with whom
EASTERBROOK and MANION, Circuit Judges, join, concurring and dissenting.
I agree that the convictions should stand (though my reasoning differs from Judge Evans’s), but not that the death sentences should be reversed. Judge Maloney, whose alleged bias is the only issue in this appeal, presided over both phases of the case. There is no basis for supposing him unbiased until the defendants were convicted, then biased at the sentencing hearing. Such a supposition offends common sense. What must be driving the outcome of the appeal is a sense of discomfort with Maloney’s antics that is too great to contemplate executions without acute distress but not too great to contemplate life sentences. That is the only meaning I can assign to Judge Evans’s reference to a “toxic mix.” For Bracy and Collins have failed to show that they were denied due process of law either at trial or in sentencing. To reverse their sentences is merely to compound Maloney’s wrongdoing. To reverse while upholding the convictions is an unprincipled splitting of the difference, rather than legal justice. It is the sort of thing an arbitrator might do or a mediator propose. It would be understandable as a settlement; it is indefensible as a judgment.
Bracy and Collins were convicted in 1981 by a jury in an Illinois state court of three gangster-style murders committed the previous year, and were sentenced to death by the jury. We affirmed the denial of federal habeas corpus relief in Bracy v. Gramley, 81 F.3d 684 (7th Cir.1996). The Supreme Court reversed, 520 U.S. 899, 117 S.Ct. 1793, 138 L.Ed.2d 97 (1997), holding that Bracy had made a sufficient showing *420under Rule 6(a) of the Rules Governing Section 2254 Cases in the United States District Courts to entitle him to conduct discovery concerning his claim that Judge Maloney had been biased. The Court remanded Collins’s case for reconsideration in light of its opinion in Bracy’s case. Collins v. Welborn, 520 U.S. 1272, 117 S.Ct. 2450, 138 L.Ed.2d 209 (1997) (per curiam).
Maloney had been convicted in a federal court in 1993 of various offenses relating to his having taken bribes from criminal defendants during a period that included the year of the petitioners’ trial. See United States v. Maloney, 71 F.3d 645 (7th Cir.1995). He had not solicited or received bribes from Bracy or Collins but they argue that he habitually came down harder on defendants who had not bribed him than he would have done had he not been taking bribes. He did this, they argue, both to deflect any suspicion that might arise in the cases in which he had accepted bribes and as a result acquitted or gone easy on the defendants that he was “soft” on criminals (which might endanger his reelection) and to increase the size and frequency of the bribes offered him.
The Supreme Court held that “if it could be proved, such compensatory, camouflaging bias on Maloney’s part in petitioner’s own ease would violate the Due Process Clause of the Fourteenth Amendment.” 520 U.S. at 905, 117 S.Ct. 1793 (emphasis added). In concluding that Bracy had presented enough evidence of such bias to entitle him to seek additional evidence through discovery, the Court focused on the contention that his trial counsel, Robert McDonnell, who had been appointed by Maloney to represent Bracy, had practiced law with Maloney before the latter had become a judge and that McDonnell “might have been appointed with the understanding that he would not object to, or interfere with, a prompt trial, so that petitioner’s case could be tried before, and camouflage the bribe negotiations in,” a contemporaneous case before Maloney. Id. at 908, 117 S.Ct. 1793. The Court pointed out that “this is, of course, only a theory at this point; it is not supported by any solid evidence of petitioner’s trial lawyer’s participation in any such plan.” Id. But if substantiated, this theory that Bracy’s “trial attorney, a former associate of Maloney’s in a law practice that was familiar and comfortable with corruption, may have agreed to take this capital case to trial quickly so that petitioner’s conviction would deflect any suspicion the rigged ... cases might attract,” id. at 909, 117 S.Ct. 1793, would support “his claim that Maloney was actually biased in petitioner’s own case.” Id. (emphasis in original). The Court rejected the view of Judge Rovner, the dissenting judge in our court, that “petitioner was entitled to relief whether or not he could prove that Maloney’s corruption had any impact on his trial. The latter conclusion, of course, would render irrelevant the discovery-related question presented in this case.” Id. at 903 n. 4, 117 S.Ct. 1793 (citation omitted). Regarding “the correctness of the various discretionary rulings cited by petitioner as evidence of Ma-loney’s bias,” the Court remarked that “many of these rulings have been twice upheld, and that petitioner’s convictions and sentence have been twice affirmed, by the Illinois Supreme Court.” Id. at 906 n. 6., 117 S.Ct. 1793
Twice the Supreme Court said that to provide a basis for relief for Bracy (and hence for Collins) compensatory bias must be shown “in petitioner’s own case.” This means that even if Maloney engaged in compensatory bias in some cases, this would not be enough to justify a conclusion that Bracy and Collins had been convicted and sentenced in violation of due process; they would have to prove that Maloney *421had been biased (“actually biased,” as the Court said) at their trial. Also noteworthy is the Court’s approving reference to the description in our panel opinion of the theory of compensatory bias as “speculative”: “The Court of Appeals, in its opinion, pointed out that this theory is quite speculative; after all, it might be equally likely that a judge who was ‘on the take’ in some criminal cases would be careful to at least appear to favor all criminal defendants, so as to avoid apparently wild and unexplainable swings in decisions and judicial philosophy.” Id. at 906, 117 S.Ct. 1793, citing 81 F.3d at 689-90.
Sometimes the temptation to bias is so great that proof of bias is not required. This is true when the judge has a substantial pecuniary stake in the outcome of the case or when he is bribed by one of the parties. See, e.g., Aetna Life Ins. Co. v. Lavoie, 475 U.S. 813, 825, 106 S.Ct. 1580, 89 L.Ed.2d 823 (1986); Del Vecchio v. Illinois Dept. of Corrections, 31 F.3d 1363, 1370-80 (7th Cir.1994) (en banc); Cartalino v. Washington, 122 F.3d 8, 11 (7th Cir.1997). Given the difficulty of peering into a judge’s mind, a high probability of bias is, in the absence of confession, the most that can ever be proved, and sometimes the objective circumstances alone are enough to establish the requisite probability or at least to establish that no ordinary person would believe that a judge would not yield to such a temptation. But it is apparent from the passages that I have quoted from the Bracy opinion that the Supreme Court does not regard the temptation to engage in compensatory bias as falling into the per se category, where proof of the temptation is enough to entitle a defendant to a new trial because the likelihood that the judge succumbed (perhaps quite unconsciously) is great. If it did fall into the per se category, as Judge Rovner had argued it should, there would have been no occasion to conduct discovery, since the existence of the temptation was conceded and the only question was whether Maloney had yielded to it, either generally or in the trial of Bracy and Collins. The Court thought it crucial to determine whether Judge Maloney had succumbed. Later we decided a case involving a different corrupt judge, Cartalino v. Washington, supra, in which the requisite proof was supplied: the bribery scheme included convicting Cartalino. There is no evidence that convicting Bracy and Collins was part of Maloney’s bribery scheming.
If the mere possibility of compensatory bias were enough to establish actual bias, all decisions by a judge who accepted bribes would be invalidated — in the case of Judge Maloney, literally thousands. That is another distinction between compensatory bias and a financial stake (or family relationship). A financial stake is case specific. The temptation it offers the judge is limited to the case in which he has a stake. His other cases are unaffected. But the theory of compensatory bias implies that all the judge’s decisions in criminal cases are fatally contaminated — the cases in which he was bribed, of course, but also the cases in which he was not bribed; and so — all his cases. The Supreme Court did not adopt and would not countenance a rule that compensatory bias can be presumed from the fact that a judge has accepted bribes in some cases. Judge Rovner’s opinion in the present round disregards the Supreme Court’s mandate. She repeats the position she took in the original appeal — the position the Court disapproved — that all of Malo-ney’s convictions (and presumably those of any other bribe-taking judge) must be set aside and that case-specific evidence of compensatory bias is always unnecessary, and indeed irrelevant. The Court made unmistakably clear that compensatory bias *422must be proved to have been operative in the particular defendant’s case. Proof of this is not impossible, as Cartalino illustrates. Bias could also be inferred, much as discrimination often is inferred, from a pattern of rulings that could not be satisfactorily explained on any hypothesis other than that of compensatory bias. The evidence need not always be case specific. Maloney was deposed as part of the discovery conducted on remand. Had he testified that he had practiced compensatory bias in all the cases in which he had not been bribed, and his testimony had been believed, or if evidence had been presented of a conspiracy to practice compensatory bias in every case in which no bribe was offered to the judge, an absence of evidence about the motive for his rulings in the trial of particular defendants who had not offered bribes would not be fatal. (It wouldn’t even matter if he didn’t remember the trial at all.) All that had to be established in the remand proceeding that the Supreme Court ordered, in order to justify ordering a new trial for Bracy, was a factual basis for inferring that Maloney probably did harbor an actual bias against him.
That could not be inferred, however, from the fact that Maloney took bribes or even from the fact, if it was a fact, that he practiced compensatory bias, for he may not have done so in every case. We do not know whether he practiced it in any case; and he would have been unlikely to practice it in every case. If he thought that a defendant was certain to be convicted and receive a severe sentence, he would have no incentive to lean in favor of the prosecution and by doing so jeopardize the conviction or sentence by making it more vulnerable to reversal on appeal. In general a corrupt criminal judge has no need to lean against criminal defendants who have not bribed him, because most criminal defendants are guilty and will be convicted anyway.
The discovery ordered by the Supreme Court drew a blank. Much of it consisted of a wild goose chase after McDonnell’s relationship to Maloney. The chase did uncover ugly evidence of criminality and mob ties of both McDonnell and Maloney, but nothing that bore on the issue of compensatory bias — except to dispel the suspicion that Maloney had appointed McDonnell to make sure that Bracy would be convicted, or that McDonnell had tried to throw the case in order to curry favor with Maloney. The judge found that McDonnell had never practiced law with Maloney and had pulled no punches in his defense of Bracy. This finding is not clearly erroneous, and so it binds this court and wipes out the theory of bias that was the focus of the Supreme Court’s discussion of the need for discovery.
It is true that during his allocution before being sentenced Maloney had spoken of the convictions and sentences of Bracy and Collins as “a credit to his record as a judge and evidence that he was not corrupt,” 79 F.Supp.2d at 907, and that this led the district judge to find (id. at 908) that
•during the same time petitioners’ case was pending, other cases were pending in which Maloney took bribes, particularly the close in time Chow and Rosario cases. Before and after this time, Malo-ney was engaged in a pattern of receiving money. Based on the evidence in the record, it is a possible and reasonable inference in this case that Thomas Maloney was motivated, at least in part, to maintain a prosecution-oriented attitude and to make pro-prosecution rulings by a desire to deflect suspicion from cases in which he accepted bribes. Other documented instances of Maloney so acting to deflect suspicion from his cor-*423nipt conduct are reported in the Hawkins and Titone cases.
This is naked conjecture, however, and so cannot be the basis of a valid factfinding. Libman Co. v. Vining Industries, Inc., 69 F.3d 1360, 1363 (7th Cir.1995); United States v. Givens, 88 F.3d 608, 613 (8th Cir.1996); Thompson v. Washington, 266 F.2d 147, 148-49 (4th Cir.1959) (per curiam); In re Kuttler’s Estate, 185 Cal.App.2d 189, 8 Cal.Rptr. 160, 169 (Cal.App.1960) (“an inference may not be based on suspicion alone, or on imagination, speculation, supposition, surmise, conjecture, or guess work.... A finding of fact must be an inference drawn from evidence rather than ... a mere speculation as to probabilities without evidence”). It was natural for Maloney, at his sentencing for accepting bribes from criminal defendants, including defendants in murder cases, to point to a case before him in which the murderers had been convicted and sentenced to death, though the jury, not he, had convicted them and had made a recommendation for death that bound him (“recommendation” is thus a misnomer). It does not follow that when he presided at trial he was thinking of how the defendants’ convictions and sentences might stave off future accusations of bribe taking, or even how they might dispel suspicions of it — if he was even aware at that time, early in his bribe taking career, that there were any suspicions; probably he was not, or he would not have continued taking bribes for nine more years. The two cases the district judge gave as examples of Maloney’s “acting to deflect suspicion from his corrupt conduct” are cases in which Maloney accepted bribes; in one he returned the bribe because he realized that he was under investigation and in the other he convicted the defendant anyway. Neither case had anything to do with compensatory bias. He returned the bribe five years after the trial of Bracy and Collins; there is no indication that he was, or thought he was, under suspicion at the time of that trial.
The district judge based his conclusion about Maloney’s motivation largely on the “Government’s Official Version of the Offense” submitted in Maloney’s criminal trial. This document, which the parties refer to as the sentencing recommendation or sentencing memorandum, is also the cornerstone of the appeal. In it the Justice Department accused Maloney (whom it called “degenerate” and “a mafia factotum”) of practicing compensatory bias. The document consists, however, of 57 single-spaced pages, and the allegation of compensatory bias appears on just one of them. It is colorful (“THOMAS MALO-NEY far surpassed the category of corrupt jurist to chart a new territory of defilement”), vivid, even plausible. But no substantiation or elaboration is offered. No cases in which Maloney may have engaged in compensatory bias are cited; no evidence, direct or circumstantial, admissible or inadmissible, that he ever engaged in the practice is offered. The Justice Department was pressing for a very long sentence (more than 20 years), and it pulled out all the stops.
Despite this “evidence” of compensatory bias, the district judge concluded that “the evidence does not establish that an interest in covering up wrongdoing or motivating larger bribe payments pervaded every action taken by Maloney as a judge. Malo-ney’s bribe taking has not been shown to have been so pervasive a part of his judicial practices that it can be assumed he was always, or even usually, motivated by his pecuniary and/or penal interests when exhibiting his prosecution-oriented tendencies.” Id. at 909 (emphasis added). This is an important finding, which not being clearly erroneous binds us. It requires (as the Supreme Court had already made *424clear) evidence that compensatory bias was at work in this case. It forbids us to rest on a presumption that compensatory bias was at work in every case in which a defendant tried before Judge Maloney was convicted.
In the light of this finding, the district judge as he was required to do examined Maloney’s rulings at the trial of Bracy and Collins and found none at the guilt phase of the trial that displayed bias. He concluded that the convictions were untainted. The conclusion is correct. For all that appears, Maloney was a prosecution minded judge for reasons unrelated to his taking bribes. That he would accept bribes to acquit criminals does not imply any affection for criminal defendants or their lawyers such that he must have been acting against character when he ruled in favor of the prosecution in cases in which he was not bribed. His conduct was appalling, his character depraved, but the bridge to the trial of Bracy and Collins is missing.
However, turning to Maloney’s rulings at the sentencing phase of the trial, the district judge found the taint of compensatory bias. The only ruling (or pair of rulings) he mentioned was Maloney’s refusal to sever Collins’s sentencing hearing from Bracy’s and hold it first in order to give Bracy’s lawyer more time to prepare for his client’s hearing. The ruling is said to have harmed Collins because it meant that the jury would hear evidence about additional murders that Bracy had committed in Arizona, murders in which Collins had not been implicated. (Bracy had not yet been convicted of the Arizona murders; later he was, and he was sentenced to death; that sentence is pending.)
Collins had not raised the issue of severance in his state-court appeal, and as a result it was treated as forfeited in the federal habeas corpus proceeding. It is not surprising that he didn’t raise the issue, because it is very difficult to see how he would have been harmed, rather than helped, by evidence that Bracy was a worse murderer than he. And so it is difficult to see how the ruling could be thought evidence of bias. But all this to one side, there is no basis for upholding Bracy’s and Collins’s convictions but setting aside their sentences. The incentive to engage in compensatory bias is stronger at the trial of guilt than at the sentencing hearing. Most criminal defendants are convicted, so a judge who wants a reputation as a tough sentencer, either to induce bribes or to avoid charges of undue lenity, will have an incentive to make rulings favorable to the prosecution, so that the defendant will not walk. Had these triple-murdering defendants been acquitted, eyebrows might have been raised. But the imposition of the death sentence is a matter of grace to be determined by the jury. Maloney would not have been “blamed” if the jury had exercised its unreviewable power of lenity and declined to recommend sentencing Bracy and Collins to death. For all we know, that is a common sequel to the conviction of defendants in capital eases in Illinois.
The only thing on which Judge Evans can hang the distinction between the sentencing phase and the guilt phase of the trial, moreover — the refusal to try Bracy and Collins separately — was before the Supreme Court when in remanding the case to us it made clear that Bracy and Collins could prevail only if discovery disclosed evidence of compensatory bias. It did not.
There is a lot of “death is different” talk in Judge Evans’s opinion. Maybe he wishes to suggest that compensatory bias has a different meaning in capital-sentencing than in other proceedings. That’s a position Bracy’s own lawyer rejected at the en banc argument. He was emphatic that compensatory bias if proved would *425invalidate a misdemeanor conviction or for that matter a judgment in a civil case; Aetna Life Ins. Co. v. Lavoie, 475 U.S. 813, 106 S.Ct. 1580, 89 L.Ed.2d 823 (1986), on which he heavily relied, was a civil case. He was right. A civil litigant and a misdemeanor defendant are entitled to an unbiased judge, just like a capital defendant. The Supreme Court’s elaborate jurisprudence on the death penalty does not include a special standard of judicial bias for capital cases only. When compensatory bias is shown, the losing party is entitled to relief regardless of the nature of the case.
The capital nature of this case is relevant only in the following very limited sense: a judge conceivably might be biased in one stage of a case but not all stages, so if there are severable stages, such as the guilt and sentencing phases of a capital case, bias at the last stage might not spill back into the earliest stage. If anything, as I have pointed out, Judge Maloney was more likely to be biased against defendants at the guilt stage of the proceeding than at the sentencing phase. There is nothing to suggest that he was indifferent to whether they were convicted but determined if they were convicted to see that they were executed. Nothing in the theory of compensatory bias or in the psychology of Maloney supports such a conjecture. Critically, there is no evidence to support it.
No evidence, but plenty of rhetoric. Judge Evans states: “It is more than a fair inference that increasing the likelihood of the imposition of the death penalty would be fine with Judge Maloney.” And: “Less concern about the fate of the defendants ... could hardly be imagined.” And: “Maloney was sublimely unconcerned about a lack of evidence in mitigation.” And: “It is pushing credibility to imagine that an experienced trial judge ... did not see this scene unfolding.” And: “Maloney [was] deliberately indifferent to the petitioners’ fates.” From this it is inferred that Maloney “deliberately let this death penalty hearing become a debacle because imposition of the death penalty on these two men would bolster his reputation as a tough judge.” But the judge does not impose the death penalty; the jury does. And of course Maloney might have wanted a reputation as a tough judge for reasons unrelated to compensatory bias. And he might not have been seeking a reputation as a tough judge — he may just have been disgusted by these defendants’ crimes, or he may have been a bad judge, or he may have thought capital punishment the right punishment for murderers, or he have been proprosecution on general principles, or all these things may have been true. In failing to canvass these possibilities, Judge Evans’s opinion reveals a lack of imagination. Furthermore, Maloney’s rulings at the guilt phase of the trial also consistently favored the prosecution, as Judge Evans’s opinion fails to make clear; it is no surprise that Malo-ney’s rulings at the sentencing hearing favored the prosecution as well. If Malo-ney was not biased in presiding at the guilt phase of the trial despite his consistent leaning in favor of the prosecution, how as a matter of logic and common sense can we have any confidence that he suddenly, inexplicably — indeed irrationally — became biased at the penalty phase?
The language that I have quoted from Judge Evans’s opinion really points in a different direction — toward a conclusion that Maloney created an “appearance of impropriety,” concretely that he gave the appearance of being determined to do in Bracy and Collins. Not only need such a determination have nothing to do with compensatory bias, but in an earlier en banc opinion of this court that Judge Evans’s opinion neglects to discuss we held *426that a judge’s mere- appearance of impropriety does not render a judgment in violation of due process. Del Vecchio v. Illinois Dept. of Corrections, supra, 31 F.3d at 1371-72; see also id. at 1389, where this ruling is elaborated. Appearances and suspicions are all that the court has going for it in this case.
The judgment of the district court should be affirmed insofar as it upheld the convictions but it should be reversed insofar as it invalidated the sentences.