DIANE P. WOOD, Circuit Judge.
Although the underlying facts of this case relate to a tragic train accident that occurred several years ago in Bourbonnais, Illinois, the question before us in this appeal is a narrow procedural one: can this court review the district court’s decision to remand the case to the state court in which it was originally filed? We conclude that the best way to interpret the district court’s order is as one finding that it had no subject matter jurisdiction over the claims that were remanded. That being the case, this court has no appellate jurisdiction over the district court’s remand order. Furthermore, even if we are wrong and the order was simply a decision to remand all supplemental state claims to the state court, we would exercise our jurisdiction to find that there was no abuse of discretion in that decision. We therefore dismiss this appeal for lack of appellate jurisdiction.
I
On March 15, 1999, an Amtrak passenger train known as the City of New Orleans (made famous in a song written by Steve Goodman and performed by Arlo Guthrie and later Willie Nelson) collided with a semi-tractor trailer loaded with steel rebar at a railroad crossing in Bour-bonnais, Illinois. The City of New Orleans derailed as a result of the collision, resulting in significant injury and the loss of many passengers’ lives. This suit, along with more than fifty others, was filed in Illinois state court by injured passengers and the estates of those killed as a result of the collision. The plaintiffs alleged neg[830]*830ligence, strict products liability and breach of implied warranties on the part of several defendants, including General Electric, which designed and manufactured the locomotive that pulled the City of New Orleans. The additional defendants were the manufacturers, owners, and operators of the equipment involved in the accident. Two from that group are relevant to this appeal: Illinois Central, which operates the tracks on which the City of New Orleans traveled, and Birmingham Steel Company, which loaded the steel rebar onto the trailer involved in this collision from its adjacent manufacturing facility. Amtrak, though not initially named as a defendant in this suit, plays a role in the issues before us as well.
Essentially, the procedural posture of this case can be boded down to a simple pattern. Plaintiffs (A) filed suit in state court raising a number of state law claims against a set of defendants (B through n). One of those defendants, GE, removed the case to federal court, asserting as a basis for removal the argument that the federal Locomotive Inspection Act, 49 U.S.C. § 20701, et seq., “completely preempted” the state law claims. (In other words, GE claimed that any lawsuit these plaintiffs were bringing was inevitably and necessarily based on a federal question, and thus removal was available under 28 U.S.C. § 1441(a) and (b).) Later, defendant Birmingham Steel filed a third-party complaint pursuant to fed. R. Civ. P. 14 against Amtrak, which is a federal instrumentality. The district court initially concluded that GE’s “complete preemption” argument was correct, and thus that the removal was properly based on the presence of a federal question (and supplemental jurisdiction over the remaining claims). The court then dismissed all claims that the plaintiffs had asserted against GE (all of which relied on state law) on preemption grounds. Finally, turning to the other defendants, the court ascertained that none of them was arguing that federal question jurisdiction existed based on the “complete preemption” idea. In the absence of a federal question, and in the absence of the complete diversity required by 28 U.S.C. § 1332, he concluded that the case had to be remanded to state court. Believing that the order of remand was erroneous, Illinois Central filed an appeal to this court. (Plaintiffs have not filed a cross-appeal from the order dismissing their claims against GE, and thus we have no occasion to rule on the correctness of the findings supporting that order.)
II
The first question we must decide is whether we have appellate jurisdiction over Illinois Central’s challenge to the remand order. After devoting scant attention to this threshold question in its opening brief, Illinois Central urged us to find such jurisdiction in its reply brief. Plaintiffs-appellees squarely challenged appellate jurisdiction in their brief. Even if they had not, however, we would have had an independent duty to satisfy ourselves that jurisdiction existed before we could proceed to the merits of the appeal.
The reason why appellate jurisdiction is problematic, at best, comes from the language of the statute governing removals and remands. In deceptively simple words, 28 U.S.C. § 1447(d) says
An order remanding a case to the State court from which it was removed is not reviewable on appeal or otherwise, except that an order remanding a case to the State court from which it was removed pursuant to section 1443 of this title [which applies only to civil rights cases] shall be reviewable by appeal or otherwise.
[831]*831The naive reader might think that this meant no appellate consideration by appeal, by writ of mandamus, or by any other device that lawyers might serve up, but that reader would be wrong. In a line of cases that began with Thermtron Products, Inc. v. Hermansdorfer, 423 U.S. 336, 96 S.Ct. 584, 46 L.Ed.2d 542 (1976), the Supreme Court has adopted a more nuanced interpretation of the statute, and it is therefore the Thermtron approach we are bound to follow.
In Thermtron itself, the Court faced an insubordinate district court judge who had remanded cases to the state court because his docket was overcrowded. The court of appeals had reviewed the propriety of these remands by means of a writ of mandamus, and the question before the Supreme Court was whether this was possible in light of the § 1447(d) ban on review “on appeal or otherwise.” Using a close reading of the statute as a whole, the Court concluded that the § 1447(d) ban on appellate consideration of remand orders had to be linked to the reasons for remand found in § 1447(c). 423 U.S. at 346, 96 S.Ct. 584. If the district court announced that its remand order was based on one of the grounds for remand recognized in § 1447(c) (essentially, defects in removal procedure or want of subject matter jurisdiction), then review was barred. If the district court’s order fairly read indicated that it was remanding for some other reason — if it was potentially ultra vires — then some form of appellate review could proceed. Id. In Thermtron itself, after accepting jurisdiction, the Court concluded that overcrowded dockets was not a permissible reason for remand. In the later case of Quackenbush v. Allstate Insurance Co., 517 U.S. 706, 711-12, 116 S.Ct. 1712, 135 L.Ed.2d 1 (1996), the Court agreed that remand in conjunction with a decision to abstain fell outside the scope of the § 1447(c) reasons and was thus reviewable despite § 1447(d); it then concluded on the merits that the remand order was unwarranted.
The Court has made it clear, however, that the Thermtron holding was not an open-ended invitation to exercise appellate review over remand decisions. To the contrary, it has three times cautioned that the Thermtron exception to § 1447(d) is to be narrowly construed. Just over a year after Thermtron was decided the Court issued a terse, per curiam opinion reversing the Fifth Circuit’s decision to reverse and remand a case to the district court for reconsideration of a remand order that had been based upon a (possibly erroneous) finding that diversity jurisdiction was lacking. See Gravitt v. Southwestern Bell Telephone Co., 430 U.S. 723, 723-24, 97 S.Ct. 1439, 52 L.Ed.2d 1 (1977). In doing so the Court stressed the fact that the appeals courts are entirely without jurisdiction to review a remand order based on jurisdictional grounds, and that nothing in Thermtron was to the contrary. Id. Nearly a decade later in Carnegie-Mellon University v. Cohill, 484 U.S. 343, 356, 108 S.Ct. 614, 98 L.Ed.2d 720 (1988), the Court described Thermtron as “a response to a clearly impermissible remand,” and hence as something that did not forbid a district court from remanding pendent state-law claims. Finally, in Things Remembered, Inc. v. Petrarca, 516 U.S. 124, 128, 116 S.Ct. 494, 133 L.Ed.2d 461 (1995), the Court held that a district court’s order remanding a bankruptcy case to state court on the basis of untimely removal fell within the review prohibition of § 1447(d). Interestingly, Justice Kennedy (joined by Justice Ginsburg) filed a concurring opinion in Things Remembered in which he characterized Carnegie-Mellon as an opinion limiting Thermtron’s scope. He specifically distinguished between the legitimacy of re[832]*832manding a pendent or supplemental state claim to state court, on the one hand, and the possibility of judicial review of such an order, on the other. See 516 U.S. at 130, 116 S.Ct. 940. The latter issue, he pointed out, was not before the Court in Things Remembered. Id.
Before turning to the further development of Thermtron in the lower courts, it is worth pausing for a moment to consider why Congress might have included § 1447(d) in the Judicial Code. The answer is apparent: it wanted to expedite the process of choosing a forum for litigation and to avoid exactly the kind of lengthy proceeding we are having in the present case. It is a one-bite-at-the-apple scheme, with the narrow Thermtron exception for truly abusive situations, plus the Quacken-bush exception for unwarranted applications of the slippery abstention doctrines. Normally, however, a litigant is entitled to remove a case from state court (a very mechanical procedure), to have the federal court test the propriety of the removal, and then to proceed with the case in either the federal court or the state court without further ado. The only thing that is at stake is the forum that will hear a claim. This is certainly not an unimportant matter, but it is not so fundamental that a second or third layer of judges must test its correctness. And as the Court has noted in a case not involving § 1447(d), an order resolving a dispute over where litigation shall proceed is not an appealable final order, nor does it fall within the exception carved out by the collateral order doctrine for significant claims that are otherwise effectively unreviewable on appeal. See Lauro Lines s.r.l. v. Chasser, 490 U.S. 495, 498, 109 S.Ct. 1976, 104 L.Ed.2d 548 (1989). This and the Supreme Court’s receptivity to arbitral fora as alternatives to courts, its acceptance of forum selection clauses in contracts, and its frequent reminders that the state courts are fully capable of adjudicating federal claims and are entitled to respect by federal courts, all support the idea that at some point litigation over the choice of a courtroom must end. Section 1447(d) establishes that endpoint for almost all cases that are removed to federal court: it comes when the district court judge makes the call, and it therefore falls outside the normal pattern of cases in which an appeal is either immediately or ultimately available. While we are certainly obliged to apply Thermtron and Quackenbush, we must also take seriously the cautionary language in Gravitt, Camegie-Mellon and Things Remembered. Otherwise, § 1447(d) will cease to play any meaningful function. Perhaps that would be good policy, but it is a policy that only Congress can implement; as long as the statute is on the books, it is binding on the federal courts.
It is against that background that we must evaluate one possible approach that arguably would support our appellate jurisdiction here. Some courts appear to follow a rule under which appellate review is permissible if a jurisdictional remand occurs as a result of “later events” in a case. See In re Amoco Petroleum Additives Co., 964 F.2d 706, 708-09 (7th Cir.1992) (“[W]e understand Camegie-Mellon to permit review when the district court believes that removal was proper and that later developments authorize remand.”); In re Shell Oil Co., 966 F.2d 1130, 1132 (7th Cir.1992) (same). See also Poore v. American-Amicable Life Ins. Co., 218 F.3d 1287, 1291 (11th Cir.2000); Trans Penn Wax Corp. v. McCandless, 50 F.3d 217, 223 (3d Cir.1995); Bogle v. Phillips Petroleum Co., 24 F.3d 758, 761-62 (5th Cir.1994); Van Meter v. State Farm Fire & Cas. Co., 1 F.3d 445, 450 (6th Cir.1993). But see Angelides v. Baylor College of Med., 117 F.3d 833, 836 & n. 3 (5th Cir.1997) (“a remand order [833]*833based on § 1447(c) jurisdictional grounds is not reviewable even if based on a post-removal event”); Linton v. Airbus Industrie, 30 F.3d 592, 599-600 & n. 38 (5th Cir.), cert. denied, 513 U.S. 1044, 115 S.Ct. 639, 130 L.Ed.2d 545 (1994) (same); Nutter v. Monongahela Power Co., 4 F.3d 319 (4th Cir.1993) (declining to address whether appellate review exists for remand for lack of subject matter jurisdiction because of post-removal occurrences).
In our view, however, it would be a mistake to think that there was a hard- and-fast rule to that effect (and we thus have no necessary quarrel with the holdings of the cases cited in the previous paragraph). All later events are not equal. Some, such as a district court’s conclusion that an initial acceptance of a claim of “complete preemption” was erroneous, reveal that the federal court never had jurisdiction of the case to begin with. Other later events, such as a bona fide loss of diversity of citizenship or an ultimate judgment for less than the jurisdictional amount, may not have any jurisdictional significance at all. Still others affect only the supplemental jurisdiction of the court. If we are dealing with the first of these categories — a later decision that jurisdiction never existed at all over the case— there is nothing in the language of § 1447(c) that makes the timing of the district court’s jurisdictional determination significant. To the contrary, that statute makes it clear that defects other than the lack of subject matter jurisdiction must be raised within 30 days after the notice of removal, but a district court must remand a case “at any time before final judgment” if it concludes that it lacks subject matter jurisdiction. 28 U.S.C. § 1447(c). Therm-tron instructed that § 1447(c) and (d) must be read together, as a unit; if that is true, then the most logical result would be to say that any remand based on a conclusion that jurisdiction was lacking at the time of removal is covered by § 1447(c), no matter when that fact becomes apparent, and thus that appellate consideration of any such remand is barred by § 1447(d).
The present case, in our view, is best regarded as one in which the court never had any jurisdiction over the claim. Granted, the court’s order is not as clear as it might be with respect to its own jurisdiction. In the order addressing Illinois Central’s motion for reconsideration of the remand order, the judge actually said “this court likely enjoys subject matter jurisdiction over this ligitation,” but it then found that Illinois Central had waived any claim that complete preemption supported jurisdiction. We agree with our dissenting colleague that the jurisdiction of the court cannot be conferred by consent, and cannot in a certain sense be defeated by waiver. On the other hand, courts normally do not force parties to file particular defenses or claims; our reported cases are replete with examples of both claims and defenses that are waived. Illinois Central’s concession here that the only pertinent form of preemption that could be asserted was conflict preemption made it clear to the court that nothing but state law claims over which it had no jurisdiction were before it. On that understanding, we read the court’s remand order as one based on an ultimate lack of its own subject matter jurisdiction over the claims. We note that the court did not say that it had jurisdiction over the claims under the supplemental jurisdiction statute, 28 U.S.C. § 1367. But the fact that the court may have made a mistake about its jurisdiction over the remaining claims does not defeat the bar found in § 1447(d). See, e.g., Gravitt, 430 U.S. at 723, 97 S.Ct. 1439; Briscoe v. Bell, 432 U.S. 404, 413 n. 13, 97 S.Ct. 2428, 53 L.Ed.2d 439 (1977).
[834]*834But wait, the defendants argue: remember that Birmingham Steel impleaded Amtrak under Rule 14 before the case was remanded. Even though the original case of A vs. B through n did not belong in federal court, what happens when defendant C files a third-party claim against Amtrak? Can the later third-party claim save federal jurisdiction over the whole suit? It cannot, for the reasons we explain in more detail in Part III of this opinion. For present purposes, however, the only important point is that the district court did not think that Amtrak saved its jurisdiction. As we just noted, the rule of nonreviewability found in § 1447(d) means that even remands based on an erroneous belief in the lack of federal subject matter jurisdiction cannot be reviewed. Thermtron, 423 U.S. at 343, 96 S.Ct. 584. Otherwise the rule means nothing at all, because appeals will be taken and sustained in those cases where the district court made a mistake, and rejected in cases where the district court was correct. Even if the district court was wrong that it lacked jurisdiction over the claims that it remanded, the remand would nevertheless be jurisdictional. The only consequence of a mistake is that the case would proceed in state court even though it could have stayed in federal court. The drafters of § 1447(d) must have known that a few such mistakes would occur, but they were willing to accept them. In this case, if Amtrak found itself before a state court as a third-party defendant, it would have the right to remove its own case under federal question jurisdiction, because it is a federal entity in which the United States owns more than 50% of the outstanding stock. See 28 U.S.C. §§ 1331, 1349. It is true that this means it might take Amtrak two steps to remain in federal court instead of one, but it also means that Congress’s intent in enacting § 1349 will not be thwarted.
Since the district court’s order does not indicate that it believed that either supplemental jurisdiction or the presence of Amtrak saved its jurisdiction, we find that the order of remand was indeed a jurisdictional one. That means it fell within the scope of § 1447(c) and thus that appellate review of the order is unavailable because of § 1447(d). Even if it is proper to rely on a temporal element to the ban on reviewability of jurisdictional remands in some circumstances, this is not one of them. To do otherwise would be to go significantly beyond the holdings of the Supreme Court and the language and purpose of § 1447(c) and (d). That is something we are not willing to do.
Ill
Because reasonable people might disagree over the best reading of the district court’s remand order, we think it prudent to address the alternative possibility that the court believed that it did have jurisdiction over the case after GE was dismissed, but that a remand was appropriate for some other reason (waiver by the defendants of the “complete preemption” theory of defense, or discretionary exercise of the power to decline supplemental jurisdiction). If that reading is the correct one, then our appellate jurisdiction would be secure and we would have to consider the merits of the district court’s decision.
The starting point for that determination would be, once again, a jurisdictional inquiry. What, if anything, supported the district court’s jurisdiction (as opposed to our own)? Diversity jurisdiction never existed, because citizens of Illinois appeared on both sides of the case. That means that jurisdiction over the case as a whole had to be based upon a federal question. GE claimed below and Illinois Central [835]*835urges on appeal that the doctrine of “complete preemption” formed the basis of federal question jurisdiction in this case. But we simply cannot agree with the district court that “complete preemption” supported the claims against either GE or the defendants who were left after GE was dismissed. That means, of course, that there was nothing in the case to which supplemental jurisdiction could attach. Our conclusion has nothing to do with the fact that the lawyers for the non-GE defendants appear to have disclaimed any reliance on complete preemption. This is a question on which appellate review must be de novo, wholly apart from the district court’s decision, party waivers, or other stipulations. See Rogers v. Tyson Foods, Inc., 308 F.3d 785, 787 (7th Cir.2002).
Federal law has not swept away all state law in this field, thereby making any claim the plaintiffs were trying to assert against GE necessarily federal. The discussions in our recent decisions in Tyson Foods, supra, and Vorhees v. Naper Aero Club, Inc., 272 F.3d 398 (7th Cir.2001), and the Supreme Court’s latest word on the question in Sprietsma v. Mercury Marine, — U.S. -, 123 S.Ct. 518, 526-30, 154 L.Ed.2d 466 (2002) (analyzing Federal Boat Safety Act’s preemptive effect), underscore how narrow the ill-named “complete preemption” doctrine is. As we stressed in Tyson Foods and Vorhees and as we stress again here, this does not mean that the federal statute in question— here the Locomotive Inspection Act — does not have preemptive force. To the contrary, ordinary conflict preemption may well exist with respect to some or all of the claims. But “conflict” preemption, as we noted in Vorhees, “is merely a defense to the merits of a claim. As such, according to the well-pleaded complaint rule, it does not provide a basis for federal question jurisdiction.” 272 F.3d at 403. We see nothing in the Locomotive Inspection Act to indicate that Congress “clearly intended completely to replace state law with federal law and create a federal forum” at the same time. Id. That is what would be necessary to create “complete” or “field” preemption, as opposed to the more routine conflict preemption. Furthermore, the district court did not inquire into the question whether “the federal claim includes the same ingredients as the state claim and provides some recovery,” id., even though that is a prerequisite to finding “complete” preemption. See Rogers, 308 F.3d at 788 (emphasizing the need to find a private right of action under the federal law in question). Both with respect to GE’s assertion of complete preemption, where the district court erred, and with respect to the other defendants, as to whom the district court properly recognized that only a question of conflict preemption within the competence of the state courts was presented, we believe that the district court itself lacked jurisdiction and thus that remand was required under § 1447(c). (Naturally, the existence of a federal defense to a state claim is not enough to support federal jurisdiction. See, e.g., Louisville & Nashville R. Co. v. Mottley, 211 U.S. 149, 152, 29 S.Ct. 42, 53 L.Ed. 126 (1908).) Without the “complete preemption” theory to support the claims against the other defendants (and some certainly had nothing to do with locomotive manufacture or safety), there simply was no federal question.
Even if federal jurisdiction was lacking at the outset, the question arises again whether jurisdiction over the case as a whole was somehow created when Birmingham Steel filed its third-party complaint against Amtrak. It is clear that an original, stand-alone lawsuit against Amtrak would fall within the original subject matter jurisdiction of the federal court, [836]*836pursuant to 28 U.S.C. §§ 1831 and 1349. Nevertheless, we must consider whether the existence of a third-party complaint affects the court’s subject matter jurisdiction over the original action. We conclude that it does not. The Supreme Court considered a closely related problem in Holmes Group, Inc. v. Vornado Air Circulation Systems, Inc., 535 U.S. 826, 122 S.Ct. 1889, 153 L.Ed.2d 13 (2002), where the question was whether the Court of Appeals for the Federal Circuit had appellate jurisdiction over a case in which the complaint did not allege a claim under the federal patent law, but a compulsory counterclaim filed pursuant to Rule 13(a) of the Federal Rules of Civil Procedure did. Because the Federal Circuit’s appellate jurisdiction extends only to claims that “arise under” the patent laws, the Court had to decide whether the counterclaim was enough to support that jurisdiction. It answered no. 122 S.Ct. at 1893. Most of its discussion is a general consideration of the well-pleaded complaint rule, and thus is just as applicable to reliance on pleadings under Rule 14 as it was to pleadings under Rule 13. If the plaintiffs well-pleaded complaint is not based on a federal question, there is no § 1331 jurisdiction. Id. at 1893-94. To allow even a compulsory counterclaim (which arguably presented a stronger case for jurisdiction than a permissive counterclaim or a permissive third-party action like the one before us) to govern jurisdiction would potentially defeat the plaintiffs choice of forum and “radically expand the class of removable cases.” Id. at 1894.
The third-party claim might, however, permit the court to exercise supplemental jurisdiction over the original case, if everything arises out of the same constitutional case or controversy. But it is not clear in the present case (because the district court never considered the question) whether all the claims of the plaintiffs arose out of the same transaction or occurrence as the Birmingham Steel third-party complaint against Amtrak. Without a better record, we cannot venture a guess on that point. This kind of constitutional “case or controversy” link is a prerequisite to the assertion of supplemental jurisdiction under 28 U.S.C. § 1367. See Board of Trustees, Sheet Metal Workers’ Nat’l Pension Fund v. Elite Erectors, Inc., 212 F.3d 1031, 1037 (7th Cir.2000). That statute authorizes supplemental jurisdiction only over claims “that are so related to claims in the action within such original jurisdiction that they form part of the same case or controversy under Article III of the United States Constitution.” 28 U.S.C. § 1367(a). The statute goes on to permit supplemental jurisdiction over claims involving the join-der or intervention of additional parties, and so the fact that very few of the parties to the original action were involved in Birmingham Steel’s third-party complaint is not necessarily a bar to the exercise of supplemental jurisdiction.
Once again, however, we will assume for the sake of argument that the third-party claim against Amtrak is part of the same Article III case as the entire rest of the action. We cannot find that the district court was compelled to exercise supplemental jurisdiction over the remainder of the case, once Amtrak was before it in one capacity. The normal rule is that district courts have discretion to retain or remand supplemental claims. Groce v. Eli Lilly & Co., 193 F.3d 496, 500-01 (7th Cir.1999). Here, Judge Holderman appears to have been proceeding on the assumption that he (probably) had jurisdiction over the original GE case. One possible interpretation of his remand order is that he was deciding what to do with the supplemental claims that remained after the GE case was resolved (although he used the language of waiver rather than this terminol[837]*837ogy). No matter what words he used, however, it is clear enough from his order that he thought the remainder of this case belonged in the state court, and thus he exercised whatever discretion he may have had and chose to remand those claims. If someone wants to litigate Amtrak’s derivative liability for this most unfortunate accident, he or she can do so in a separate federal court action. In fact, now that even more eases have been filed, it is possible that there will be an eventual consolidation of cases in the Northern District of Illinois. As of now, however, there is no rule that permits federal courts to reach out and prevent state courts from proceeding with claims under state law that do not fall within the jurisdiction of the federal courts. The Anti-Injunction Act, 28 U.S.C. § 2288, has no exception for this situation, even though this type of consolidation might be efficient. Furthermore, we have just been reminded by the Supreme Court that a desire to protect federal jurisdiction is not enough to support the removal of essentially state law claims. See Syngenta Crop Prot., Inc. v. Henson, — U.S. -, 123 S.Ct. 366, 154 L.Ed.2d 368 (2002).
IV
For these reasons, we dismiss this appeal for want of appellate jurisdiction. In the alternative, we conclude that the district court’s remand was proper because of a lack of its own subject matter jurisdiction over the case, or (resolving all issues as generously as possible in Illinois Central’s favor), there was no abuse of discretion in the decision to relinquish these state law claims among non-diverse parties.
Appeal Dismissed.