EASTERBROOK, Chief Judge.

A state judge erroneously disqualified one of the two lawyers representing Neftaly Rodriguez, who was convicted of murder and sentenced to 29 years in prison. Three years ago we held that this disqualification violated the sixth amendment (applied to the states by the fourteenth) by depriving him of his chosen counsel, but that such an error does not automatically vitiate the conviction. Rodriguez v. Chandler, 382 F.3d 670 (7th Cir.2004) (Rodriguez I). A writ of habeas corpus should issue, we held, only if the error had an adverse effect on the defense.

After the Supreme Court denied certiorari, 543 U.S. 1156, 125 S.Ct. 1303, 161 L.Ed.2d 124 (2005), the district court first ruled that Rodriguez bore the burden of persuasion on this issue and then granted summary judgment against him after concluding that he had not adduced any reason to think that the erroneous disqualification affected the trial. Rodriguez v. Chandler, No. 02 C 2184 (N.D.Ill. June 21, 2006).

Five days later, the Supreme Court held in United States v. Gonzalez-Lopez, - U.S. -, 126 S.Ct. 2557, 165 L.Ed.2d 409 (2006), that an erroneous failure to allow a defendant’s chosen lawyer to represent him at trial violates the Constitution without regard to consequences—that it is a “structural error” that always requires a new trial if an objection has been preserved. The district court then issued a writ of habeas corpus requiring Illinois to release Rodriguez unless he is retried within 120 days. We stayed that order to give the state time for appeal.

In this court the parties limited their briefs to a single question: whether Gonzalez-Lopez governs not only cases in which defendant hires one lawyer at a time (the situation in Gonzalez-Lopez) but also the situation in which defendant has two lawyers. The state’s theory is that one “counsel of choice” is enough; Rodriguez responds that Illinois does not limit a (paying) defendant to just one lawyer and therefore cannot disqualify counsel to come down to that number.

Rodriguez has the better of this argument. After the district court in Gon *865 zalez-Lopez refused to allow Low to represent defendant at trial, he hired Dickhaus—who was, we must assume, his second choice. Rodriguez had choices 1 and 2 (Brent and Grimaldi) on hand from the start; when one was knocked out, the other represented him at trial. That’s exactly what happened for Gonzalez-Lopez. Illinois does not argue that Grimaldi was # 1 in Rodriguez’s ranking and Brent #2, and it is hard to see how a court could distinguish them. It would not have mattered in Gonzalez-Lopez if Dickhaus had been asked first and had turned Gonzalez-Lopez away because of a scheduling conflict, only to take the case later when Low could not handle the trial. The Supreme Court did not distinguish among degrees of preference. Fact is that both Gonzalez-Lopez and Rodriguez found two lawyers they trusted but were allowed to use the services of only one. The injury may have been greater for Rodriguez, who thought that having two lawyers would help (they could share the work, and two sets of ideas may be better than one) but ended up represented at trial by just one.

But does Gonzalez-Lopez apply in the first place? Which decisions supply the governing law, no less than whether the collateral attack is untimely, is a subject that a court may and often should raise on its own. See Day v. McDonough, 547 U.S. 198, 126 S.Ct. 1675, 164 L.Ed.2d 376 (2006). Rodriguez’s conviction became final years before Gonzalez-Lopez was announced, and we know from both Teague v. Lane, 489 U.S. 288, 109 S.Ct. 1060, 103 L.Ed.2d 334 (1989), and 28 U.S.C. § 2254(d)(1) that a federal court should apply the law as it was when the state’s process ended, not the law as it is when the federal judiciary turns to the subject. Wfiien Rodriguez’s conviction became final, no decision of the Supreme Court of the United States treated mistaken disqualification as a structural error. Our opinion of 2004 discussed the state of the law during the 1990s and the early part of this decade; it shows that the law was unsettled, which normally means that there cannot be federal collateral relief. There was a conflict among the circuits about how to handle this situation, and although four Justices concluded in Gonzalez-Lopez that our solution was sound, see 126 S.Ct. at 2568 (Alito, J., dissenting), five Justices took a view more favorable to Rodriguez. Resolution of such uncertainties does not apply retroactively.

We asked the parties to file supplemental memoranda concerning the retroactive application of Gonzalez-Lopez, and after considering the parties’ arguments we conclude that the case must be resolved under the law as we stated it in 2004.

True enough, we held in Rodriguez I that the doctrine of harmless error, as applied on collateral review, specifies how federal courts evaluate the consequences of state courts’ mistakes; it is not a rule concerning what state courts must do and therefore may change without being thought impermissibly retroactive. See 382 F.3d at 674. The Supreme Court regularly elaborates on doctrines of harmless error without thinking that it needs to justify their application to old cases under the Teague standard. See, e.g., Brecht v. Abrahamson, 507 U.S. 619, 113 S.Ct. 1710, 123 L.Ed.2d 353 (1993); O’Neal v. McAninch, 513 U.S. 432, 115 S.Ct. 992, 130 L.Ed.2d 947 (1995).

What the Court did in Gonzalez-Lopez, however, differs from a refinement of harmless-error analysis for the federal judiciary. It held, instead, that a primary meaning of the sixth amendment is that judges must respect defendants’ choice of lawyer unless a good reason requires disqualification. The Court made it clear that its decision with respect to both de *866 fendants’ legal entitlement and the structural-error analysis was driven by the Constitution. Gonzalez-Lopez did not articulate a rule that applies only in federal court, leaving states to devise their own approach independently. The Court announced a constitutional norm, applicable in all courts within the jurisdiction of the United States. And a new rule telling states how their own courts must behave rarely applies to prosecutions that became final before the rule’s announcement.

A new rule of constitutional law applies retroactively in collateral proceedings only if it is substantive (that is, if it places primary conduct outside the reach of the criminal sanction) or is a “ ‘watershed rul[e]’ of criminal procedure implicating the fundamental fairness and accuracy of the criminal proceeding.” Saffle v. Parks, 494 U.S. 484, 495, 110 S.Ct. 1257, 108 L.Ed.2d 415 (1990); Whorton v. Bockting, - U.S. -, 127 S.Ct. 1173, 1180, 167 L.Ed.2d 1 (2007). Gonzalez-Lopez does not fit in either category. It does not insulate any conduct from criminal punishment. It isn’t a “watershed” equivalent to Gideon v. Wainwright, 372 U.S. 335, 83 S.Ct. 792, 9 L.Ed.2d 799 (1963); Rodriguez does not deny that he enjoyed the assistance of competent counsel. And it does not protect “the fundamental fairness and accuracy of the criminal proceeding.” Quite the contrary, Gonzalez-Lopez rejected the approach we had taken in Rodriguez I precisely because, in the majority’s view, the right to be represented by counsel of choice protects an interest in autonomy rather than the accuracy of the trial. The holding of Gonzalez-Lopez is that the right’s nature requires the trial to be rerun even when there was no prejudice. A rule so stated is not applied retroactively on collateral attack, a point made earlier this year by Whorton, which held that Crawford v. Washington, 541 U.S. 36, 124 S.Ct. 1354, 158 L.Ed.2d 177 (2004), is not retroactive.

This conclusion leaves the question we posed in 2004, reflecting the law as it then was: whether Rodriguez suffered any adverse effect from the mistaken disqualification. Unfortunately, neither side’s appellate brief devotes much space to that question; counsel were too dazzled by Gonzalez-Lopez. The discussion that does appear in the briefs does not distinguish clearly between the need to show some kind of effect as an element of the constitutional wrong, and the evaluation of harmless error if the Constitution has been violated. These need to be analyzed separately, as Gonzalez-Lopez did. Our original opinion’s failure to keep these subjects distinct may have influenced the parties’ briefing choices on appeal.

Rodriguez I treated adverse effect as a modification of the “prejudice” component in Strickland v. Washington, 466 U.S. 668, 104 S.Ct. 2052, 80 L.Ed.2d 674 (1984), and thus as an element of the constitutional wrong. This implies that the petitioner bears the burden of persuasion, as the district court held, but our opinion did not address that subject and it may need further consideration. We also invite the parties to address whether the district court needed to hold an evidentiary hearing before reaching a decision about adverse effect, and whether there is any independent role for harmless-error analysis if an adverse effect should be established. Counsel also should ensure that both of the lawyers representing Rodriguez were retained (as we have been assuming), given the Supreme Court’s view that “the right to counsel of choice does not extend to defendants who require counsel to be appointed for them.” Gonzalez-Lopez, 126 S.Ct. at 2565. See also Wheat v. United States, 486 U.S. 153, 159, 108 S.Ct. 1692, 100 L.Ed.2d 140 (1988). Accordingly, we direct counsel for the state to file a brief addressing these subjects by August 2, 2007. Counsel for Rodriguez will have *867 until August 23, 2007 to respond, and a reply brief may be filed no later than September 4, 2007.