements under the governing law, the burden on each element, and the record.

Four omission classes follow.

Two notes on the table.

O2 sits upstream of everything else. Retrieval, generation, and verification are all invoked by questions. An element that no one asks about produces no query, no citation, and no error. It is the only failure that no downstream instrument can observe even in principle.

O4 is distinct from O1 in a way that matters for search design. An O1 query asks for authority on a rule. An O4 query asks for authority on an argument: a specific inference from rule to conclusion that some court has already considered and declined to draw. The two queries retrieve different documents. A lawyer who runs only the first will not find the second.

2. Why Dispositive Motions Make Completeness Computable

General legal research has no completeness denominator. Ask what the law of promissory estoppel is in a given state, and there is no principled answer to how many authorities make the answer complete.

Summary judgment is different in kind. The movant must show that there is no genuine dispute as to any material fact and that it is entitled to judgment as a matter of law, and materiality is defined by the governing substantive law. The substantive law supplies an element list, which has three useful properties.

• Enumerable in advance. The elements can be listed before research begins. The list does not depend on what a search returns.
• Conjunctive. A movant must prevail on every element; a nonmovant needs a genuine dispute on only one. An element the brief never addresses is therefore a detectable structural gap, independent of the quality of anything the brief does say.
• Decomposable. Each element requires a legal standard (which requires authority) and a factual showing (which requires record citation). Completeness decomposes into a matrix of cells that can be checked one at a time.

The record side is sharper still. Rule 56(c)(1)(A) requires citation to particular parts of materials in the record, and Rule 56(c)(3) provides that the court need consider only the cited materials, though it may consider others. Courts of appeals have repeatedly declined to require district courts to search the record for uncited evidence. The Tenth Circuit has described the concern as courts becoming advocates who comb the record and make a party's case for it. The practical consequence: at summary judgment, a record fact that exists and is not cited is a record fact the court is permitted to disregard.

A motion to dismiss has a narrower specification (elements against allegations rather than evidence), but the same matrix applies with the record column replaced by the pleading.

3. The Protocol

The protocol has eight steps. Steps 0 and 1 construct the specification. Steps 2 through 6 fill it. Step 7 terminates it.

Step 0. Fix the frame

Record, before any search:

• the forum, and the governing law for each claim (in diversity, the forum state's choice of law rules select it);
• the stage and the standard the motion invokes;
• the burden on each claim at trial, which determines the summary judgment route available to the movant.

Output: a frame record F. Every posture error described in the companion paper, Right Law, Wrong Stage, originates in an unrecorded or incorrect F.

Step 1. Enumerate the element set E

For each claim and each affirmative defense at issue, list the elements from controlling authority in the forum. Tag each element with the party bearing the trial burden and the route that follows: in federal court, the Celotex showing of an absence of evidence or the Adickes affirmative showing; in California, the Aguilar requirement that even a no-evidence showing be made with evidence.

Cross-check the list against the forum's pattern jury instructions where they exist (CACI in California; the circuit's pattern instructions in federal court). A pattern instruction is not authority, but a discrepancy between it and your element list is a signal to look again.

Step 2. Build the matrix M

Construct M = E × {Standard, Controlling authority, Record, Adverse, Rejected}.

Each cell must end in exactly one of three states:

No cell may be blank. A blank cell is not an empty result. It is a question that was never asked.

Step 3. Authority sweep (O1)

For each element, retrieve top-down by hierarchy: the court of last resort in the forum, then the intermediate appellate court whose decisions bind the trial court, then persuasive authority. Run every authority through a citator. Then apply the posture filter: for each authority, confirm that the proposition cited does not depend on a standard, record, burden, or forum different from the motion's.

Step 4. Record sweep (O3)

For each element with a contested fact, list every record location bearing on it: deposition pages and lines, declarations, discovery responses, Bates ranges. Cite pinpoints. For the nonmovant, the evidence that creates the genuine dispute must be cited, not merely present in the record. In districts with a separate statement rule, every response paragraph must carry its own citation, since several local rules deem uncontroverted facts admitted.

Step 5. Adverse sweep (O1, adverse subclass)

For each controlling authority in M, retrieve the decisions that limited it, distinguished it, or carved an exception from it. Search each rule's exceptions directly rather than waiting for them to surface in a topical search. Model Rule 3.3(a)(2) requires disclosure of directly adverse controlling authority known to the lawyer and not disclosed by opposing counsel; this step is how the lawyer comes to know it.

Step 6. Negative-space sweep (O4)

Restate each argument in the draft as a single proposition of the form rule R, applied to fact pattern P, yields conclusion C. Search for that proposition joined with rejection language in the forum: rejected, unpersuasive, declined to extend, not persuaded. Search the assigned judge's prior rulings on the same issue where they are available. For each hit, record the court's stated reason. The draft must then either distinguish the rejection or drop the argument.

Step 7. Terminate and log

The protocol terminates when no cell of M is blank.

It does not prove completeness. Nothing can, over an open corpus of authority. What it does is convert completeness from an unfalsifiable state of confidence into an inspectable artifact: a frame, an element list, a filled matrix, and a dated search log. That artifact can be reviewed by a supervising partner, defended to a client or carrier, and reconstructed if a court later asks how the brief was made.

4. Running the Protocol With AI

The protocol does not depend on any tool. It does change how a tool should be used.

One design rule governs the whole table: ask the tool cell-level questions, not document-level ones. A system asked to "draft the summary judgment motion" answers the questions it infers and gives no signal about the ones it skipped. A system asked to fill cell (element 3, adverse authority) either fills it or reports that it could not, and either result is visible.

This matters because of how people use reliable automation. Human-factors research on automation bias has found that operators working with a highly reliable aid miss more of the events the aid fails to flag. A draft that returns from a checker marked clean is exactly such an aid. The matrix counteracts the effect by making every unasked question look like what it is: an empty cell.

5. Self-Check Metrics

Before filing, compute three quantities. They are coarse, and they are sufficient.

• Element coverage = elements addressed ÷ elements in E. The only acceptable value is 1.
• Blank cells in M. The only acceptable value is 0.
• Unlogged negative results. Every "searched, none found" cell must carry its query and date. The only acceptable count of unlogged cells is 0.

None of these measures whether the brief is persuasive. Persuasion is a different problem. These measure whether the brief was built against a specification, which is the precondition for everything else.

Sources: Fed. R. Civ. P. 56(a), 56(c)(1)(A), 56(c)(3); Anderson v. Liberty Lobby, Inc., 477 U.S. 242, 248 (1986); Celotex Corp. v. Catrett, 477 U.S. 317, 323, 325 (1986); Adickes v. S.H. Kress & Co., 398 U.S. 144, 157-58 (1970); Aguilar v. Atlantic Richfield Co., 25 Cal.4th 826, 854 (2001); Adler v. Wal-Mart Stores, Inc., 144 F.3d 664, 671-72 (10th Cir. 1998); Model Rules of Prof'l Conduct R. 3.3(a)(2); Parasuraman & Manzey, Complacency and Bias in Human Use of Automation, 52 Human Factors 381 (2010).