Driving a loaded tractor-trailer weighing 20 tons, Billy Bammerlin struck the right rear corner of another rig at approximately 25 miles per hour. The left half of Bammer-lin’s cab decelerated rapidly; pushed by the weight of the trailer, the right half of the tractor (which was not in contact with the other vehicle) pivoted away. The cab disintegrated. Bammerlin wound up on the ground with serious injuries, which have three possible explanations:

1. Bammerlin was not wearing a seat belt.

2. Bammerlin was belted but was crushed in the cab by the mass of the trailer he was hauling.

3. Bammerlin was belted, escaped serious injury in the cab, and was ejected and injured when he hit the ground.

Bammerlin argues for possibility (3), which poses the question how the seat belt failed.[*900] A jury awarded Bammerlin $500,000, implicitly finding that he fastened his seat belt and rejecting possibility (2). Navistar concedes that the evidence permitted the jury to reject (1) and (2) but contends that the record nonetheless does not support possibility (3).

Bammerlin contends that Navistar, which built the tractor, designed the seat belt assembly improperly by anchoring one end to a door pillar and the other to the engine housing. When the cab broke apart, the engine housing and seat belt anchor separated from the portion of the cab containing the driver’s seat, releasing tension on the belt. Bammerlin adds that even if this placement of the anchorage was not negligent, the assembly did not comply with federal safety standards. In support of this second theory of defect, Bammerlin offered the testimony of several expert witnesses, each of whom testified that the Navistar vehicle did not comply with the federal standards, as the witness understood them. Navistar replied that these witnesses misunderstood the standards and asked the district judge to strike the testimony. The judge declined and, instead of interpreting the standards for the jurors, allowed them to reach their own conclusion about the meaning of the standards. That was a serious mistake. The meaning of federal regulations is not a question of fact, to be resolved by the jury after a battle of experts. It is a question of law, to be resolved by the court. Harbor Insurance Co. v. Continental Bank, 922 F.2d 357, 366 (7th Cir.1990); Specht v. Jensen, 853 F.2d 805 (10th Cir.1988) (en banc); United States v. Bashes, 649 F.2d 471, 478-79 (7th Cir.1980). But cf. United States v. Bilzerian, 926 F.2d 1285, 1294-95 (2d Cir.1991). These safety standards do not have the meaning Bammer-lin’s experts gave them.

For example, one of Bammerlin’s experts testified that the seat belt assembly was defective because it did not comply with Federal Motor Vehicle Safety Standard (fmvss) 208, 49 C.F.R. § 571.208, which the witness understood to require that the assembly withstand a crash into a fixed barrier at any speed up to and including 30 m.p.h. See fmvss 208 S4.1.1.3. That is indeed what S4.1.1.3 says—but that part of fmvss 208 (like all other subparts of S4.1) applies only to passenger cars. See fmvss 208 S4.1. Trucks with a gross weight exceeding 10,000 pounds are governed by subpart S4.3 of fmvss 208, which lacks any provision comparable to subpart S4.1.1.3, for the simple reason that the greater mass of larger vehicles may make it infeasible to comply with the same standards applicable to passenger cars. Trucks exceeding 10,000 pounds need to comply with the design standards of fmvss 209, 49 C.F.R. § 571.209, rather than the rules applicable to passenger ears. The district court should have excluded this testimony and the corresponding theory of liability.

Another of Bammerlin’s experts testified that the Navistar truck failed fmvss 209 .and 210, as he understood them. This expert believed that the belt anchorage did not comply with fmvss 209 S4.1(f) and S4.4(a) because in a test it separated at 4,210 pounds of force rather than 5,000. He believed that it failed to comply with fmvss 210 S4.2.1 for the same reason: given the separation of the anchorage at less than 5,000 pounds of force, the expert believed that the assembly could not satisfy the requirement in S4.2.1 that “the pelvic portion of a Type 2 seat belt assembly shall withstand a 5,000-pound force when tested in accordance with S5.1.” The difficulty with this testimony is that the relevant portions of fmvss 209 and 210 speak of 5,000 pounds of force applied to the assembly or belt. Because each assembly has at least two anchorages, a 5,000-pound force applied to the belt subjects each anchorage to considerably less than that amount. † Once again the district court either should have excluded[*901] the testimony or instructed the jury that as a matter of law Navistar’s seat belt assembly complied with the federal requirements.

Both experts conceded on the stand that they were unfamiliar with the legal interpretation of the safety standards and that their test protocols did not conform to those specified by the National Highway Transportation Safety Administration. The district judge conceived of this as a problem of credibility: Navistar could argue to the jury that its experts, who were familiar with fmvss 208, 209, and 210, should be believed. It is not simply a credibility issue. A district judge should assure himself, before admitting expert testimony, that the expert knows whereof he speaks. Daubert v. Merrell Dow Pharmaceuticals, Inc., - U.S. -, 113 S.Ct. 2786, 125 L.Ed.2d 469 (1993). Plaintiff’s experts did not. All questions of testing method to one side, however, the initial step here was one of legal interpretation. What do the safety standards mean? The district judge should have resolved that question and provided the jury with the proper answer, so that experts for each side could address their testimony to the governing standards. By treating the meaning of the rules as if it were an issue of fact, and the reliability of the tests as if it were an issue of credibility, the district judge left the jury adrift and permitted it to return a verdict on a basis that may have been legally and factually flawed. We cannot say that the error was harmless. As the district court observed, the evidence on Bammerlin’s other theory (that the anchorage was defective because attached to the engine tunnel housing) was sharply divided. For all we can tell, the jury may well have resolved that question in Navistar’s favor and returned a verdict for Bammerlin only because it believed that the assembly flunked fmvss 208, 209, or 210.

Compliance with federal vehicle safety standards does not prevent a plaintiff from appealing to theories of products liability under the applicable state law of torts. 15 U.S.C. § 1597; Myrick v. Freuhauf Corp., 13 F.3d 1516 (11th Cir.1994). Thus we come to the question whether a jury could find the seat belt anchorage defective under Indiana law (which the parties agree applies) because it was attached to the engine housing rather than the floor of the cab. Bammerlin seems to think that the answer must be “yes” because, had the anchorage been attached to the floor, it would not have pulled apart in this accident. That a product failed in a particular accident does not necessarily show “defect,” however. Placing the anchorage in the engine housing increased the risks in a particular type of crash (one in which the cab disintegrates) but may reduce the risk in other kinds of crashes. The best evidence of net effects would be a study of seat belt failures in trucks designed with the cab over the engine (the kind involved here). Some anchor the seat belt in the cab floor, some in the engine housing. Does the latter fail more frequently than the former? Is the difference statistically significant? If an anchorage in the engine housing is indeed defective, then the answer to both questions should be “yes.” Surprisingly, none of the expert witnesses in the case provided these data. Both the lawyers and the experts seemed to think that “defect” is a question of first principles, to be resolved by jurors as if they were engineers designing the first truck in the world rather than observers asking whether the design of a particular truck unduly increased the risk of injury. Jurors are not engineers, and data on accident rates speak more loudly than abstract arguments. See Carroll v. Otis Elevator Co., 896 F.2d 210 (7th Cir.1990) (concurring opinion). But because neither side supplied data, the jurors were left to rely on intuition, and we agree with the district court that reasonable jurors could conclude that the placement of the anchorage was defective. The disagreement between Jackson v. Warrum, 535 N.E.2d 1207 (Ind.App.1989), and Masterman v. Veldman’s Equipment, Inc., 530 N.E.2d 312 (Ind. App.1988), about the proper way to articulate the standard of products liability in Indiana does not affect this conclusion. See Miller v. Todd, 551 N.E.2d 1139 (Ind.1990) (citing both Jackson and Masterman with approval).

Navistar disputes this conclusion on the ground that Bammerlin’s injury depended not only on the failure of the anchorage (which loosened the belt) but also on Bam-merlin’s getting out of the loose belt. Bam-[*902] merlin’s experts testified that the failure of the anchorage caused a tether bar (which carried some of the load) to bend, relaxing the webbing, which popped out from under Bammerlin’s abdomen, exposing the buckle, which was then hit and unlatched by a flying object. Navistar responds that folds of tummy fat are not part of the truck’s safety equipment, that the buckle was designed to be exposed, and that, if the buckle itself was not defective (no one says that it was), the failure of the anchorage cannot have rendered the vehicle defective. This is not, however, a complete answer.

Suppose the probability of the latch opening in a crash is 0.0001 if both anchors hold and 0.0002 if one anchor fails. Neither probability can be reduced by a redesign of the latch, which therefore cannot be called “defective.” Suppose further that the loss if the latch opens in an accident is $500,000. Then the expected costs per vehicle attributable to belt opening are $50 if both anchors hold and $100 if only one anchor holds. If it costs, say, an extra $10 to ensure that an anchor holds (as by securing it to the cab’s floor rather than its engine tunnel), then a prudent designer will incur the cost—and the vehicle is defective if an anchor fails. This is nothing but an application of Learned Hand’s formula for negligence, B<PL (where B is the burden of precautions, L the loss if there is an accident that the precautions could have prevented, and P the probability of an accident if the precautions are not taken). United States v. Carroll Towing Co., 159 F.2d 169, 173 (2d Cir.1947). Our court has applied Judge Hand’s approach in many kinds of negligence actions, see McCarty v. Pheasant Run, Inc., 826 F.2d 1554, 1556 (7th Cir.1987), and as the definition of a product defect in Indiana depends on general principles of negligence, see Miller, 551 N.E.2d at 1141, we have no reason to think that state would see things otherwise. Thus the fact that the probability of a particular failure is low is no defense if the costs of protecting against it are even lower. See ALI, Restatement of the Law—Torts: Products Liability § 2(b) and Reporters’ Notes at 4(M5, 123-25 (Tent. Draft No. 1, 1994).

All of this supposes, however, that a flying object did unlatch Bammerlin’s seat belt. Otherwise the failure of the anchorage did not cause his injuries, and causation is part of the plaintiffs case under Indiana law. Crull v. Platt, 471 N.E.2d 1211, 1215 (Ind. App.1984). See also Products Liability § 6(a). Bammerlin need not prove causation directly. He did not have high-speed cameras running in the cab at the time of the accident, and nothing else could have caught a flying object in the act. Instead he proceeded by eliminating the alternatives. We know that he wound up outside the cab. How did he get there? Navistar’s theory is that he was not wearing a seat belt. Bam-merlin countered with his say-so, plus a physician’s testimony that some of his injuries are more consistent with wearing a belt than with the hypothesis that he was not wearing one. (“Seat belt burn,” the physician called it.) A biomechanics expert added that the injuries are most consistent with Bammer-lin’s being belted for part of the time during the crash. This is enough to permit a rational jury, see Mayer v. Gary Partners & Co., 29 F.2d 330 (7th Cir.1994), to find that the belt disengaged during the crash. Engineers trying to understand a disaster often follow causal chains (“failure trees”) until they find one that can account for the calamity. Sherlock Holmes observed that “when you have eliminated the impossible, whatever remains, however improbable, must be the truth”. A. Conan Doyle, The Sign of Four ch. 6. Courts need not disdain a method that both engineers and detectives find useful.

Bammerlin produced evidence that could lead a rational jury to eliminate the hypotheses inconsistent with his favored theory, which in turn permits an inference that his hypothesis is true. Although the jury’s verdict was unreliable, for reasons we have explained, Bammerlin is entitled to try again.

REVERSED AND REMANDED.