Ismael Ornelas-Ledesma and Saul Ornelas were convicted of illegal possession of a controlled substance with intent to distribute it and were sentenced to 60 and 63 months in prison, respectively. The appeal challenges the denial of their motion to exclude from evidence the cocaine seized from "their" 1981 Oldsmobile. We say "their" Oldsmobile, although it was registered to someone else, because the government does not question their right, akin to that of a tenant, bailee, borrower, or other lawful occupier or possessor, to object to the seizure. The cases are quicker to draw the analogy in the case of the driver of a vehicle, United States v. Garcia, 897 F.2d 1413, 1418-19 (7th Cir.1990); United States v. Soto, 988 F.2d 1548, 1553 (10th Cir.1993); United States v. Miller, 821 F.2d 546, 548-49 (11th Cir.1987), than in the case of a passenger, Rakas v. Illinois, 439 U.S. 128, 148-49, 99 S.Ct. 421, 433, 58 L.Ed.2d 387 (1978); United States v. Lechuga, 925 F.2d 1035, 1037, 1041 n. 3 (7th Cir.1991); United States v. Roberson, 6 F.3d 1088, 1091 (5th Cir.1993), but the government does not ask us to make that distinction here.

The defendants argue that the stop which led to their consenting to the search of the car violated the Fourth Amendment, invalidating the consent and hence the search; and that independently of this the seizure of the drugs, which was from the inside of the door of the car, violated the Fourth Amendment because, even if a superficial search of the car was proper as an incident of the stop, the officers needed and lacked probable cause to open the compartment and seize its contents.

Police officers in Milwaukee keep a regular watch on motels, looking for drug runners. Cruising through a motel parking lot one day in 1992, a detective named Pautz, a two-year veteran of the Milwaukee County Sheriff's Drug Enforcement Unit, spotted a 1981 two-door Oldsmobile with a California license plate. Pautz was interested. Two-door General Motors cars of that vintage are believed by law enforcement authorities to be drug traffickers' favorites (though not their only favorites, United States v. Sharpe, 470 U.S. 675, 682 n. 3, 105 S.Ct. 1568, 1573 n. 3, 84 L.Ed.2d 605 (1985); United States v. Ocampo, 890 F.2d 1363, 1366 (7th Cir.1989)) because it is easy to conceal drugs in them. And California, like other states on the eastern, western, and southern borders of the United States, is a state from which drugs are shipped to other states, a "source" state. So Pautz called up his dispatcher on his car radio and asked him to find out whom the car was registered to. Now in fact it was registered to "Ornelas, Miguel Ledesma," of San Jose, California, who may or may not be related to one of the defendants. But we do not know what exactly the dispatcher told Pautz or what exactly Pautz understood--whether it was "Miguel Ledesma Ornelas" or "Miguel Ornelas Ledesma." Confusion on this score is understandable, though not justifiable. Spanish naming conventions are confusing to non-Hispanic Americans. When a Hispanic has two surnames, such as Ornelas Ledesma or Ledesma Ornelas, the first is his father's last name and the second his mother's maiden name. The first is primary, and the second subordinate--exactly the reverse of the middle and last names of non-Hispanics. To a Hispanic, therefore, "Ornelas, Miguel Ledesma," would denote Miguel Ornelas Ledesma rather than, as a non-Hispanic would expect, Miguel Ledesma Ornelas. But if the motor vehicle authorities, the dispatcher, or Pautz were unfamiliar with Spanish naming conventions (and Pautz testified that he was unfamiliar with them), "Ledesma" and "Ornelas" could easily get reversed.

Pautz, believing that he had the name of the car's registered owner, entered the motel and checked the registry, which showed that an Ismael (not Miguel) Ornelas had registered at 4:00 a.m. and had been accompanied by another man. Pautz then called his partner, Detective Hurrle, to join him. When Hurrle arrived, the two officers called the Milwaukee office of the Drug Enforcement Administration and asked it to run a NADDIS check on Miguel Ledesma Ornelas from San Jose and on Ismael Ornelas. NADDIS (Narcotic and Dangerous Drug Information System) is a computerized compilation of the federal Drug Enforcement Administration's information on known and suspected drug traffickers. NADDIS reported two "hits." One was on Miguel Ledesma Ornelas, also known as Miguel Lemus-Ledesma, identified by NADDIS as a heroin dealer in El Centro, California, which is hundreds of miles from San Jose, the residence of the registered owner of the car. The other "hit" was on Ismael Ornelas, Jr., of Tucson, Arizona, reported by the computer to be a 1000-kilogram-per-month cocaine dealer, although no "wants" or warrants were outstanding against him. The officers did not attempt to obtain descriptions of the two suspected dealers, neither of whom in fact is one of or, as far as we know, related to the defendants.

Because the car was an older GM two-door, because its occupants had checked in at the motel at 4:00 a.m. without having made reservations in advance, because there were two persons rather than just one and at least one of them was Hispanic, and because they apparently had come from a "source state"--and of course because of the "hits" on similar or identical names--the officers decided to stop and question the men when they left the motel. At this point a third officer arrived, Luedke, bringing with him a drug-sniffing dog; and Pautz left. After a time, two men emerged from the motel and entered the Oldsmobile. The officers parked their cars on either side of it. Hurrle tapped the window on the driver's side, identified himself as a police officer, and asked both men for identification. Saul Ornelas, who was in the driver's seat, produced a driver's license in that name; Ismael Ornelas-Ledesma, who was in the front passenger's seat, produced a driver's license in the name of Ismael Ornelas, residing in Martinez, California. Hurrle asked whether there were any drugs in the car. The occupants said "no." Hurrle asked them for permission to search the car, and they gave their permission. The two men were not placed under arrest but the government concedes that a reasonable person in their position would not have felt free to leave. So the encounter was a so-called "Terry stop" (after Terry v. Ohio, 392 U.S. 1, 88 S.Ct. 1868, 20 L.Ed.2d 889 (1968)), Michigan v. Chesternut, 486 U.S. 567, 573, 108 S.Ct. 1975, 1979, 100 L.Ed.2d 565 (1988); INS v. Delgado, 466 U.S. 210, 216, 104 S.Ct. 1758, 1762, 80 L.Ed.2d 247 (1984); United States v. Adebayo, 985 F.2d 1333, 1338-39 (7th Cir.1993), a semi-arrest that is lawful under the Fourth Amendment if but only if the officers had a "reasonable suspicion supported by articulable facts" that the persons stopped were engaged in criminal activity. United States v. Sokolow, 490 U.S. 1, 7, 109 S.Ct. 1581, 1585, 104 L.Ed.2d 1 (1989).

Officer Luedke inspected the interior without canine assistance and noticed a loose panel on the passenger's door. One of the screws by which the panel was fastened to the door was--Luedke testified--rusty. To him this meant that it had been removed recently. He didn't explain why, but maybe the idea is that the removal of the screw would have scraped off its chrome coating, which protected it from rusting. Suspecting that drugs were concealed behind the panel, Luedke pried it open. His hunch was correct; there was a package containing two kilograms of cocaine wrapped in aluminum foil and paper. Ornelas and Ornelas-Ledesma were then arrested.

Clearly, were it not for the NADDIS hits, the officers would not have had grounds for reasonable suspicion that the defendants were drug traffickers. Not only is every circumstance on which the officers relied other than the hits innocent taken by itself--many Americans (approximately one in eight) are Californians, many Californians are Hispanic, many Americans drive two-door General Motors cars, many people check into motels very late at night (or early in the morning), many travel in pairs rather than alone, and many do not make advance reservations--but the confluence of these circumstances is pretty innocuous as well, especially since many of the circumstances are correlated rather than independent. Hispanics are disproportionately concentrated in California, and having on average lower incomes than non-Hispanic Americans are doubtless more likely than other Americans to drive two-door rather than four-door cars, older rather than newer cars, and American rather than foreign cars. They are more likely to drive than to fly and, we imagine, less likely to make reservations in advance at motels, since cheap motels don't advertise much or have 800 numbers. Nothing is more common than for people taking long trips to drive until they're tired and then--often at very odd hours--to check in at the nearest motel, of course without a reservation. And people who drive long distances late at night prefer to have someone with them. Because the "suspicious" circumstances (other than the NADDIS hits) are so strongly correlated with each other, were they considered sufficient by themselves to justify a stop the practical consequence would be that a very large proportion of all Hispanic Americans would be vulnerable to being stopped on suspicion of drug trafficking. Hispanics would be second-class citizens in the eyes of the police. Although a brief investigatory stop is less intrusive than an arrest, it is sufficiently redolent of police-state tactics for courts to insist in the name of the Fourth Amendment that its use be circumscribed.

So we must ask what the NADDIS hits added to a "drug courier profile" that seems only a little better than a dragnet for Hispanics. It would be nice to know something about NADDIS, but the government successfully opposed the defendants' efforts to obtain discovery aimed at determining the character and reliability of the information in the NADDIS data base, and as a result the record is bare of evidence about it. At argument the government's lawyer, while saying that she has used NADDIS herself, disclaimed any knowledge about the system except how to access it. We do not know how many names are in it currently, where exactly the names come from, whether any of the information inputted into the system is screened for reliability before being entered, whether anyone checks to make sure that errors are not made in inputting information, whether information is updated systematically, and (a closely related question) whether information discovered to be stale or inaccurate is removed from the system. Although a number of judicial opinions mention NADDIS in passing--obviously it is widely used in drug enforcement--we have found no considered judicial assessment of its reliability, although we have found occasional judicial expressions of concern about that reliability. United States v. $7,850 United States Currency, 7 F.3d 1355, 1358 (8th Cir.1993); United States v. $215,300 United States Currency, 882 F.2d 417, 419 and n. 2 (9th Cir.1989) (per curiam); United States v. Saperstein, 723 F.2d 1221, 1232 (6th Cir.1983).

Maybe NADDIS is no better than a vast compendium of rumors, errors, and libels: garbage in, garbage out. That seems unlikely. It would not be heavily used by drug enforcement authorities if it were merely a random sample of the American population. Which is not to say, however, that it is highly reliable; concern that it may not be is heightened by the (scanty) secondary literature, which depicts NADDIS as an unselective, unweeded repository of unsubstantiated allegations, often dated. See UPI Release, "VIP Names in Drug Agency's Computer Files," July 3, 1984; Vanessa Jo Grimm, "Behemoth DEA Database Tracks Drug Smugglers," Government Computer News, July 8, 1991, p. 85. According to the UPI Release, as of 1984 NADDIS contained 1.5 million names, obtained from debriefings of informants and suspects and from surveillance and intelligence reports of various law enforcement agencies.

As an example of NADDIS's possible unreliability, we note that according to it Ismael Ornelas, Jr., of Tucson, a 1000 kilogram a month dealer--a large dealer--is not wanted by any law enforcement authority in the country. Maybe that Ornelas (who is neither Saul Ornelas nor Ismael Ornelas-Ledesma nor the registered owner of the Olds) died years ago or is some practical joker's or desperate informant's fictional creation or is an honest man falsely accused by an enemy. Because the reliability of information stored in NADDIS is unknown, that information must, on the record before us, be reckoned no more reliable than that of an informant not known to be reliable. Therefore, since an uncorroborated tip from such an informant cannot by itself furnish probable cause for an arrest or search, Illinois v. Gates, 462 U.S. 213, 227, 103 S.Ct. 2317, 2326, 76 L.Ed.2d 527 (1983), an unverified, uncorroborated entry in NADDIS cannot do so either. United States v. $7,850 United States Currency, supra, 7 F.3d at 1358; United States v. $215,300 United States Currency, supra, 882 F.2d at 419 and n. 2; United States v. Saperstein, supra, 723 F.2d at 1232. The issue here, it is true, is not probable cause, but reasonable suspicion, because we are dealing with a stop rather than an arrest. But the Supreme Court has held that an uncorroborated anonymous tip, even when it comes from law enforcement authorities, does not by itself justify a stop. United States v. Hensley, 469 U.S. 221, 232, 105 S.Ct. 675, 682, 83 L.Ed.2d 604 (1985). The statement in United States v. Troka, 987 F.2d 472, 474 (7th Cir.1993), that "the police department was a reliable source," must be understood in context (the police department was a reliable source in the circumstances of that case ), rather than as a quixotic effort to overrule the Supreme Court.

That cannot be the end of our analysis, however. NADDIS furnished two tips, not one. "Ismael Ornelas," the name on the motel registry, was very similar to "Ismael Ornelas, Jr.," the name in NADDIS. And while the names Miguel Ornelas Ledesma and Miguel Ledesma Ornelas are as different from each other as Francis Scott Key and Francis Key Scott, the names could easily have been reversed in entering them in the computer, just as "Jr." could easily have been left off a motel registry. So there was some reason to believe that the registered owner of the Olds was the Miguel Ornelas Ledesma, or Miguel Ledesma Ornelas, identified by NADDIS as a suspected drug dealer. Even if NADDIS is not terribly reliable, some or for that matter many entries in it may be accurate or at least approximately so. And the fact that both the registered owner of the car and the name on the motel registry corresponded more or less to names of suspected drug dealers in the NADDIS database was impressive. One "hit" might well be coincidence; it was much less likely that two--one corresponding to the name in the hotel registry, presumably that of either the driver or the passenger; the other to the name of the registered owner, who might or might not be one of the two men in the motel--were, though this depends in part on how common the names "Ornelas" and "Ledesma" are. Paradoxically, the very unreliability of NADDIS (if it is unreliable) might strengthen the inference that the two men in the Olds were involved in the drug trade; inverting last and middle names is just the kind of mistake that one would expect in a carelessly maintained database. A mistaken premise can furnish grounds for a Terry stop, if the officers do not know that it is mistaken and are reasonable in acting upon it. Illinois v. Rodriguez, 497 U.S. 177, 184-86, 110 S.Ct. 2793, 2799-2800, 111 L.Ed.2d 148 (1990); United States v. De Leon-Reyna, 930 F.2d 396, 399 (5th Cir.1991) (en banc) (per curiam); United States v. Walraven, 892 F.2d 972, 974-75 (10th Cir.1989).

The tips were mutually corroborating; and the other circumstances on which the officers relied (the make and type of the car, the state it came from, etc), insufficient as they were by themselves to create a suspicion reasonable enough, substantial enough, to justify the infringement of personal liberty that is brought about by an investigative stop, may, when taken together with the NADDIS "tips," have tipped the balance in favor of a finding of reasonable suspicion. United States v. Esieke, 940 F.2d 29, 34 (2d Cir.1991), and United States v. Morin, 665 F.2d 765 (5th Cir.1982), factually similar cases although stronger for the government, so suggest. And the Supreme Court held in Alabama v. White, 496 U.S. 325, 110 S.Ct. 2412, 110 L.Ed.2d 301 (1990), that an anonymous tip can create reasonable suspicion if corroborated. The form the corroboration took in that case was verification of a prediction made by the informer, but there is no magic in a particular form of corroboration. The circumstances that first aroused Pautz's suspicions were corroborative of the "anonymous tips" furnished by NADDIS in the sense that they made it likelier than otherwise that NADDIS was, in this instance anyway, accurate. Standing alone, the drug courier profile that first aroused Pautz's suspicions was very little. Standing alone, one NADDIS "hit" could not, in the absence of any evaluation of NADDIS's reliability, be thought much either. The second "hit," however, added to the credibility of the first; and together the profile and the "hits," though still not enough evidence of crime to furnish probable cause to search the car, could satisfy the lesser showing required for a Terry stop.

Although the question is close, we conclude that the district court did not commit clear error in holding that there was a lawful stop. And clear error is the test; the proposition that review of a district court's determination of reasonable suspicion (if it is a Terry stop) or probable cause (if it is an arrest, search, or seizure) is plenary, embraced in such cases as United States v. Jaramillo, 891 F.2d 620, 626 (7th Cir.1989), is no longer tenable after United States v. Spears, 965 F.2d 262 (7th Cir.1992), where, overruling inconsistent precedents, this court held that the standard for judicial review of determinations of probable cause was clear error. There is no basis for distinguishing in this regard between probable cause and reasonable suspicion. They are adjacent points on a continuum.

Since the stop was lawful, we need not decide whether, had it been unlawful, the defendants' consent to search would nevertheless have been valid even though (a relevant factor, though not a decisive one, Schneckloth v. Bustamonte, 412 U.S. 218, 231-33, 249, 93 S.Ct. 2041, 2049-51, 2059, 36 L.Ed.2d 854 (1973)) the officers failed to advise the defendants of their right not to consent. United States v. Valdez, 931 F.2d 1448, 1451 (11th Cir.1991); United States v. Valencia, 913 F.2d 378, 381 (7th Cir.1990). But we do not think the denial of the motion to suppress the drugs seized from the door of the Olds can be upheld on the basis of the record compiled in the district court. The government does not argue that a Terry stop justifies so intrusive a search. The only lawful purpose of a search incident to a Terry stop is to protect the officers from the danger that the persons they have stopped will reach for a weapon, Minnesota v. Dickerson, --- U.S. ----, ----, 113 S.Ct. 2130, 2136, 124 L.Ed.2d 334 (1993); Michigan v. Long, 463 U.S. 1032, 1049-51, 103 S.Ct. 3469, 3481-82, 77 L.Ed.2d 1201 (1983), and ordinarily and here the officers do not have to dismantle a car to satisfy themselves on that score. Nor does the government argue that the consent to search that Ornelas and Ornelas-Ledesma gave the officers was consent to dismantle the car. United States v. Garcia, supra, 897 F.2d at 1419-20; but cf. United States v. Pena, 920 F.2d 1509, 1515 (10th Cir.1990). The seizure was lawful, therefore, only if the search yielded information that gave Officer Luedke probable cause to believe that contraband was secreted behind the loose panel. If so, no search warrant was required, even for a search that would require taking the car apart. United States v. Ross, 456 U.S. 798, 102 S.Ct. 2157, 72 L.Ed.2d 572 (1982).

In arguing that there was probable cause for the search that discovered the drugs secreted in the car door, the government places great weight on the loose panel and the rusty screw. But the screw was not rusty. The magistrate judge to whom the motion to suppress was referred for a recommended ruling looked at the screw and found no indication of rust. The screw is in the appellate record and we have looked at it ourselves. There is not the slightest trace of rust or of anything that looks like rust. It is a Phillips screw and the head contains scratches that suggest it might have been removed and reinserted, but Luedke was adamant that it was signs of rust that caused him to believe that the panel had been removed recently and then put back in place. Having inspected the screw, the magistrate judge naturally disbelieved Luedke's testimony that the screw had looked rusty. Without discussing the bearing of Luedke's testimony about the loose panel on the issue of probable cause, the magistrate judge concluded that the seizure of the drugs had not been supported by probable cause. But noting the presence of the drug-sniffing dog, he found that the drugs would have been discovered anyway and therefore upheld the seizure on the ground of inevitable discovery. Nix v. Williams, 467 U.S. 431, 104 S.Ct. 2501, 81 L.Ed.2d 377 (1984).

The district judge upheld the seizure, too, but not on the magistrate judge's ground but rather on the ground that the allegedly loose panel alone, which the magistrate judge had not discussed, furnished probable cause. The government defends Luedke's testimony about the rust by arguing that it was dark inside the car (though it was broad daylight) and maybe the screw looked rusty though it wasn't; we are more accustomed to being told of the remarkable eyesight of the police that enables them to see tiny objects at great distances in the dark. United States v. Sophie, 900 F.2d 1064, 1073 (7th Cir.1990); United States v. Gilliard, 847 F.2d 21, 24 (1st Cir.1988); cf. United States v. Pinto-Mejia, 720 F.2d 248, 262 (2d Cir.1983). In any event, the magistrate judge was not required to believe Luedke or recast his testimony in a believable form.

The magistrate judge concluded that Luedke did not have probable cause to seize the drugs; the district judge concluded that he did. Can we uphold the district judge's determination, in view of this disagreement? The standard for judicial review of a determination of probable cause is, as we have noted already, clear error; and this is so whether the determination is made by a magistrate judge or by a district judge. United States v. Spears, supra; United States v. Adebayo, 985 F.2d 1333, 1337 n. 2 (7th Cir.1993). A qualification is necessary, however, for the case in which the magistrate judge, rather than determining probable cause en route to issuing a warrant later challenged by a motion to suppress evidence seized pursuant to it, is merely recommending the disposition of a motion referred to him by a district judge to suppress evidence obtained as a result of a search, or an arrest, made without a warrant. In the first case, the relation between district judge and magistrate judge is that of reviewing court to court of first instance, and the usual standard of appellate review of findings of fact and of the application of a legal standard to such findings applies. In the second case--which is our case--not only may the district judge (with inapplicable exceptions) "accept, reject, or modify, in whole or in part, the [magistrate judge's] findings or recommendations," 28 U.S.C. Sec. 636(b)(1); he is required "to make a de novo determination of those portions of the [magistrate judge's] report or specified proposed findings or recommendations to which objection is made." Id.

The question whether the district judge's power of de novo determination authorizes him to reject a magistrate judge's finding of credibility, without the district judge's hearing the witness himself, was left open in United States v. Raddatz, 447 U.S. 667, 681 n. 7, 100 S.Ct. 2406, 2415, 65 L.Ed.2d 424 (1980), though with a broad hint, picked up in Proffitt v. Wainwright, 685 F.2d 1227, 1237, 1241 (11th Cir.1982), that the answer is "no." See also Grassia v. Scully, 892 F.2d 16, 19-20 (2d Cir.1989); but cf. LoConte v. Dugger, 847 F.2d 745, 750 (11th Cir.1988). We need not answer the question here. Rather than reject the magistrate judge's finding that Luedke's testimony about the rusty screw should not be believed, the district judge found probable cause for the seizure in a part of Luedke's testimony that the magistrate judge had not relied upon--the testimony that the panel was loose. A number of cases, illustrated by United States v. Lugo, 978 F.2d 631, 637 (10th Cir.1992), treat a loose door panel as a telltale sign of drug running--but not as sufficient in itself to constitute probable cause. (In United States v. Garcia, supra, 897 F.2d at 1416, the officers saw packages peeping out from behind the loose panel.) The facts of this case show why. The car was eleven years old when Officer Luedke conducted the search. We mean no disrespect to Detroit in observing that the interior of a 1981 Olds unlike that of a Rolls Royce is apt to be rather worn and battered after eleven years. The fact that a door panel is loose does not by itself create a reasonable likelihood that the panel conceals a secret compartment containing contraband. State v. Swanson, 172 Ariz. 579, 838 P.2d 1340, 1345-48 (App.1992). Here as in most cases there was more, but not much more: mainly the NADDIS "hits" of unproven reliability.

We need not decide, however, whether the NADDIS hits (remember that there were two, and that this enhances their reliability), plus the fact that the defendants fit the drug courier profile, plus a loose door panel add up to probable cause--whether in other words the loose panel could turn reasonable suspicion into probable cause. For we do not know whether Luedke's testimony about the door panel should be believed. The only judicial officer who heard him testify was the magistrate judge, who failed to indicate whether he believed the testimony. Since the magistrate judge disbelieved a closely related part of Luedke's testimony, and the district judge did not overturn that credibility finding, it can hardly be assumed that the magistrate judge would surely have believed Luedke's testimony about the door panel, or that if he had disbelieved it would surely have been reversed by the district judge. Nor did the latter base his determination on a viewing of the car; there was no viewing. Even if, contrary to Proffitt, and to decisions involving the parallel case of administrative review of administrative law judges' credibility determinations, e.g., Universal Camera Corp. v. NLRB, 340 U.S. 474, 495-97, 71 S.Ct. 456, 468-469, 95 L.Ed. 456 (1951); NLRB v. Augusta Bakery Corp., 957 F.2d 1467, 1475-76 (7th Cir.1992); Ortiz-Salas v. INS, 992 F.2d 105, 108 (7th Cir.1993), a district judge can reject a magistrate judge's determinations of credibility without hearing the witnesses' testimony himself, we do not think that he can be permitted to ignore the issue of credibility altogether. We do not think that he can accept challenged testimony given before another judicial officer without giving a reason for resolving the issue himself rather than referring the matter for initial determination by the other officer.

This oversight would be academic if the magistrate judge's alternative ground for upholding the seizure--inevitable discovery by Luedke's dog--was so clearly correct that we could uphold it without the benefit of the district judge's views, as in United States v. Lewis, 621 F.2d 1382, 1387 (5th Cir.1980). We cannot. No evidence was presented concerning the dog's capabilities. For all we know he is an infallible sniffer of two kilograms of cocaine wrapped and hidden behind the door panel of a car; but we cannot take judicial notice of this fact. We don't know and have not been told anything about this dog except his name (Merlin), although our own research reveals that a Milwaukee drug-sniffing dog named Merlin, presumably the same beast, has on at least one occasion detected drugs in a piece of luggage when there weren't any. Schaefer v. United States, 656 F.Supp. 631, 632 (E.D.Wis.1987). If Merlin is that good, no doubt he could sniff the cocaine in the door of the defendants' car, but the government has not had the audacity to argue that a dog who can detect nonexistent drugs is an infallible inevitable discoverer of existent drugs. We may be too hard on Merlin; maybe he sniffed the residue of drugs in the luggage in the Schaefer case; maybe it was a different dog with the same name. But the government in its brief, without going so far to confess error on the magistrate judge's alternative ground (that is, inevitable discovery), does not defend it, even though defendant Ornelas had attacked it at some length in his opening brief. The government does not, for example, respond to Ornelas's damaging quotation from the testimony of Detective Hurrle: "I don't know if he [the dog] necessarily would have hit on this particular cocaine." (To which it can be added that we don't know whether the dog would have been directed to sniff in or about the car if Luedke's search had disclosed no suspicious circumstances.) We cannot affirm on a ground that has been waived, Crane v. Indiana High School Athletic Ass'n, 975 F.2d 1315, 1319 (7th Cir.1992); Frederick v. Marquette Nat'l Bank, 911 F.2d 1, 2 (7th Cir.1990), and the government has waived the defense of inevitable discovery by failing to defend the magistrate judge's determination. Perhaps it did this because it believed that the determination could not be defended. It put all its eggs in the probable-cause basket, and the eggs are broken.

The motion to suppress should not have been denied, though it is open to the government on remand to ask the district judge either to hear Luedke's testimony on the loose panel himself (or at least view the car's interior) or refer the matter to the magistrate judge for a determination of the credibility of that testimony. If Luedke's testimony is disbelieved, the motion to suppress should be granted. If it is believed, the district court must decide whether the loose panel, in conjunction with the NADDIS tips as corroborated by the circumstances that first aroused Officer Pautz's suspicions, adequately established probable cause to believe that drugs were concealed behind the panel. The court may wish to consider the evidence of the loose panel (if that evidence is believed) in conjunction with the Philips screw. For while it is plain that it is not rusty, it does look as if it had been scratched in just the way one would expect if it had been removed and reinserted, though its appearance could we suppose reflect its removal by Officer Luedke or other handling since. The test for probable cause is objective. It is not what the officer in question actually believed but what a reasonable officer in his position would have believed. Scott v. United States, 436 U.S. 128, 138, 98 S.Ct. 1717, 1723, 56 L.Ed.2d 168 (1978); Mahoney v. Kesery, 976 F.2d 1054, 1057-58 (7th Cir.1992). A reasonable officer would not have believed the screw rusty but might have believed it showed (other) evidence of having been removed and reinserted. That is a matter to be explored on remand.

The judgments of conviction are vacated and the case is remanded for further proceedings consistent with this opinion. The other grounds raised by the appellants have too little merit to warrant discussion.

VACATED AND REMANDED.