GORSUCH, Circuit Judge.

Garbage in, garbage out. Everyone knows that much about computers: you give them bad data, they give you bad results. There was a time when the enforcement of traffic laws depended on officers lying in wait behind billboards watching cars flow past. Today, officers nearly as often rely on distant computer databases accessed remotely from their dashboards, stopping passersby when the computer instructs. But what if the computer turns out to be a good deal less reliable than the officer's eagle eye? What if the computer suggests you've broken the law only because of bad data — garbage in, garbage out? Today's case requires us to wrestle with these questions for the first time, bringing the Fourth Amendment face-to-face with Charles Babbage.1

Our case began when Kansas Trooper Andrew Dean, a regular before this court, sat watching traffic along I-70. At some point, a minivan caught his eye. There was nothing special about the minivan except maybe the fact it bore a Colorado temporary 30-day registration tag. But even that didn't suggest anything amiss: the tag looked genuine enough, it was displayed in the right place, its expiration date hadn't yet passed. All the same, Trooper Dean decided to verify the tag with a law enforcement database. He called in the tag number to a dispatcher who soon replied "that's a negatory on record, not returning." Because of — and only because of — the dispatcher's "no return" report, Trooper Dean turned on his lights and stopped the minivan. After a brief discussion, the trooper sought and received permission to conduct a search, one that eventually yielded a secret compartment containing over a pound of methamphetamine and Mr. Esquivel-Rios's trial and conviction on federal drug charges.

Before and after trial the defense argued that Trooper Dean's traffic stop violated the Fourth Amendment and that all evidence found during the stop should be suppressed. But the district court disagreed. In its view, the trooper had reasonable suspicion to believe the minivan was displaying a forged tag. In Kansas, as elsewhere, vehicles must be registered with some lawful authority. When a law enforcement database yields no information about a registration tag, the court reasoned, that raises a non-trivial possibility the tag wasn't lawfully issued in the first place but falsified in some way.

The district court's reasoning was right as far as it went. This court and others have regularly upheld traffic stops based on information that the defendant's vehicle's registration failed to appear in a law enforcement database — at least when the record suggested no reason to worry about the database's reliability. See, e.g., United States v. Cortez-Galaviz, 495 F.3d 1203, 1205-06 (10th Cir.2007); United States v. Garcia-Ballesteros, 74 F.3d 1250, 1996 WL 3920, at *1 (10th Cir.1996) (unpublished table decision); see also United States v. Stephens, 350 F.3d 778, 779-80 (8th Cir.2003).

The difficulty we face in this case stems from that last and critical qualification. When Trooper Dean asked about the minivan's temporary tag, the dispatcher replied not only that the tag yielded a "no return" response from the queried database. The dispatcher also added that "Colorado temp tags usually don't return." And this is a piece of evidence our cases haven't confronted before: evidence admitted by a district court suggesting that the database on which the officer relied to justify his stop might bear a real problem — a problem that might mean a "no return" doesn't suggest criminal conduct but only some bureaucratic snafu.

In Cortez-Galaviz we took pains to reserve judgment on the question what to do with an investigative detention based on a report from a gravely unreliable database. The defendant in that case asserted the database there was unreliable but produced no evidence suggesting so much. In rejecting the defendant's motion to suppress, we emphasized that his theory was promising but his proof lacking: if, we said, a future defendant could bear the burden of making "a demonstration" based on evidence that the database in question "is unreliable," that "might well form a persuasive basis for a suppression motion." 495 F.3d at 1209; see also United States v. Mounts, 35 F.3d 1208, 1213 n. 4 (7th Cir. 1994) (if the defendant had shown that "a large percentage of registered automobiles... did not appear on Georgia's computer system and that this fact was common knowledge" the case might be different, but in that particular case the defendant "did not make such a showing").

Today, we find ourselves facing a good deal more directly the problem we anticipated — and on which we reserved judgment — in Cortez-Galaviz.

Of course, nothing in life is perfect. Neither does anyone expect or even want some sort of maniacally perfect, all-knowing, all-seeing HAL 9000 computer in the government's hands — a situation that would itself no doubt raise Fourth Amendment questions. Instead, the law expects and takes account of human (and computational) frailties, requiring less, far less, than perfect certainty of a traffic violation before an officer may initiate a brief investigatory stop. The Supreme Court tells us that only "reasonable suspicion" is needed for a traffic stop to comply with the Fourth Amendment's guarantee against unreasonable seizures.

That standard requires the officer to possess a "particularized and objective" basis for thinking unlawful activity is afoot. United States v. Cortez, 449 U.S. 411, 417, 101 S.Ct. 690, 66 L.Ed.2d 621 (1981). But it requires "considerably less" than a preponderance of the evidence and "obviously less" than that required for probable cause to effect an arrest. United States v. Sokolow, 490 U.S. 1, 7, 109 S.Ct. 1581, 104 L.Ed.2d 1 (1989); United States v. Johnson, 364 F.3d 1185, 1194 (10th Cir. 2004). To satisfy the reasonable suspicion standard, an officer need not "rule out the possibility of innocent conduct," or even have evidence suggesting "a fair probability" of criminal activity. Poolaw v. Marcantel, 565 F.3d 721, 736 (10th Cir.2009); see also United States v. Vercher, 358 F.3d 1257, 1261 (10th Cir.2004). Equally important, reasonable suspicion can be shown by evidence that is inherently less reliable in kind than the sort of evidence needed to establish probable cause. Alabama v. White, 496 U.S. 325, 330, 110 S.Ct. 2412, 110 L.Ed.2d 301 (1990); Johnson, 364 F.3d at 1194. Concerns with the "quality" of the information law enforcement possesses, too, sometimes may be offset by its "quantity." White, 496 U.S. at 330, 110 S.Ct. 2412. So, for example, worries about a computer database's reliability might diminish if officers also possess other facts independently suggestive of unlawful activity. Given all this, it is pretty plain that no one — and no computer — has to be perfect, or even close to perfect, to provide or help provide legally sufficient grounds for a traffic stop.

What do these general principles mean for our particular case? We know that Trooper Dean relied only on the database report to support his stop: he had no other quantum of evidence that might mitigate shortcomings of the database. We know the database undoubtedly qualifies as "objective" evidence in the sense that it is independent from the trooper's beliefs or biases. Neither is it alleged to evince anyone else's. But it is a much harder question whether the computer's "no return" report, taken in light of the dispatcher's comment, qualifies as "particularized" evidence that supplied Trooper Dean some "specific[]" reason to think that the minivan might be engaged in criminal activity. See Cortez, 449 U.S. at 418, 101 S.Ct. 690.

It's a harder question because at this point we know almost nothing about the database on which everything hinges. We don't know how it operates, why it "usually" fails to yield information about Colorado temporary tags — or, for that matter, even its name. We don't know if officials systemically decline to place information about Colorado temporary tags into the database because of their fleeting 30-day life span. We don't know if instead information from Colorado takes, say, two weeks (half the usual life span of a temporary tag) to reach the database. We don't know if Colorado temporary tags "usually don't return" thanks to some other glitch altogether. Neither do we know whether a difference exists between a tag that doesn't "return" and one that isn't "on file," another term Trooper Dean reported hearing before but also isn't exactly sure about.

To add to the confusion, we can't even be sure what it means that Colorado temporary tags usually don't return. The dispatcher's use of the term "usually" (rather than "always") suggests some Colorado temporary tags do "return" with registration information. But to say a problem usually occurs suggests it occurs more often than not. Cf. 19 Oxford English Dictionary 361 (2d ed.1991). So it seems Colorado temporary tags don't return over half the time. But still we don't know how much more than half the time the problem occurs. And we don't know why it occurs. Though threadbare, the record seems to suggest "no returns" occur as a result of at least two very different causes — unlawful conduct by the driver (forgeries and the like) and some sort of bureaucratic bumbling (whether failure to enter into the database in a timely fashion or something else altogether). Yet, how often "no returns" occur and how often they are attributable to wrongful conduct as opposed to bureaucratic mischief can make all the difference.

Consider some examples. Imagine that 200 of every 1000 Colorado temporary tags (20% of the total) are forged, and for this reason do not "return" when the database is queried. Imagine too that 301 legitimate tags similarly fail to return for innocuous reasons. That means 501 or just more than 50% of temporary tags will not return, and it means there's about a 40% chance a non-returning tag is a forgery. That chance is very likely chance enough to satisfy the reasonable suspicion standard. After all, in Terry v. Ohio, perhaps the paradigmatic investigative detention case, Officer McFadden only saw men hovering around a store window, gathering in small groups, walking away, and rejoining a couple blocks away. 392 U.S. 1, 23, 88 S.Ct. 1868, 20 L.Ed.2d 889 (1968). Equivocal though the situation surely was, the Court concluded that "[i]t would have been poor police work indeed for an officer of 30 years' experience in the detection of thievery from stores in this same neighborhood to have failed to investigate this behavior further." Id. Maybe much the same might be said for a 21st century trooper who fails to follow up on a computer lead with this sort of track record.

But change the assumptions about the underlying data and the resulting intuition changes too. Suppose only 5 of every 1000 temporary tags in Colorado (0.5%) are forged and 496 don't return for innocuous reasons. Once again, just over half the tags will not return, but less than 1% of tags that fail to return in this scenario are illegitimate (5 out of 501). And the suggestion of wrongdoing diminishes even further as the number of innocuous no returns increases. So if legitimate Colorado temporary tags are almost never placed in the database and if forged temporary tags account for an even smaller percentage of the population of temporary tags — both plausible possibilities — getting a "no return" result may tell a reasonable officer next to nothing: virtually every query would yield the same "no return" message whether the tag is legitimate or not. And it is hard to imagine how a "no return" report in those circumstances could form a "particularized" basis to suspect wrongdoing. What goes into the database will much affect the reasonableness of a search relying on it: garbage in, garbage out (GIGO).

We mention all this not to suggest district courts must become statisticians to evaluate the question of reasonable suspicion in every case. See United States v. Ludwig, 641 F.3d 1243, 1251 (10th Cir. 2011) (citing Orin Kerr, Why Courts Should Not Quantify Probable Cause 14 (George Washington University Public Law & Legal Theory Paper No. 543, Mar. 27, 2011)). Often (usually) factors relevant to the analysis "aren't subject to easy quantification." See id. But in a case where the stop is based solely on a database with apparent credibility problems, it's also difficult not to engage with these sorts of questions.

And that's the real shortcoming we see in the district court's disposition in this case. In its order rejecting Mr. Esquivel-Rios's motion to suppress, the district court treated this case as just another one where the database wasn't challenged. It recounted the dispatcher's comment that Colorado temporary tags "usually don't return" in the background section of its order but made no mention of the fact, let alone considered its significance, in the section devoted to legal analysis. Neither did the district court mention the dispatcher's comment in the legal analysis section of its order rejecting Mr. Esquivel-Rios's motion for a new trial, though that motion extensively attacked the database's reliability. Simply put, the district court failed to engage with evidence seeming to call the database into question.

Compounding the problem, the district court seemed to think the evidence about the database was much more favorable to the government than it was. In rejecting the motion to suppress, the district court found that Trooper Dean "had encountered Colorado temporary tags before, and had them come back on file when he ran them through dispatch." On its face, this fact seems to suggest the database might be more reliable than the dispatcher thought. But as the government acknowledges, the record does not support the district court's finding about Trooper Dean's testimony. As the government has admitted, "the only passage in the record that could arguably support" the district court's finding is Trooper Dean's testimony that "I've experienced [Colorado temporary tags] returning not on file." If the transcript's to be believed, Trooper Dean's testimony doesn't suggest he had experienced Colorado temporary tags returning before, but more nearly the opposite. On its face, the trooper's testimony seems to corroborate the dispatcher's comment, giving only more reason to worry that we might have a GIGO problem lurking here.

The Fourth Amendment reasonable suspicion analysis requires a careful consideration of the "totality of the circumstances." Cortez, 449 U.S. at 417, 101 S.Ct. 690. Yet in a case that hinges entirely on the reliability of a computer database, the district court overlooked one critical circumstance — the dispatcher's comment — casting doubt on its reliability. And the district court misstated another critical circumstance doing the same thing — Trooper Dean's comment about his experience with the database. All of the relevant circumstances, thus, were not considered and the district court's ruling cannot stand as issued.

The question of remedy remains. It may be within our discretion to try to take up the reasonable suspicion question for ourselves, armed now with a candid appreciation of the facts in the record. But it is no less an option to remand the case to the district court for it to undertake that task in the first instance, and in this case we think that course the more prudent. This court lacks the fact-finding powers necessary to try to answer even some of the many questions we've identified. Without some effort to address those questions, without further data about the database, any answer we might offer at this stage could itself suffer a GIGO problem. Meanwhile, the district court can permit the parties to explore further how (un)reliable the database is, make factual findings based on their presentations, and with more and better information in hand increase the chance the case will be resolved justly. Even if the district court refuses to take additional evidence on remand, a course we would not choose or advise but one admittedly within its discretion, the reasonable suspicion question is sufficiently close on the record as it currently stands, and sufficiently novel and important, that before issuing a definitive ruling this court would benefit from a district court judgment that takes account of the relevant facts previously overlooked.

One way or the other, then, we think it wise in this case to follow the more restrained remedial path of correcting the district court's error and allowing it to revisit the case in that new light — a path we and others have found valuable in the past. See, e.g., United States v. Salerno, 505 U.S. 317, 325, 112 S.Ct. 2503, 120 L.Ed.2d 255 (1992) ("we think it prudent to remand" when court "declined to consider fully" the parties arguments); Metro. Stevedore Co. v. Rambo, 515 U.S. 291, 301, 115 S.Ct. 2144, 132 L.Ed.2d 226 (1995) (remanding for circuit court to address arguments raised before it but not addressed in its ruling); United States v. Little, 18 F.3d 1499, 1506 (10th Cir.1994) (en banc) (because "the district court ... failed to fully explore the totality of the circumstances surrounding the encounter" the court remanded "for further proceedings, employing the correct legal standards"); United States v. Wright, 485 F.3d 45, 54 (1st Cir.2007) (remanding and suggesting district court consider location of crime as among circumstances for reasonable suspicion analysis).

We are not unmindful that our invitation to the district court to accept additional evidence about the database might be seen as affording the government a chance to introduce more evidence to support its cause. But additional evidence in this case might prove no less and maybe even more helpful to Mr. Esquivel-Rios: it is he, after all, who bears the burden of proving a Fourth Amendment violation and he who shoulders the consequences of an insufficiently developed record. See, e.g., Cortez-Galaviz, 495 F.3d at 1208-09, 1211 (rejecting database challenge because of insufficiently developed record). The simple fact is, no one knows where the facts may lead or who they will help. More importantly, whatever the effect on the parties, our responsibility is to ensure an important and novel legal question bound to affect many cases and parties — the interaction between the reasonable suspicion standard and the use of law enforcement databases that are both imperfect and now so prevalent — gets the full vetting it deserves so this court ultimately might be in a position to offer a judgment with the degree of confidence the question merits.

If, at the end of the day, the district court on remand finds reasonable suspicion absent and the Fourth Amendment violated, it should proceed to the question (raised by the government below but also as yet unaddressed by the district court) whether exclusion is an appropriate remedy, following the mode of analysis dictated by the Supreme Court in Davis v. United States, ___ U.S. ___, 131 S.Ct. 2419, 180 L.Ed.2d 285 (2011), and Herring v. United States, 555 U.S. 135, 129 S.Ct. 695, 172 L.Ed.2d 496 (2009). Even if the district court on remand doesn't find a Fourth Amendment violation, we believe it would still be prudent for that court to consider the remedial question. Doing so in this and other similarly close cases can help avoid the potential need for (further) remands. See United States v. Master, 614 F.3d 236, 243 (6th Cir.2010) (remanding for district court in the first instance to "re-examin[e] the facts and balanc[e] the interests as required by Herring").2

Besides challenging the lawfulness of his traffic stop, Mr. Esquivel-Rios raises three other points in this appeal, all concerning the administration of his trial. None, however, proves so close as the Fourth Amendment question.

First, Mr. Esquivel-Rios objects to the government's introduction of testimony from two fellow drug traffickers discussing his involvement in a pair of other drug deals. Mr. Esquivel-Rios summarily asserts, with almost no development or reference to authority, that this evidence wasn't relevant and was unfairly prejudicial. Even assuming without granting this is enough to preserve the issue, we do not see any basis for reversal.

Under our case law, evidence of other crimes may be admissible only if it satisfies a four-part test: (1) it must be introduced only for a proper purpose under Federal Rule of Evidence 404(b); (2) it must be relevant; (3) its probative value must not be substantially outweighed by its potential for unfair prejudice under Federal Rule of Evidence 403; and (4) the district court must, on request, instruct the jury to consider the evidence only for the purpose for which it was admitted. See United States v. Wilson, 107 F.3d 774, 782 (10th Cir.1997).

The district court held the other crimes evidence in this case satisfied each of these requirements. The methamphetamine that formed the basis for the current charges against Mr. Esquivel-Rios was found in a secret compartment of the minivan he was driving, but the minivan bore other passengers: Mr. Esquivel-Rios's girlfriend, his brother, and two children. To prevail at trial, the government had to show that the secreted drugs belonged to Mr. Esquivel-Rios specifically and that he intended to distribute them, all despite the defendant's suggestion that he had no idea the drugs were in the minivan and they really belonged to his brother. In these circumstances, the district court found the evidence of Mr. Esquivel-Rios's prior drug deals relevant and admissible for a proper purpose — to show his knowledge of the drugs' presence and his intent to distribute them. The court also found that the probative value of the past crimes evidence outweighed the prejudicial effect. And the court gave a limiting instruction, admonishing the jury that evidence of past crimes was permissible not to show propensity but only to show intent and knowledge.

We are unable to fault any of this. Federal Rule of Evidence 404(b) lists knowledge and intent as proper purposes for the admission of other crimes evidence. See id.; United States v. Mares, 441 F.3d 1152, 1156 (10th Cir.2006); United States v. Cherry, 433 F.3d 698, 701 (10th Cir. 2005). This court has said that prior crimes evidence may be relevant and admissible for purposes of showing knowledge or intent "as long as the uncharged acts are similar to the charged crime and sufficiently close in time." United States v. Zamora, 222 F.3d 756, 762 (10th Cir.2000). In this case, the uncharged acts not only occurred less than a year before the traffic stop at issue, they also occurred on the defendant's property and, much as here, they involved a large amount of drugs that suggested an intent to distribute. We do not doubt that the evidence was prejudicial "in the sense that it rebutted [Mr. Esquivel-Rios's] theory of defense," but we have explained that that alone is not usually enough to establish unfair prejudice. Mares, 441 F.3d at 1159. Neither has Mr. Esquivel-Rios identified any other form or special degree of prejudice we should consider, nor does he fault the district court's limiting instruction in any way. In similar circumstances, this court has upheld the use of other crimes evidence and we do not see any way we might faithfully reach a different result here. See id.; Cherry, 433 F.3d at 702; United States v. Ramirez, 63 F.3d 937, 943 (10th Cir.1995) ("This court has repeatedly held that the use of prior drug involvement to show plan, motive or intent in a drug trafficking offense is appropriate." (internal quotation marks omitted)).

Second, Mr. Esquivel-Rios objects to the admission at trial of expert testimony from a government agent about the significance of a book, Santa Biblia de la Santa Muerte ("Holy Bible of Saint Death"), and a tattoo on his shoulder depicting an image of Santa Muerte ("Saint Death"). The government agent testified that these icons are closely associated with drug traffickers, thought by some to "protect[]" them from detection and "guid[e]" their efforts. Mr. Esquivel-Rios argues that this evidence should have been excluded as irrelevant under Federal Rules of Evidence 402 and 403.

Whatever conceivable error might have occurred, however, was harmless. Quite apart from this tangential testimony, the evidence amassed at trial against Mr. Esquivel-Rios on the drug trafficking charge was, frankly, pretty compelling. The evidence showed that, after his traffic stop, he lied to law enforcement about the purpose of his trip, the minivan containing the drugs was registered to him, he was driving it, a large amount of drugs was involved (suggesting that it might be difficult for the driver and owner to be blind to their presence), and Mr. Esquivel-Rios later admitted to a law enforcement agent that he knew "things about drug trafficking... [and] illegal activities." Federal Rule of Criminal Procedure 52(a) precludes us from reversing on the basis of trial errors that don't affect a defendant's substantial rights. All in all, given all the evidence produced at trial we think the Santa Muerte testimony fits that description: indeed, we would find any error here harmless even beyond a reasonable doubt, though Mr. Esquivel-Rios does not ask us to do so and does not even answer the government's harmless error argument. Our holding on this score again comports with results this court has regularly reached in similar cases in the past. See, e.g., United States v. Berry, 717 F.3d 823, 832-33 (10th Cir.2013); Wilson, 107 F.3d at 785-86; United States v. Hill, 60 F.3d 672, 681 (10th Cir.1995); United States v. Barbee, 968 F.2d 1026, 1033-35 (10th Cir. 1992).

Third and finally, Mr. Esquivel-Rios contends the prosecutor engaged in reversible misconduct in his closing argument. He points to two passages in particular. In the first, the prosecutor said this:

Later, the prosecutor added this:

Mr. Esquivel-Rios complains that in the first passage the prosecutor improperly appealed to social policy rather than the facts of the case and that in the second passage the prosecutor improperly vouched for the credibility of his witnesses. Counsel, however, failed to object to the prosecutor's statements before the district court, leaving us to review only for plain error, a standard that requires Mr. Esquivel-Rios to prove four things: (1) an error, (2) that is plain, which (3) affects the defendant's substantial rights, and which (4) seriously affects the fairness, integrity, or public reputation of judicial proceedings. United States v. Fleming, 667 F.3d 1098, 1103 (10th Cir.2011). Applying this stringent test, we see no way to reverse. Assuming without granting an error lurks here somewhere, it isn't plainly evident and Mr. Esquivel-Rios's arguments fail at the latest by the second step.

Mr. Esquivel-Rios says the prosecutor's remarks in the first passage run afoul of United States v. Taylor, 514 F.3d 1092 (10th Cir.2008). There we admonished prosecutors to avoid "confus[ing] the task of the jury — as finder of fact — with the task of elected officials — as the authors of social policy.... [T]he amelioration of society's woes is far too heavy a burden for the individual defendant to bear." Id. at 1095 (internal quotation marks omitted). But it's not at all plain that's what the prosecutor was up to in this case. While the prosecutor surely rambled a good deal and it's hard to say what exactly his point was, it's at least plausible to think his purpose in the first passage was to focus the jury on the presumption of innocence and the burden of proof, matters favorable in many ways to the defense.

Mr. Esquivel-Rios says the prosecutor's remarks in the second passage violate our directions in Thornburg v. Mullin, 422 F.3d 1113 (10th Cir.2005). There we said that "[a]rgument or evidence is impermissible vouching only if the jury could reasonably believe that the prosecutor is indicating a personal belief in the witness'[s] credibility." Id. at 1132. Again, however, it is not plain that the prosecutor was doing anything of the sort in this case. The second passage lends itself equally if not more to the innocuous conclusion that the prosecutor was pleading with the jury to credit the government's witnesses based on the jury's own observations. Whatever other possibly untoward intimations might be buried in the prosecutor's remarks in either passage, they simply aren't anything close to plain.

At the end of it all, we see merit in only one of Mr. Esquivel-Rios's arguments on appeal: his first. That argument may yet prove dispositive — if the traffic stop proves to be unlawful and if suppression of evidence about the drugs found during the stop proves to be the correct remedy, the results of Mr. Esquivel-Rios's otherwise lawful trial — where evidence about the drugs was introduced against him — will be called in question. At this point, however, those contingencies remain for the district court to reconsider, informed now by a full appreciation for all the circumstances surrounding the "no return" report and their consequences for the database's reliability. The case is remanded for further proceedings consistent with this opinion.