Direction is a property of a holding, not of a case.

Lawyers speak of "defense cases" and "plaintiff's cases," but opinions do not wear team jerseys. An opinion is a bundle of holdings, and they can cut in opposite directions on the same page.

Blank v. Kirwan is cited in California demurrers for the rule that a demurrer admits material facts properly pleaded but not contentions or conclusions, a rule the defense likes. The same page holds that denying leave to amend is an abuse of discretion when there is a reasonable possibility the defect can be cured, which is the plaintiff's way out. Aguilar v. Atlantic Richfield Co. lets a defendant win summary judgment by showing that the plaintiff lacks, and cannot reasonably obtain, needed evidence, and the same opinion requires the defendant to make that showing with evidence. Cite the first half, ignore the second, and opposing counsel can quote your own authority back to you.

For machines this is a representation problem. A holding and its exception share vocabulary, facts and doctrine, so they sit next to each other in any similarity space. Retrieval that is blind to stance returns both as relevant and cannot tell the user which one hurts. The computational jurisprudence tradition identified directionality as a requirement long ago: HYPO treated argument as adversarial and CATO made factors side-relative. Document-level systems that lack it suffer what we call directional opacity.

The fix is to record stance on each holding, so that "cases against my position" becomes a query the system can actually answer. If direction is left for the model to infer on the fly, the lawyer is back to skimming a topical pile and hoping.