Todd A. Brecht was serving time in Georgia for felony theft when his sister Molly Hartman paid the restitution for Brecht’s crime and took temporary custody of him while he awaited an opening in a halfway house. This created a sticky situation because Molly’s husband Roger was the district attorney for Buffalo County, Wisconsin. There cannot be many public prosecutors who want to have a felon about the house. The awkwardness of putting up the black sheep of the family was augmented by the fact that Brecht, a heavy drinker, is also homosexual, which Roger disapproved. The Hartmans explained to Brecht, when they picked him up at the prison gate on October 10, 1985, that there was to be no drinking or homosexual activity so long as he remained at their home. They reiterated that theme at least twice more before October 17, when Brecht shot and killed Roger Hartman. Wisconsin convicted Brecht of murder and sentenced him to life imprisonment. A district court has ordered him released. 759 F.Supp. 500 (W.D.Wis.1991). Whether the writ of habe-as corpus should issue depends on the standard of harmless error employed on collateral review.
I
While Roger Hartman was at work on October 17 and Molly Hartman was shopping, Brecht raided the Hartmans’ liquor cabinet for beer and brandy, and their gun cabinet for a rifle. Out in the back yard, Brecht opened fire on some cans. When Roger returned, Brecht shot him in the back and fled in Molly’s car. Roger crawled up the hill to a neighbor’s house to summon help (the phone on the ground floor of the Hartman home was inoperable, Brecht having taken a receiver off the hook upstairs). Help came, but the injury was too serious; Roger died on November 11. Roger named his assassin to the neighbor and an officer who arrived with the medics. Brecht’s initial flight ended after he drove the car into a ditch. When a police officer stopped to inquire, Brecht asserted that Molly Hartman knew about the wreck and had called a tow truck. That lie was coupled with silence about the shooting. Brecht hitched a ride to Winona, Minnesota, where the police nabbed him in a shopping center. Brecht first lied about his identity and when that did not work asked to talk with “somebody that would understand me.” He said nothing else.
Come the trial, Brecht admitted shooting Roger Hartman but contended that the rifle discharged by accident. According to Brecht, Roger’s return caught him by surprise; he did not want Roger to see him with a gun and raced to take the rifle upstairs before Roger entered the house. He tripped as he was heading toward the stairs, Brecht testified, and the gun went off as he fell. He tried to find Roger to offer aid, Brecht maintained, but Roger had left the house. Only after seeing Roger at the neighbor’s door did Brecht drive away in panic, according to his tale.
It is unlikely, but not impossible, that someone would run with his finger on the trigger of a cocked bolt-action rifle. The *1365state thought the path of the bullet disposi-tive. The coroner testified that the bullet entered Roger Hartman’s upper body on a horizontal to slightly downward trajectory, inconsistent with a claim that the weapon discharged as its bearer was falling. The rifle was found on the patio near the pool (where Roger was shot), not inside near the stairs; a live round was jammed in the chamber (from which a jury could have inferred that Brecht tried to fire another shot). In further response to the defense of accident, the prosecutor pointed to (a) Brecht’s failure to summon aid for Roger; (b) his flight; (c) his lies to the police; (d) his failure to say to anyone, at any time before trial, that the shooting was accidental; and (e) his homosexuality. The first four were offered to impeach Brecht’s testimony, and the fifth to establish a motive — hostility to Roger, reciprocating Roger’s hostility to Brecht’s sexual preference and the ban on its practice in the Hart-mans’ home. The prosecutor also obtained a ruling from the trial court that if Brecht called his probation officer in Wisconsin to testify that Brecht lacks a violent disposition, the state could offer into evidence Brecht’s conviction for passing worthless checks. Brecht elected not to call the probation officer.
The Court of Appeals of Wisconsin reversed Brecht’s conviction. State v. Brecht, 138 Wis.2d 158, 405 N.W.2d 718 (1987). It gave three reasons: first, that the prosecutor’s use of Brecht’s silence violated both the self-incrimination clauses of the state and federal Constitutions and the due process clause of the federal Constitution; second, that the references to Brecht’s homosexuality were unduly prejudicial because the prosecutor was not able to show that Roger Hartman communicated antipathy explicitly to Brecht; third, that the court’s ruling on the motion in limine was wrong because worthless checks do not establish a propensity for violence. The court concluded that these errors were not harmless.
The Supreme Court of Wisconsin reversed in turn, unanimously reinstating the conviction. 143 Wis.2d 297, 421 N.W.2d 96 (1988). Relying on State v. Sorenson, 143 Wis.2d 226, 421 N.W.2d 77 (1988), the court rebuffed all claims under the self-incrimination clause of the state constitution, just as Jenkins v. Anderson, 447 U.S. 231, 235-39, 100 S.Ct. 2124, 2127-30, 65 L.Ed.2d 86 (1980), had held that use of silence to impeach a testifying defendant does not violate the self-incrimination clause of the federal Bill of Rights. Turning to the due process inquiry, the court observed that prosecutors may use as impeachment silence preceding the delivery of the warnings required by Miranda v. Arizona, 384 U.S. 436, 86 S.Ct. 1602, 16 L.Ed.2d 694 (1966). See Fletcher v. Weir, 455 U.S. 603, 102 S.Ct. 1309, 71 L.Ed.2d 490 (1982). Nothing in the record implies that Brecht was told, until his arraignment, that he had a right to remain silent. It follows from Jenkins and Fletcher, the court held, that the prosecutor could impeach Brecht’s testimony at trial by observing that he did not tell anyone before his arraignment the story he told on the stand at trial. Silence at and after the arraignment was another matter under Doyle v. Ohio, 426 U.S. 610, 96 S.Ct. 2240, 49 L.Ed.2d 91 (1976), the court thought, because advice that you have a right to be silent implies that silence may not be used against you. Twice the prosecutor asked Brecht whether he told anyone at any time before trial that the shooting was accidental; twice the prosecutor remarked in closing argument that Brecht told no one; these four references, the court thought, invited the jury to draw an adverse inference from post-arraignment silence and thus violated the rule established in Doyle.
This did not avail Brecht, however, because the court deemed four references in a four-day trial to be peripheral; the error, the court held, was harmless beyond a reasonable doubt. Error in the ruling on the motion in limine the court put into the same category. As for the evidence of Brecht’s homosexuality: the court held this no error at all. Although Molly Hartman testified that her husband had not “openly expressed]” his disapproval of homosexuality to Brecht, and that there had been no set-to, she also testified that her husband *1366“didn’t want a brother-in-law who was a homosexual.” (The district court’s opinion reproduces this part of her testimony, 759 F.Supp. at 505.) The Supreme Court of Wisconsin concluded that Roger Hartman’s “disapproval was conveyed to Brecht by the restrictions Hartman and his wife imposed on Brecht while he lived in their home. These restrictions included that Brecht was to refrain from homosexual conduct and from inviting homosexual friends to the Hartman home. Thus, it is not unreasonable to infer that Brecht’s homosexuality, which prompted restrictions on his lifestyle, could furnish a motive for the shooting.” 143 Wis.2d at 321, 421 N.W.2d at 105.
The district court continued this game of Ping-Pong by issuing a writ of habeas corpus. Chief Judge Crabb agreed with the Court of Appeals of Wisconsin on all issues: she concluded that the comments on Brecht’s silence violated the due process clause, that the evidentiary rulings were erroneous, and that none of the errors is harmless beyond a reasonable doubt. We summarize only her grounds of disagreement with the Supreme Court of Wisconsin.
Concerning evidence of homosexuality, the district judge wrote: “Discussing a person’s sexual preference and setting restrictions on that person’s behavior while staying in one’s home are not equivalent to communicating disapproval or condemnation to that person.” 759 F.Supp. at 509. The references to Brecht’s homosexuality therefore were irrelevant, and, given the subject matter, prejudicial.
Disagreement on whether the references to post-arraignment silence were harmless was driven largely by the different questions the courts asked. The Supreme Court of Wisconsin asked whether these four references added substantial force to the many proper references to pre-arraignment silence and the lies Brecht told to the police; it concluded that the incremental effect was small. The district court, by contrast, asked whether the evidence against Brecht was irrefutable, stating that “the Court of Appeals for the Seventh Circuit requires that the other evidence of guilt be ‘overwhelming’ before concluding that a constitutional error was harmless.” 759 F.Supp. at 508. It then concluded that the evidence against Brecht is less than overwhelming: there was no apparent motive, and the flight may have been inspired by panic rather than a desire to escape. The court wrote that the evidence about the trajectory of the bullet “may be inconsistent with [Brecht’s] claim that the shooting was an accident, but it is not conclusive.” 759 F.Supp. at 508.
The district court ordered Brecht released unless retried within 60 days. We stayed this order and now become the fourth court in sequence to disagree with its immediate predecessor.
II
This case is here under 28 U.S.C. § 2254(a), which says that a court shall entertain a petition for habeas corpus “only on the ground that [the petitioner] is in custody in violation of the Constitution or laws or treaties of the United States.” “A federal court may not issue the writ on the basis of a perceived error of state law.” Pulley v. Harris, 465 U.S. 37, 41, 104 S.Ct. 871, 875, 79 L.Ed.2d 29 (1984). See also, e.g., Lewis v. Jeffers, — U.S. -, 110 S.Ct. 3092, 3102, 111 L.Ed.2d 606 (1990); Smith v. Phillips, 455 U.S. 209, 221, 102 S.Ct. 940, 948, 71 L.Ed.2d 78 (1982); Rose v. Hodges, 423 U.S. 19, 22, 96 S.Ct. 175, 177, 46 L.Ed.2d 162 (1975); Bates v. McCaughtry, 934 F.2d 99 (7th Cir.1991); Bell v. Duckworth, 861 F.2d 169 (7th Cir.1988); Jones v. Thieret, 846 F.2d 457 (7th Cir.1988). Whether the probative force of Brecht’s homosexuality exceeded its potential for improper use, and whether the prosecutor may put a defendant’s prior convictions before the jury, are staples in the law of evidence. Errors in admitting or excluding evidence do not violate the Constitution unless the trial as a whole is so infected that the verdict is no longer reliable. Cupp v. Naughten, 414 U.S. 141, 94 S.Ct. 396, 38 L.Ed.2d 368 (1973); Lisenba v. California, 314 U.S. 219, 228, 62 S.Ct. 280, 286, 86 L.Ed. 166 (1941). But see McGuire v. Estelle, 902 F.2d 749 (9th Cir.1990), re*1367hearing denied, 919 F.2d 578 (1990), cert. granted, — U.S-, 111 S.Ct. 1071, 112 L.Ed.2d 1177 (1991). The district court did not ask whether Wisconsin conducted a trial that was, all things considered, adequate to sift guilt from innocence; it asked instead whether there were errors, and whether these were harmless beyond a reasonable doubt. Given Donnelly v. De-Christoforo, 416 U.S. 637, 94 S.Ct. 1868, 40 L.Ed.2d 431 (1974), and Darden v. Wainwright, 477 U.S. 168, 106 S.Ct. 2464, 91. L.Ed.2d 144 (1986), not to mention Cupp and Lisenba, that is the wrong approach.
Even taking the subjects in isolation, as the district judge did, does not call into question the validity of the state’s judgment. Whether the Hartmans’ interdict on homosexual acts and homosexual friends at their home conveyed to Brecht Roger Hartman's reprehension of his sexual preference, and thus perhaps engendered Brecht’s hostility, is a nice question of inference. A state may choose to give this question to a jury. The district judge substituted her assessment for that of the Supreme Court of Wisconsin: she did not say that its opinion is unreasoned or beyond the bounds of dispute; she simply disagreed. That is not an appropriate stance for a federal judge engaged in collateral review of a state conviction. “[R]espect for a state court’s findings of fact and application of its own law counsels against the sort of de novo review undertaken” by the district court. Lewis, 110 S.Ct. at 3102.
United States v. Gillespie, 852 F.2d 475, 478-79 (9th Cir.1988), which the district court cited for the proposition that evidence of homosexuality is prejudicial, is another kettle of fish. In a federal prosecution for sexually molesting a three-year-old girl, the prosecution introduced evidence that the defendant had a homosexual relationship with his adoptive father, inviting the inference that anyone departing from the sexual norm is likely to violate little girls. There is no reason to think that homosexuals are likely to be pedophiles; the court properly thought the introduction of this evidence an abuse of discretion under Fed.R.Evid. 404(b). Gillespie was not a constitutional decision, and the use made of homosexuality in that case bears no resemblance to its use in this one.
Now for the trial court’s ruling that if the probation officer testified, the prosecution could use Brecht’s bad-check convictions in cross-examination. The Supreme Court of Wisconsin has determined authoritatively that the court erred under state law. It does not follow that Wisconsin has violated the Constitution of the United States. The district court did not adduce any authority for its conclusion. Brecht’s brief in this court likewise offers no authority, although he does try to distinguish Michelson v. United States, 335 U.S. 469, 69 S.Ct. 213, 93 L.Ed. 168 (1948), the state’s principal case.
Had the identical subject arisen in federal court, there would have been no reversible error, period. Luce v. United States, 469 U.S. 38, 105 S.Ct. 460, 83 L.Ed.2d 443 (1984), holds that a defendant who wishes to contest a ruling allowing the use of a conviction in cross-examination must present the testimony and endure the cross-examination. Otherwise, the Court concluded, it is impossible to put in context the use made of the conviction and to evaluate a claim of prejudice. Just so here. We have no idea — since the probation officer did not testify, and the convictions were never laid before the jury — how they would have been used, and with what effect, if any, on the trial. We also do not know how valuable was the testimony foregone.
It is, at all events, difficult to accept the idea that the Constitution forbids a state to inform the jury of a defendant’s prior convictions. Brecht testified, and a federal jury normally learns of a testifying defendant’s crimes. Fed.R.Evid. 609(a). Wisconsin does things differently, which may account for the prosecutor’s effort to introduce the convictions when examining the probation officer, but this difference hardly makes anathema of the knowledge. Especially not when the witness was going to be the defendant’s probation officer; it does seem apt to ask such a person what crime the defendant was on probation for. The *1368jury knew from the circumstances of Brecht’s presence in the Hartman home that Brecht was a convicted felon; learning that the felony was writing bad checks (as opposed to, say, mayhem or selling drugs) is an unlikely prospect for constitutional error in a murder case. We turn, then, to the prosecutor’s comment on Brecht’s failure to say at any time before trial that the shooting was a mishap.
Ill
“Common law traditionally has allowed witnesses to be impeached by their previous failure to state a fact in circumstances in which that fact naturally would have been asserted. Each jurisdiction may formulate its own rules of evidence to determine when prior silence is so inconsistent with present statements that impeachment by reference to such silence is probative.” Jenkins, 447 U.S. at 239, 100 S.Ct. at 2129 (citations omitted). If Brecht indeed shot Roger Hartman by accident, he had every reason to tell the police early and often. How else was he to win release from prison—or to ensure that investigators preserved evidence that would support his version of events? Why would an innocent man keep his sister in the dark about the circumstances of her husband’s death? Would he not at least apologize for the accident? Silence supports an inference that it took Brecht a while to cook up a story. Fabricated defenses are best withheld until after the state has rested, not only to catch the opposition off guard but also to avoid any blatant inconsistencies with the facts the prosecution adduces.
But for Doyle v. Ohio, the prosecutor would have been entitled to ask the questions and argue the case precisely as he did. Doyle holds that however natural the inference from silence may be, the prosecutor may not undermine the implications of the Miranda warnings. A suspect told that he has a right to remain silent, Doyle concluded, may not be bushwhacked by an argument that silence implies guilt. Fletcher establishes that Doyle applies exclusively to silence after the suspect has been informed about the privilege of silence. Nothing in this record implies that Brecht received such advice before his arraignment. Thus, as the Supreme Court of Wisconsin held, questions and arguments about silence before arraignment are proper, while the questions and remarks in closing argument concerning whether Brecht told his story “at any time” before trial stand condemned by Doyle, for they referred in part to post-warning silence. Our question is whether the error was harmless.
A
Whether the error was harmless depends, for reasons discussed later, on what kind of entitlement Doyle establishes. Does Doyle implement a part of the Bill of Rights, or is it instead a constitutional prophylaxis?
One possibility is that comment on silence violates the self-incrimination clause of the fifth amendment, applied to the states by the fourteenth. No state may compel a person to incriminate himself, which sets up an argument that an inference from silence is a form of compulsion to speak. The defendant made just such an argument in Jenkins, and the Court rejected it—at least when the accused testifies, as Brecht did. 447 U.S. at 235-39, 100 S.Ct. at 2127-30. The privilege against compulsory self-incrimination may be waived. Raffel v. United States, 271 U.S. 494, 46 S.Ct. 566, 70 L.Ed. 1054 (1926), holds that a defendant who takes the stand waives the privilege to the extent of authorizing questions and comments along the line: Why didn’t you tell this story earlier? Raffel testified at his second trial after keeping mum at the first. Jenkins holds that Raffel is equally applicable when the silence precedes trial. Justice Stevens, concurring in Jenkins, doubted the vitality of Raffel (treating testimony as a “waiver” is strained) but concluded that comment on silence is not compulsion to speak, so there is no privilege to waive. 447 U.S. at 241-45, 100 S.Ct. at 2130-33. One way or the other, Jenkins shows that what happened at Brecht’s trial did not violate the self-incrimination clause.
*1369Doyle suggests a different rationale: Comment on silence threatens the conviction of the innocent, and thus violates the due process clause. The Court remarked that “every post-arrest silence is insolubly ambiguous”, 426 U.S. at 617, 96 S.Ct. at 2244, because the suspect may be exercising his constitutional right to silence. This theme came from United States v. Hale, 422 U.S. 171, 95 S.Ct. 2133, 45 L.Ed.2d 99 (1975), in which the Court found post-arrest silence too ambiguous to be an appropriate basis of inferences in federal trials. Yet if the suspect was exercising his right to silence, he could explain as much to the jury and so defuse the prosecutor's argument. That silence can carry multiple messages does not make the ambiguity insoluble. Few pieces of evidence standing alone support confident conclusions. Trial is an occasion for the cumulation of evidence so that one strand may illuminate another. Bourjaily v. United States, 483 U.S. 171, 179-80, 107 S.Ct. 2775, 2780-81, 97 L.Ed.2d 144 (1987); United States v. Martinez de Ortiz, 907 F.2d 629, 633 (7th Cir.1990) (in banc).
In the years since Doyle the Court has thought better of its suggestion that no argument from silence may be presented to a jury. Jenkins allowed a prosecutor to urge inferences from pre-arrest silence, remarking that ambiguity “is a question of state evidentiary law.” 447 U.S. at 239 n. 5, 100 S.Ct. at 2130 n. 5. Fletcher extended this treatment to post-arrest silence by suspects who did not receive Miranda warnings. Although one could have said that most suspects know of their privilege to remain silent whether or not the police tell them, and that this knowledge creates the same ambiguity mentioned in Doyle, the Court held that difficulties of inference are subjects for state law. Fletcher limited the ambiguity rationale of Doyle to federal cases, observing that states may draw such inferences as their domestic law permits unless the trial as a whole is fundamentally flawed. 455 U.S. at 604-05, 102 S.Ct. at 1310-11. South Dakota v. Neville, 459 U.S. 553, 564-66, 103 S.Ct. 916, 922-24, 74 L.Ed.2d 748 (1983), reinforced this conclusion in holding that a state may draw adverse inferences from a driver’s refusal to allow a blood-alcohol test. Then Wainwright v. Greenfield, 474 U.S. 284, 106 S.Ct. 634, 88 L.Ed.2d 623 (1986), put the icing on the cake by holding that states may not draw even well-justified inferences from post-warning silence. The prosecutor tried to distinguish Doyle by observing that the silence in Greenfield’s case was not ambiguous, to which the Court replied that probativeness is irrelevant. 474 U.S. at 293-94 & n. 12, 106 S.Ct. at 639-40 & n. 12. The reference to ambiguity in Doyle, the Court now observed, just “added weight”, id. at 294, 106 S.Ct. at 640, to the opinion’s real foundation — the desire to prevent the state from undercutting the value of the Miranda warnings. Doyle itself said as much, 426 U.S. at 617-18 n. 8, 96 S.Ct. at 2244-45 n. 8, when responding to the dissent of Justice Stevens, who wrote for the Court in Greenfield.
Justice White first offered the anti-mousetrapping ground in a brief opinion concurring in Hale, 422 U.S. at 182-83, 95 S.Ct. at 2139-40. Doyle and all of the opinions in its wake have quoted and adopted this view as the Court's. Doyle put it this way:
[Wjhile it is true that the Miranda warnings contain no express assurance that silence will carry no penalty, such assurance is implicit to any person who receives the warnings. In such circumstances, it would be fundamentally unfair and a deprivation of due process to allow the arrested person’s silence to be used to impeach an explanation subsequently offered at trial.
426 U.S. at 618, 96 S.Ct. at 2245. Jenkins, which allows reference to pre-arrest silence, and Fletcher, which allows reference to post-arrest silence if the officers neglected to give Miranda warnings, emphasize that the state did not afford — and therefore could not renege on — any assurance that silence is costless. Cf. Anderson v. Charles, 447 U.S. 404, 408, 100 S.Ct. 2180, 2182, 65 L.Ed.2d 222 (1980) (“Doyle does not apply to cross-examination that merely inquires into prior inconsistent statements. Such questioning makes no unfair use of silence, because a defendant who voluntar*1370ily speaks after receiving Miranda warnings has not been induced to remain silent.”)- Greenfield reinforces this understanding in the course of explaining, 474 U.S. at 292-94, 106 S.Ct. at 639-40, that even reliable inferences from silence are forbidden because they amount to governmental self-contradiction.
Doyle thus rests on a single proposition: Having implied when giving warnings that silence is safe, the government may not reverse course at trial. The rule of Doyle reduces the chance that Miranda warnings will injure the persons they are supposed to protect. Doyle makes the world safe for Miranda. But the warnings required by Miranda are not themselves part of the Constitution. They are designed to inform the suspects of their entitlements and so reduce the likelihood of subtle coercion of people in custody. Oregon v. Elstad, 470 U.S. 298, 304-09, 105 S.Ct. 1285, 1290-93, 84 L.Ed.2d 222 (1985); New York v. Quarles, 467 U.S. 649, 654, 104 S.Ct. 2626, 2630, 81 L.Ed.2d 550 (1984); Michigan v. Tucker, 417 U.S. 433, 442-46, 94 S.Ct. 2357, 2362-65, 41 L.Ed.2d 182 (1974). Doyle is therefore a prophylactic rule designed to protect another prophylactic rule from erosion or misuse. It has nothing to do with the defendant’s guilt and everything to do with insisting that the government play straight with those it prosecutes.
B
Whether the prosecution’s violation of Doyle’s rule in Brecht's case is harmless beyond a reasonable doubt is a question about which reasonable persons can disagree — have disagreed. The district judge and the three judges of the state’s appellate court believe that the error was not harmless under this standard. Six justices of the Supreme Court of Wisconsin reached the contrary conclusion. (The seventh, Justice Ceci, did not participate in the decision.) The district judge thought her inability to deem the evidence “overwhelming” dispositive on the question. Some opinions of this court support such a perspective. E.g., United States ex rel. Savory v. Lane, 832 F.2d 1011, 1020 (7th Cir.1987); Dudley v. Duckworth, 854 F.2d 967, 972 (7th Cir.1988). “Beyond a reasonable doubt” is an exacting standard. See Yates v. Evatt, — U.S.-, 111 S.Ct. 1884, 114 L.Ed.2d 432 (1991); Arizona v. Fulminante, — U.S. -, 111 S.Ct. 1246, 1257-61, 113 L.Ed.2d 302 (1991). The state, by contrast, submits that errors are harmless when their marginal effect is slight even though the evidence leaves room for debate. This position, too, has some support in our decisions. E.g., Phelps v. Duckworth, 772 F.2d 1410, 1413-15 (7th Cir.1985) (in banc); Fencl v. Abrahamson, 841 F.2d 760, 769 (7th Cir.1988).
We could trudge through the record and add to the score, making the count anywhere from 9-4 for harmlessness (if we should agree with the Supreme Court of Wisconsin) to 7-6 for the opposite conclusion. The antecedent question, however, is whether collateral review of state convictions in federal court is nothing more than a replay of the direct appeal, as one federal district judge and then three federal appellate judges duplicate the work of the more numerous state judges. Although arguing that duplication of effort should end in its favor, Wisconsin also urges us to revisit the question whether a federal court should apply the “reasonable doubt” standard, derived from Chapman v. California, 386 U.S. 18, 87 S.Ct. 824, 17 L.Ed.2d 705 (1967), on collateral review of Doyle violations. United States ex rel. Miller v. Greer, 789 F.2d 438 (7th Cir.1986) (in banc), used the reasonable doubt test, but the Supreme Court reversed our judgment on other grounds under the name Greer v. Miller, 483 U.S. 756, 107 S.Ct. 3102, 97 L.Ed.2d 618 (1987), and so deprived our opinion of authority. The Court granted certiorari to identify the proper standard on collateral attack but did not reach the issue because, it held, there had been no error in the first place. Id. at 761 n. 3, 765 n. 6, 107 S.Ct. at 3106 n. 3, 3108 n. 6. Justice Stevens, who alone considered the question on which review had been granted, concluded that a distinct standard of harmlessness applies on collateral review. Id. at 767-69, 107 S.Ct. at 3109-11 (concurring opinion). Two other Justices have cit*1371ed that opinion with approval. Duckworth v. Eagan, 492 U.S. 195, 212, 109 S.Ct. 2875, 2884, 106 L.Ed.2d 166 (1989) (O’Connor & Scalia, JJ., concurring). No Justice has taken a contrary view.
Not long ago a majority of this court treated the question as open. In Hunter v. Clark, 934 F.2d 856 (7th Cir.1991) (in banc), two judges concluded that the reasonable doubt standard should not be employed in the collateral enforcement of prophylactic rules. Id. at 865-66 (concurring opinion). Four other judges called the concurring opinion “persuasive” and remarked that it “may well be correct” but found it unnecessary to essay a final decision. Id. at 859 n. 1 (plurality opinion). The other five members of this court did not discuss the question. See also Hanrahan v. Greer, 896 F.2d 241, 244-45 (7th Cir.1990) (calling question open but avoiding resolution because argument had not been preserved in district court); Fencl, 841 F.2d at 766-67 n. 6 (flagging question but avoiding resolution). Judges of this court have been bandying the subject about for six years, ever since Phelps was decided in September 1985. Although Savory, 832 F.2d at 1016-17 & n. 2, reiterated the approach of Miller, subsequent developments, especially the reservation in Hunter, permit renewed attention. No interest of the judicial system (or the parties) would be served by postponing yet again resolution of a question the Supreme Court thought certworthy in 1987. The State has asked us to apply the standard of United States v. Lane, 474 U.S. 438, 449, 106 S.Ct. 725, 732, 88 L.Ed.2d 814 (1986), and Kotteakos v. United States, 328 U.S. 750, 776, 66 S.Ct. 1239, 1253, 90 L.Ed. 1557 (1946), under which only an error that exerts substantial influence on the course of the trial and produces actual prejudice to the accused vitiates the judgment. By ignoring this request, Brecht has forfeited any possibility that the state neglected to preserve the claim in the district court. (The defendant made a vigorous waiver argument in Han-rahan, quite unlike Brecht’s strategy. We do not know whether Wisconsin made this argument to the district judge and shall not endeavor to find out.)
■ The standard of “harmlessness” has never been a unitary one. Contrast Lane with Chapman. It varies with the nature of the underlying right, the tools available to correct the violation at trial, and the number of layers of review. Chapman establishes that the standard an appellate court should employ on direct review is whether the error is harmless beyond a reasonable doubt. Does this carry over to collateral attack? The Supreme Court has never addressed that question.
We know from Stone v. Powell, 428 U.S. 465, 96 S.Ct. 3037, 49 L.Ed.2d 1067 (1976), that § 2254 does not require federal courts on collateral attack to employ for every constitutional claim the same standards state courts must use. Stone holds that when a defendant invokes the exclusionary rule that has been devised to implement the fourth amendment, a federal court should inquire only whether the state gave the defendant a full and fair opportunity to litigate the claim. Error in application does not support a writ of habeas corpus. Stone observed that the exclusionary rule is an extra-constitutional device that helps motivate adherence to the fourth amendment, but that exclusion is not itself compelled by the Constitution. See also, e.g., United States v. Leon, 468 U.S. 897, 104 S.Ct. 3405, 82 L.Ed.2d 677 (1984); cf. Henry P. Monaghan, The Supreme Court, 1974 Term — Foreword: Constitutional Common Law, 89 Harv.L.Rev. 1 (1975). It followed, the Court held, that federal courts engaged in collateral review need not replicate the efforts of the judges of the state tribunals; it is enough to ensure that states adjudicate these claims. It may well be that the Court will establish a similar system for claims under Miranda, which like the exclusionary rule depends on prudential rather than textual or historical support. Justice O’Connor (joined by Justice Scalia) so argued in Eagan, 492 U.S. at 205-14, 109 S.Ct. at 2881-86, and the majority left the issue open. 492 U.S. at 201 n. 3, 109 S.Ct. at 2879 n. 3. If Miranda ultimately joins the exclusionary rule, claims under Doyle are logical companions, for just as Miranda creates a buffer *1372around the self-incrimination clause, so Doyle creates a buffer around Miranda.
Selecting a deferential standard by which to assess harmlessness is an alternative to confining collateral review as tightly as Stone does — an alternative that leaves the federal court with substantially greater power to ensure that the state court is taking seriously its obligation to enforce the substantive rules. It is a middle ground between federal review that duplicates the direct appeal and federal review that is, after the fashion of Stone, next to no review.
For most of this nation’s history “next to none” was an accurate description of collateral review of state judgments. Section 14 of the Judiciary Act of 1789 gave the federal courts jurisdiction to issue writs only when persons were held in federal custody. 1 Stat. 81-82. In 1867 a Congress concerned about the behavior of states that had joined the Confederacy extended federal power to state prisoners. 14 Stat. 385-86. In 1867, as in 1789, the writ ran only to release persons held without authority— that is to say, without “jurisdiction.” See Paul M. Bator, Finality in Criminal Law and Federal Habeas Corpus for State Prisoners, 76 Harv.L.Rev. 441, 463-69 (1963). One convicted by a tribunal with jurisdiction over subject-matter and person could obtain no collateral relief, however grievously the tribunal erred. Shortly after Congress extended federal power to state prisoners, the Supreme Court began pouring more into the vessel of “jurisdiction”. In 1915 the Court equated a kangaroo court with lack of jurisdiction, Frank v. Mangum, 237 U.S. 309, 330-35, 35 S.Ct. 582, 588-90, 59 L.Ed. 969 (1915), and by 1942 the distinction between “no jurisdiction” and “grievous error” had eroded so substantially that the Court discarded the limitation. Waley v. Johnston, 316 U.S. 101, 62 S.Ct. 964, 86 L.Ed. 1302 (1942) (conviction based on coerced confession). Not for a further 11 years, however, did the Court first use collateral review as a supplement to direct appeal, re-deciding procedural questions (such as the validity of the indictment) already fully considered by the state court; until Brown v. Allen, 344 U.S. 443, 73 S.Ct. 397, 97 L.Ed. 469 (1953), it had been necessary to show an error of special gravity that undermined confidence in the verdict. That is to say, the device through which the Court expanded review beyond strictly jurisdictional claims was to equate omission of those procedures central to the protection of the innocent with the lack of jurisdiction.
If the fifteen years that followed Brown witnessed an extraordinary extension of federal review of state judgments, the fourteen years since Wainwright v. Sykes, 433 U.S. 72, 97 S.Ct. 2497, 53 L.Ed.2d 594 (1977), have experienced a movement in the other direction. Sykes held that forfeitures in state court also block review in federal court unless the defendant shows both external cause and prejudice — not simply that the error had an adverse effect, but that it could threaten the conviction of the innocent. Cf. Henry J. Friendly, Is Innocence Irrelevant? Collateral Attack on Criminal Judgments, 38 U.Chi.L.Rev. 142 (1970). Stone effectively eliminated collateral review of claims under the fourth amendment. Recently the Court overruled Fay v. Noia, 372 U.S. 391, 83 S.Ct. 822, 9 L.Ed.2d 837 (1963), one of the milestones in the expansion of collateral review. Coleman v. Thompson, — U.S.-, 111 S.Ct. 2546, 2565, 115 L.Ed.2d 640 (1991) (observing that Fay “was based on a conception of federal/state relations that undervalued the importance of state procedural rules.”). See also, e.g., McCleskey v. Zant, — U.S. -, 111 S.Ct. 1454, 113 L.Ed.2d 517 (1991) (enlarging the doctrine of “abuse of the writ”).
Today — as for most of our history — the Court puts considerations of finality and federalism at the forefront of any discussion of the scope of collateral review. One who seeks that review must justify the federal role; it is not enough to cry “Constitution!”. There are many shades of constitutional rules; those that protect the accuracy of the fact-finding process receive the greatest protection; rules that may be found in the text (such as the self-incrimination clause), even though they promote values other than accurate resolution of *1373factual contests, receive greater protection than doctrines that courts have devised in the penumbras of constitutional rules. We deal with such a penumbral doctrine — one that until Brown would not have been open at all on collateral review.
In deciding how (even whether) a federal court should examine the state’s decision concerning such a supplementary doctrine, we could recite themes of federalism and finality — but these are so trumpeted in recent opinions such as McCleskey and Coleman that repetition would be tendentious. See also, e.g., Teague v. Lane, 489 U.S. 288, 308-09, 109 S.Ct. 1060, 1073-74, 103 L.Ed.2d 334 (1989); Kuhlmann v. Wilson, 477 U.S. 436, 452-55, 106 S.Ct. 2616, 2626-28, 91 L.Ed.2d 364 (1986). One dominant theme is that doctrines protecting the innocent, and claims of actual innocence, receive special protection. See Murray v. Carrier, 477 U.S. 478, 496, 106 S.Ct. 2639, 2649, 91 L.Ed.2d 397 (1986) (the “cause and prejudice” preconditions to collateral review of claims forfeited in state court do not apply “where a constitutional violation has probably resulted in the conviction of one who is actually innocent.”); Kuhlmann, 477 U.S. at 452-54, 106 S.Ct. at 2626-27 (showing of probable innocence is condition of second review of same issue). Three additional practical considerations stand out.
First, collateral review followed by a new trial means delay, often extended. Passage of time diminishes the courts’ ability to resolve factual disagreements accurately. Trials long after the event, when memories have faded, are less likely to convict the guilty and acquit the innocent. More: at a new trial there will be opportunities for new errors, which may be as bad as the error in Brecht’s trial. Sometimes inaccuracy hurts the defendant, but more often it counts against the prosecution, which bears the burden of persuasion, and thus against the public interest in the accurate enforcement of the law. When the constitutional blunder diminished the accuracy of the factfinding process, then retrial even after many intervening years may make a just resolution more likely; when the constitutional entitlement is one that tolerates a reduction in accuracy (as does the self-incrimination clause) in pursuit of other objectives, again the costs of deferred trial may be worth bearing. When the rule neither promotes accuracy nor can be located in the Constitution itself — when the foundation of the doctrine is avowedly prudential — then courts must take into account all of the costs as well as the benefits. Delay makes it substantially more costly to enforce a prophylactic rule such as Doyle on collateral attack than on direct appeal, which implies a corresponding adjustment in doctrine. See Thomas W. Merrill, The Common Law Powers of the Federal Courts, 52 U.Chi.L.Rev. 1, 53 (1985).
Second, federal courts, knowing the high costs of ordering retrials many years after the events, may be inclined to dilute the standard of harmlessness while retaining the “reasonable doubt” formula, excusing errors that would not be tolerated on direct appeal. A careful reader of our opinions would find some traces of this. E.g., Hanrahan v. Thieret, 933 F.2d 1328, 1336 (7th Cir.1991). State courts reading these opinions then relax their own vigilance, an unhappy consequence that could be avoided by distinguishing the standards of harmlessness on direct and collateral review.
Third, while federal courts re-do the decisions already made by state courts, federal business languishes. Time to review closed cases may be had only at the expense of time to decide pending ones. Serving as super-appellate tribunals in the state system, federal courts necessarily devote less time to cases that have yet to receive an initial decision. Less time per case means an increased chance of error. It also means a longer queue, during which disputes fester and memories fade, impairing the federal courts’ ability to resolve accurately cases pending on their primary dockets. In some federal districts civil trials have become exceptional events. There are many causes of this, but each contributes at the margin, and each segment of judicial business must justify its imposition on resources the judicial system cannot control or augment.
*1374Selecting a standard for assessing harmless error, no less than deciding when the prosecutor can use evidence the Constitution says he should not have at all, requires an assessment of the marginal benefits and costs of enforcement. The difference between a Chapman standard of review on collateral attack and a Kotteakos standard is not likely to impair the enforcement of Doyle. Recall that Kotteakos supplies the standard — whether an error “had substantial and injurious effect or influence in determining the jury’s verdict”, 328 U.S. at 776, 66 S.Ct. at 1253; see also Lane, 474 U.S. at 449, 106 S.Ct. at 732 — that federal courts use in evaluating claims that federal statutes and rules have been violated in federal trials. No one believes that it affords insufficient protection to these federal norms. Under any of the competing definitions, the substantive rule (and review under the most stringent standard in state court) informs prosecutors that an effort to draw the jury’s attention to the defendant’s silence reduces the chance of conviction, and tells trial judges how to respond to the prosecutors’ efforts. These two messages, unaffected by the standard of harmless error, are the principal advantages of the rule. When the rule is as far removed from protecting the innocent as is Doyle, there is no need for the strict enforcement that the Chapman standard creates.
C
What we have said would be academic if Congress or the Supreme Court has directed the application of the Chapman standard on collateral review. But neither institution has done this. Congress has been silent on the standard of review. Indeed, with the exception of 28 U.S.C. § 2254(d), Congress has been silent on most procedural questions. See John C. Jeffries, Jr. & William J. Stuntz, Ineffective Assistance and Procedural Default in Federal Habeas Corpus, 57 U.Chi.L.Rev. 679, 707-09 (1990). Chapman was a direct appeal, as are most of the other cases in which the Supreme Court applied the “reasonable doubt” standard to state convictions. That the standard of harmless error may vary between direct and collateral appeal is a commonplace. For example, a violation of Fed.R.Crim.P. 11, which establishes procedures for taking guilty pleas, is enforceable on direct appeal by almost automatic reversal. McCarthy v. United States, 394 U.S. 459, 89 S.Ct. 1166, 22 L.Ed.2d 418 (1969). On collateral attack, however, Rule 11 is enforceable only to the extent that the disregard of its provisions makes the proceedings fundamentally unfair. United States v. Timmreck, 441 U.S. 780, 99 S.Ct. 2085, 60 L.Ed.2d 634 (1979). The difference is not attributable to the nonconstitutional status of Rule 11, for the Rule is a “law,” and the pertinent statute, 28 U.S.C. § 2255, allows collateral attack on the ground that a person is in custody “in violation of the Constitution or laws of the United States”. See Davis v. United States, 417 U.S. 333, 94 S.Ct. 2298, 41 L.Ed.2d 109 (1974). The change in the.standard of review instead rests on the conclusion that one round of review is sufficient unless something has gone very wrong indeed. Collateral review is not supposed to be a replay of the direct appeal, Sunal v. Large, 332 U.S. 174, 181-82, 67 S.Ct. 1588, 1592-93, 91 L.Ed. 1982 (1947), a point emphasized when the Court held in United States v. Frady, 456 U.S. 152, 102 S.Ct. 1584, 71 L.Ed.2d 816 (1982), that blunders that could support reversal as “plain error” on direct appeal do not support collateral relief.
The same may be said for review of state convictions. See Henderson v. Kibbe, 431 U.S. 145, 154 & n. 13, 97 S.Ct. 1730, 1737 & n. 13, 52 L.Ed.2d 203 (1977), which holds that only in the rarest of cases will an instruction to which no objection was made at trial support collateral attack, even though the same instruction might have been “plain error” on direct appeal. Ordinary trial errors (even constitutional ones) that might lead to reversal on appeal do not support collateral review in federal court unless the trial as a whole has miscarried, impugning the accuracy of the judgment. See, e.g., Darden, Donnelly, Cupp, and Lisenba. Expressions throughout the Doyle line imply that the Court believes these cases salient. E.g., Fletcher, 455 *1375U.S. at 605, 102 S.Ct. at 1311 (citing Cupp for the proposition that the ambiguity of silence supports a new trial only when it degrades the accuracy of the fact-finding apparatus); Miller, 483 U.S. at 765-67, 107 S.Ct. at 3108-10 (applying the Donnelly approach to determine whether a prosecutor’s efforts to argue inferences from silence violate Doyle at all).
So far as we can tell, only twice has the Supreme Court used the Chapman standard on collateral attack. The first time, Rose v. Clark, 478 U.S. 570, 106 S.Ct. 3101, 92 L.Ed.2d 460 (1986), the central question was whether an unconstitutional burden-shifting instruction ever could be harmless error. The Court held that it may be deemed harmless under the Chapman standard. The second time, Yates v. Evatt, required the Court to refine its holding in Rose, which was not entirely consistent in its explanation of Chapman. See 111 S.Ct. at 1892 & n. 8. Yates held that the Supreme Court of South Carolina erred in withholding collateral relief when it employed a standard of harmlessness at variance with Chapman’s. Yates illustrates the value of the Chapman standard when the constitutional rule is designed to protect the innocent from wrongful conviction. South Carolina’s trial court had charged the jury in a murder case not only that malice could be presumed from the use of a deadly weapon but also that malice could be presumed when the death occurs in the course of an unlawful act. Both of these presumptions had been declared unconstitutional because they allowed the state to evade its burden of proving guilt beyond a reasonable doubt. See Yates v. Aiken, 484 U.S. 211, 108 S.Ct. 534, 98 L.Ed.2d 546 (1988); Sandstrom v. Montana, 442 U.S. 510, 99 S.Ct. 2450, 61 L.Ed.2d 39 (1979); In re Winship, 397 U.S. 358, 90 S.Ct. 1068, 25 L.Ed.2d 368 (1970). The state court, on remand from Yates v. Aiken, proceeded to misstate the facts and all but thumb its nose at the opinion reversing it. (It had earlier denied relief to Yates without even citing the federal cases — and this despite a remand by the Supreme Court with instructions to apply Francis v. Franklin, 471 U.S. 307, 105 S.Ct. 1965, 85 L.Ed.2d 344 (1985). See 111 S.Ct. at 1890.) So the Supreme Court again granted review and reversed, explicating the meaning of Chapman in burden-shifting cases.
Yates v. Evatt presents a striking contrast with our case. First, Yates dealt with constitutional rules designed to protect the innocent. Such rules give the strongest justification for independent federal review. Doyle is a penumbral rule. Second, Yates addressed a state court apparently unwilling to accept the federal norm; Wisconsin, by contrast, has applied Doyle conscientiously. Federal courts should discourage recalcitrance and reward full consideration — which they cannot if they review every constitutional claim from scratch. Yates reinforces rather than shakes our conviction that the “reasonable doubt” standard, compelling the state to justify the judgment with the same burden it bears before the jury in the first instance, is out of place in the collateral enforcement of prophylactic rules.
D
We hold that the standard of harmlessness on collateral enforcement of prophylactic rules is the one stated in Kotteakos and Lane — whether the error “had substantial and injurious effect or influence in determining the jury’s verdict.” Kotteakos, 328 U.S. at 776, 66 S.Ct. at 1253; see also Lane, 474 U.S. at 449, 106 S.Ct. at 732. The Court added in Kotteakos that “if one is left in grave doubt, the conviction cannot stand.” 328 U.S. at 765, 66 S.Ct. at 1248. This amply protects the federal interest — and the defendant’s — in enforcing rules that are unrelated to the protection of the innocent.
We recognize that Bass v. Nix, 909 F.2d 297, 304-05 & n. 14 (8th Cir.1990), has elected not to apply the Kotteakos-Lane approach to collateral review of Doyle errors. Bass is founded on a belief that the “insoluble ambiguity” language of Doyle, which is directed to the protection of the innocent, still represents the Supreme Court’s view. As we have shown, it does not; today the probative value of inferences from silence is a question of state law. *1376Because Bass rests on a misunderstanding of the justification of the Doyle rule, we respectfully decline to follow it.*
Application of our approach is straightforward. Brecht does not contend that the four references to his post-arraignment silence had “substantial and injurious effect” and could not do so with a straight face, given the many more, and entirely proper, references to his silence preceding arraignment. Brecht’s sole submission, tracking the district court’s conclusion, is that once any constitutional error has occurred, the judgment must be set aside unless the evidence is “overwhelming”, which Brecht submits it is not. As we have rejected that approach, we conclude that the violation of Doyle does not support a writ of habeas corpus.
Reversed
Because this opinion creates a conflict among the circuits and changes the standard used in this circuit, it was circulated before release to all judges in active service. See Circuit Rule 40(f). A majority did not favor hearing in banc. Chief Judge Bauer and Judges Wood, Cudahy, Flaum, and Ripple voted to hear the case in banc.