The Constitution entrusts state legislatures with the primary responsibility for drawing congressional districts, and legislative redistricting is an inescapably political enterprise. Claims that a map is unconstitutional because it was drawn to achieve a partisan end are not justiciable in federal court. By contrast, if a legislature gives race a predominant role in redistricting decisions, the resulting map is subjected to strict scrutiny and may be held unconstitutional. These doctrinal lines collide when race and partisan preference are highly correlated. This Court has endorsed two related propositions when navigating this tension. First, a party challenging a map’s constitutionality must disentangle race and politics to show that race was the legislature’s “predominant” motivating factor. Miller v. Johnson, 515 U. S. 900, 916. Second, the Court starts with a presumption that the legislature acted in good faith. To disentangle race from other permissible considerations, plaintiffs may employ some combination of direct and circumstantial evidence. Cooper v. Harris, 581 U. S. 285, 291. Where race and politics are highly correlated, a map that has been gerrymandered to achieve a partisan end can look very similar to a racially gerrymandered map. Thus, in Easley v. Cromartie, 532 U. S. 234, the Court held that the plaintiffs failed to meet the high bar for a racial-gerrymandering claim when they failed to produce an alternative map showing that a rational legislature sincerely driven by its professed partisan goals would have drawn a different map with greater racial balance. Id., at 258. Without an alternative map, the Court also found it difficult for plaintiffs to defeat the starting presumption that the legislature acted in good faith.

   Following the 2020 Census, South Carolina was tasked with redrawing its congressional district maps because of population shifts in two of its seven districts—Districts 1 and 6. The State Senate subcommittee responsible for drawing the new map issued a statement explaining that the process would be guided by traditional districting principles along with the goal of creating a stronger Republican tilt in District 1. To draw the new maps, the Senate turned to Will Roberts, a nonpartisan staffer with experience in drawing reapportionment plans. Roberts’ plan (the Enacted Plan) achieved the legislature’s political goal by increasing District 1’s projected Republican vote share by 1.36% to 54.39%. The plan also raised the black voting-age population (BVAP) from 16.56% to 16.72%. The legislature adopted the plan, and the Governor signed it into law.

   The National Association for the Advancement of Colored People and District 1 voter Taiwan Scott (Challengers), challenged the plan, alleging that it resulted in racial gerrymanders in certain districts and in the dilution of the electoral power of the State’s black voters. The three-judge District Court held that the State drew District 1 with a 17% BVAP target in mind in violation of the Equal Protection Clause and that this putative use of race to draw District 1 unlawfully diluted the black vote.

Held:

  1. The District Court’s finding that race predominated in the design of District I in the Enacted Plan was clearly erroneous. Pp. 12–35.

   (a) Because the State’s principal legal argument—that the District Court did not properly disentangle race from politics—is an attack on the factual basis of the District Court’s findings, this case can be disposed on clear-error grounds. The District Court clearly erred because the Challengers did not satisfy the demanding burden of showing that the “legislature subordinated traditional race-neutral districting principles . . . to racial considerations.” Miller, 515 U. S., at 916. The Challengers provided no direct evidence of a racial gerrymander, and their circumstantial evidence is very weak. Instead the Challengers relied on deeply flawed expert reports. And the Challengers did not offer a single alternative map to show that the legislature’s partisan goal could be achieved while raising the BVAP in District 1. Pp. 12–13.

   (b) The District Court’s factual findings in this case are reviewed for clear error. Because the racial predominance test has a very substantial legal component that must take account of the Court’s prior relevant decisions, special care must be exercised in reviewing the relevant findings of fact. Pp. 13–14.

   (c) The District Court’s heavy reliance on four pieces of evidence was seriously misguided in light of the appropriate legal standard and repeated instructions that a court in a case such as this must rule out the possibility that politics drove the districting process. None of the facts on which the District Court relied to infer a racial motive is sufficient to support an inference that can overcome the presumption of legislative good faith. First, the District Court concluded that the legislature deliberately sought to maintain a particular BVAP because the maps that produced the sought-after partisan goal all had roughly the same BVAP. But the mere fact that District 1’s BVAP remained around 17%, despite all the changes made during the redistricting process, proves very little. The tight correlation between the legislature’s partisan aim and District 1’s BVAP is substantiated by the District Court’s own findings. The Challengers could not point to a single map in the record that would satisfy the legislature’s political aim with a BVAP above 17%. The District Court disregarded the presumption of legislative good faith by drawing an inference that the State acted in bad faith based on the racial consequences of a political gerrymander in a jurisdiction in which race and partisan preference are very closely related. Second, the District Court inferred a racial motive from the fact that the Enacted Plan moved more voters out of District 1 than were needed to comply with the one person, one vote rule, and that the Enacted Plan split a few counties. But the high priority that the legislature gave to its partisan aim can explain these decisions. Third, the District Court clearly erred when it concluded that the legislature’s real aim was racial based on the movement of certain predominantly black Charleston precincts from District 1 to District 6. Again, the legislature’s partisan goal can easily explain this decision. Fourth, the District Court placed excessive weight on the fact that several legislative staffers admitted to viewing racial data at some point during the redistricting process. The District Court cited no evidence that could not also support the inference that politics drove the mapmaking process and provided no explanation why a mapmaker who wanted to produce a version of District 1 that would be safely Republican would use data about voters’ race rather than their political preferences. Pp. 14–19.

   (d) The four expert reports relied upon by the Challengers are flawed because they ignored traditional districting criteria such as geographical constraints and the legislature’s partisan interests. Allen v. Milligan, 599 U. S. 1, 34. The report of Dr. Kosuke Imai made no effort to disentangle race from politics. It also failed to consider “core district retention,” a term referring to “the proportion of districts that remain when a State transitions from one districting plan to another.” Id., at 21. The report of Dr. Jordan Ragusa did attempt to disentangle race from politics, but its analysis has two serious defects. First, each of his three models failed to control for contiguity or compactness. Second, he used an inferior method of measuring a precinct’s partisan leanings by counting absolute votes rather than a party’s relative share of the vote. The report of Dr. Baodong Liu purported to show that race rather than politics explains District 1’s design, but Dr. Liu’s methodology was plainly flawed. Like Dr. Ragusa, Dr. Liu failed to account for contiguity and compactness. And while this defect alone is sufficient to preclude reliance, Dr. Liu also used inferior data to measure a district’s partisan tilt—i.e., data from the 2018 off-cycle gubernatorial primaries. Finally, the report of Dr. Moon Duchin, like that of Dr. Imai, did not account for partisanship or core retention and was based on an assessment of the map as a whole rather than District 1 in particular. Thus, her report has no probative force with respect to the Challengers’ racial-gerrymandering claim regarding District 1’s boundaries. Pp. 19–29.

   (e) The District Court also critically erred by failing to draw an adverse inference against the Challengers for not providing an adequate alternate map. By showing that a rational legislature, driven only by its professed mapmaking criteria, could have produced a different map with “greater racial balance,” Cromartie, 532 U. S., at 258, an alternative map can perform the critical task of distinguishing between racial and political motivations when race and partisanship are closely entwined. Moreover, an alternative map is easy to produce. The District Court mistakenly held that an alternative map is relevant only for the purpose of showing that a remedy is plausible. A plaintiff’s failure to submit an alternative map should be interpreted by courts as an implicit concession that the plaintiff cannot draw a map that undermines the legislature’s defense. Pp. 30–31.

  2. Because the same findings of fact and reasoning that guided the court’s racial-gerrymandering analysis also guided the analysis of the Challengers’ independent vote-dilution claim, that conclusion also cannot stand. The District Court also erred in conflating the two claims. A plaintiff pressing a vote-dilution claim cannot prevail simply by showing that race played a predominant role in the districting process, but rather must show that the State “enacted a particular voting scheme as a purposeful device to minimize or cancel out the voting potential of racial or ethnic minorities.” Miller, 515 U. S., at 911. In other words, the plaintiff must show that the State’s districting plan “has the purpose and effect” of diluting the minority vote. Shaw v. Reno, 509 U. S. 630, 649. In light of these two errors in the District Court’s analysis, a remand is appropriate. Pp. 34–35.

Reversed in part and remanded in part.

 Alito, J., delivered the opinion of the Court, in which Roberts, C. J., and Gorsuch, Kavanaugh, and Barrett, JJ., joined, and in which Thomas, J., joined as to all but Part III–C. Thomas, J., filed an opinion concurring in part. Kagan, J., filed a dissenting opinion, in which Sotomayor and Jackson, JJ., joined.

The dispute here involves a conflict between two contracts executed by petitioner Coinbase, Inc., operator of a cryptocurrency exchange platform, and respondents, who use Coinbase. The first contract—the Coinbase User Agreement that respondents agreed to when they created their accounts—contains an arbitration provision with a delegation clause. Per this provision, an arbitrator must decide all disputes under the contract, including whether a given disagreement is arbitrable. The second contract—the Official Rules for a promotional sweepstakes respondents entered—contains a forum selection clause providing that California courts “shall have sole jurisdiction of any controversies regarding the [sweepstakes] promotion.” Respondents ultimately filed a class action in the U. S. District Court for the Northern District of California, alleging that the sweepstakes violated various California laws. Coinbase moved to compel arbitration based on the User Agreement’s delegation clause. The District Court determined that the Official Rules’ forum selection clause controlled the parties’ dispute and accordingly denied the motion. The Ninth Circuit affirmed.

Held: Where parties have agreed to two contracts—one sending arbitrability disputes to arbitration, and the other either explicitly or implicitly sending arbitrability disputes to the courts—a court must decide which contract governs. Pp. 4–9.

  (a) The Federal Arbitration Act “reflects the fundamental principle that arbitration is a matter of contract.” Rent-A-Center, West, Inc. v. Jackson, 561 U. S. 63, 67. Given that arbitration agreements are simply contracts, the first question in any arbitration dispute must be: What have these parties agreed to? Parties can form multiple levels of agreements concerning arbitration, and thus can have different kinds of disputes. At a basic level, parties can agree to send the merits of a dispute to an arbitrator. The merits of a dispute is a first-order disagreement. The parties may also have a second-order dispute—“whether they agreed to arbitrate the merits”—as well as a third-order dispute—“who should have the primary power to decide the second matter.” First Options of Chicago, Inc. v. Kaplan, 514 U. S. 938, 942. Pp. 4–5.

  (b) This case involves a fourth kind of dispute: What happens if parties have multiple agreements that evidence a conflict over the answer to the third-order question of who decides arbitrability? That question can be answered as to these parties only by determining which contract applies. Homing in on the conflict between the delegation clause in the first contract and the forum selection clause in the second, the question becomes whether the parties agreed to send the given dispute to arbitration. And that question must be answered by a court.

  Coinbase asks the Court to revisit the Ninth Circuit’s bottom-line conclusion below, but its arguments are unpersuasive. First, Coinbase argues that the Ninth Circuit should have applied the so-called severability principle—under which “an arbitration [or delegation] provision is severable from the remainder of the contract,” Buckeye Check Cashing, Inc. v. Cardegna, 546 U. S. 440, 445–446—and considered only arguments specific to the User Agreement’s delegation provision. But the severability rule does not require that a party challenge only the arbitration or delegation provision. Rather, where a challenge applies “equally” to the whole contract and to an arbitration or delegation provision, a court must address that challenge. Rent-A-Center, 561 U. S., at 71.

  Coinbase next contends that, as a matter of California state law, the Ninth Circuit erroneously held that the Official Rules’ forum selection clause superseded the User Agreement’s delegation provision. That issue is outside the scope of the question presented, and the Court does not address it.

  Finally, the Court does not believe its ruling here will invite chaos by facilitating challenges to delegation clauses. Regardless, where the parties have agreed to two contracts, a court must decide which contract governs. To hold otherwise would be to impermissibly elevate a delegation provision over other forms of contract. See ibid. Pp. 5–8.

55 F. 4th 1227, affirmed.

 Jackson, J., delivered the opinion for a unanimous Court. Gorsuch, J., filed a concurring opinion.

These cases concern the application of the Armed Career Criminal Act to state drug convictions that occurred before recent technical amendments to the federal drug schedules. ACCA imposes a 15-year mandatory minimum sentence on defendants who are convicted for the illegal possession of a firearm and who have a criminal history thought to demonstrate a propensity for violence. As relevant here, a defendant with “three previous convictions” for “a serious drug offense” qualifies for ACCA’s enhanced sentencing. 18 U. S. C. §924(e)(1). For a state crime to qualify as a “serious drug offense,” it must carry a maximum sentence of at least 10 years’ imprisonment, and it must “involv[e] . . . a controlled substance . . . as defined in section 102 of the Controlled Substances Act.” §§924(e)(1), (2)(A)(ii).

   Under the categorical approach, a state drug offense counts as an ACCA predicate only if the State’s definition of the drug in question “matche[s]” the definition under federal law. Shular v. United States, 589 U. S. 154, 158. The question presented is whether a state crime constitutes a “serious drug offense” if it involved a drug that was on the federal schedules when the defendant possessed or trafficked in it but was later removed.

   Petitioners Justin Rashaad Brown and Eugene Jackson were separately convicted of the federal crime of possession of a firearm by a convicted felon in violation of §922(g)(1). In both cases, an ACCA enhancement was recommended based on prior state felony drug convictions. And both defendants argued that their prior convictions did not qualify as “serious drug offense[ s].”

   Brown’s presentence report identified several Pennsylvania drug convictions, including four convictions for possessing marijuana with intent to distribute. At the time of Brown’s marijuana convictions, the federal and Pennsylvania law definitions of marijuana matched. But while Brown’s federal §922(g)(1) charge was pending, Congress modified the federal definition of marijuana. Because the federal and state definitions did not fully match when Brown was sentenced, Brown argued that his marijuana convictions no longer qualified as “serious drug offense[s]” for purposes of the ACCA sentencing enhancement.

   Jackson’s presentence report identified several prior Florida convictions, including convictions in 1998 and 2004 for possession and distribution of cocaine. In 2015, the Federal Government amended the federal definition of cocaine, so the federal and Florida definitions no longer matched when Jackson committed his §922(g)(1) offense. Like Brown, Jackson argued that these prior convictions no longer qualified as “serious drug offense[s].” In both cases, the District Courts disagreed and sentenced petitioners to enhanced sentences, and the respective appellate courts ultimately affirmed.

Held: A state drug conviction counts as an ACCA predicate if it involved a drug on the federal schedules at the time of that conviction. Pp. 4–19.

  (a) The parties propose three different answers to the question whether the federal and state definitions of a drug must match when the state crime is committed or at some later point in time. The Government argues that a prior state drug conviction qualifies if the federal and state definitions of the relevant drug matched when the defendant committed the state crime. Jackson argues instead that the definitions must match when the defendant violates the federal felon-in-possession statute. Finally, Brown contends that the definitions must match when the defendant is sentenced for the federal felon-in-possession offense. Pp. 4–7.

  (b) Precedent and statutory context support the Government’s interpretation. ACCA gauges what a defendant’s “history of criminal activity” says about his or her “culpability and dangerousness.” McNeill v. United States, 563 U. S. 816, 823. In previous cases, the Court has held that ACCA requires sentencing courts to examine the law as it was when the defendant violated it. This “backward-looking” approach, id., at 820, supports the Government’s interpretation. And the plain language of the statute points to the same conclusion. Section 924(e)(2)(A)(i), which immediately precedes the provision at issue, defines a “serious drug offense” to include, among other things, “offense[s] under the Controlled Substances Act.” A later change in a federal drug schedule does not change the fact that an offense “under the [CSA]” is a “serious drug offense.” The Government’s interpretation would treat state offenses “involving . . . a controlled substance (as defined in [the CSA])” like those federal offenses “under the [CSA].” Petitioners’ interpretations, by contrast, would treat those federal and state offenses differently, i.e., the federal offense would remain an ACCA predicate, but the state offense would not. Pp. 7–9.

  (c) The Government’s interpretation also best fulfills ACCA’s statutory objectives. In Congress’s view, defendants who have repeatedly committed ACCA predicate offenses are “especially likely to inflict grave harm when in possession of a firearm,” so ACCA imposes a higher punishment when they do so. Wooden v. United States, 595 U. S. 360, 375. Because a defendant’s “history of criminal activity” does not “cease to exist” merely because the crime was later redefined, McNeill, 563 U. S., at 823, it makes sense to ask whether a prior offense met ACCA’s definition of seriousness at the time it was committed. Brown’s and Jackson’s contrary arguments misunderstand the theory on which ACCA is based. A prior drug conviction for an offense punishable by 10 years’ imprisonment augurs a risk of future dangerousness even if the drug is no longer considered dangerous. Indeed, in McNeill, the Court found “absurd” petitioner’s argument that a later reduction in the maximum sentence for his offense reflected a legislative judgment that his prior offense was less serious than previously thought. Id., at 822. The “subsequent chang[e] in state law” did not “erase [the] earlier conviction.” Id., at 823. And it was the fact of that earlier conviction—not the legislature’s subsequent judgment—that ACCA was concerned with, because that fact “demonstrate[d]” the defendant’s “culpability and dangerousness.” Ibid. Pp. 9–12.

  (d) Petitioners various other arguments are unpersuasive. Pp. 12–19.

   (1) Relying on the so-called reference canon, Jackson claims that ACCA “incorporates [the] schedules . . . ‘as [they] exis[t] whenever a question under [ACCA] arises.’ ” Brief of Petitioner Jackson 32. The reference canon provides that a statutory reference to a “general subject” incorporates “the law on that subject as it exists whenever a question under the statute arises.” Jam v. International Finance Corp., 586 U. S. 199, 209 (emphasis added). But a reference “to another statute by specific title or section number”—such as ACCA’s reference to 21 U. S. C. §802—“in effect cuts and pastes the referenced statute as it existed when the referring statute was enacted.” Ibid. Even assuming that there may be contexts in which references to specific statutory provisions may be considered general, it is hard to see the phrase “as defined in section 102 of the Controlled Substances Act” as anything but a specific reference. Jackson’s alternative argument—that his reading is required by the “settled legal principle” that “the law that sets the penalty for a federal crime is the law in place when the crime was committed”—simply begs the question what §924(e)(2)(a)(ii) means. Pp. 12–14.

   (2) Brown suggests that present-tense language in ACCA’s definition of a “serious drug crime”—language such as “involving” and “as defined in”—indicates a present-day focus requiring courts to look to the drug schedules in effect at the time of federal sentencing. The Court rejected that approach in McNeill, holding that ACCA requires a historical inquiry into the state law at the time of that prior offense and that “[u]se of the present tense . . . did not suggest otherwise.” 563 U. S., at 820. Brown also claims that his reading is required by United States v. Schooner Peggy, 1 Cranch 103, 110, which says that when the law changes while a case is in progress, the case must be decided under the new law. But §924(e)(2)(A)(ii) has not changed at any point in the litigation. Pp. 15–17.

   (3) Petitioners’ additional arguments do not persuade. Petitioners assert that this Court’s interpretation is underinclusive because it would preclude ACCA enhancements for state offenses involving drugs added to the federal lists only after the state crimes were committed. But none of the parties’ interpretations captures all cases involving career criminals. Petitioners next suggest that the Government’s interpretation is absurd because it would exclude all state drug convictions before the CSA’s enactment in 1970. But there are reasons Congress might have chosen not to court either federal or state drug convictions that occurred before 1970. Petitioners also argue that the Government’s interpretation would unduly burden courts and defendants by requiring them to undertake the laborious task of digging up old federal drug schedules and comparing those to the state laws the defendants violated, but petitioners overstate the difficulty of this task. Finally, petitioners contend that the rule of lenity counsels in favor of their interpretations. But lenity applies only if a statute remains grievously ambiguous, and here context, precedent, and statutory design adequately show “ ‘what Congress intended.’ ” United States v. Castleman, 572 U. S. 157, 173. Pp 17–19.

No. 22–6389, 47 F. 4th 147, and No. 22–6640, 55 F. 4th 846, affirmed.

 Alito, J., delivered the opinion of the Court, in which Roberts, C. J., and Thomas, Sotomayor, Kavanaugh, and Barrett, JJ., joined. Jackson, J., filed a dissenting opinion, in which Kagan, J., joined, and in which Gorsuch, J., joined as to Parts I, II, and III.

Notes
¹ Together with No. 22–6640, Jackson v. United States, on certiorari to the United States Court of Appeals for the Eleventh Circuit.