SUTTON, J., delivered the opinion of the court, in which OBERDORFER, D.J., joined.
CLAY, J. (pp. 1189-93), delivered a separate concurring opinion.
OPINION
A jury convicted Earl Ray Lyell of assault with intent to commit murder, Mich. Comp. Laws § 750.83, and a judge sentenced him to thirty to sixty years in prison as an habitual offender, Mich. Comp. Laws § 28.1083. He now petitions for a writ of habeas corpus, claiming that the trial court (1) coerced the jury into reaching a guilty verdict by improperly polling the jury and (2) exhibited bias and partiality that denied him a fair trial. Although we reject Lyell’s claim that the jury polling violated his constitutional rights, we agree that the trial judge’s conduct violated his due-process right to a fair trial. We reverse and conditionally grant the writ.
I.
The facts giving rise to Lyell’s arrest and trial are straightforward. According to the prosecution, Lyell repeatedly stabbed Anthony Nimeth after learning that Nimeth intended to tell the authorities about Lyell’s involvement in a high-speed car chase with the police. According to Lyell, Nimeth tried to rob him at knife point, leaving Lyell no choice but to stab Nimeth in self defense.
The trial was anything but straightforward. Throughout the proceedings, the trial judge and Lyell’s defense counsel, Hart, clashed verbally, frequently in the presence of the jury. See People v. Lyell, No. 214100, slip op. at 3, 2001 WL 671474 (Mich.Ct.App. May 4, 2001). The clashes began in voir dire when the judge interrupted Hart’s questioning of prospective jurors on several occasions (without prior objection from the prosecution) in order to accuse him of being repetitive, see, e.g., Tr. Day 1 at 61, 109, or argumentative, see, e.g., id. at 63, 96, 99, and they continued through Hart’s direct and cross-examination of witnesses. At the same time, the judge rarely interrupted the prosecutor; and when she did interrupt him (at least on two occasions), she did so in order to assist him. See Tr. Day 2 at 68; Tr. Day 5 at 28. Our review of the record reveals that over the course of the six-day trial, the judge interrupted Lyell’s counsel— *1180without prompting or objection from the prosecution — roughly 40 times.
On three occasions the court assumed control of witness questioning in a manner suggesting that the judge favored the prosecution’s case. Tr. Day 2 at 68 (urging the prosecutor to ask a question even though the prosecutor believed it called for hearsay), 154 (interrupting Hart’s attempt to impeach the witness with previous statements made to police and sua sponte eliciting details — via 12 separate questions — not revealed on direct); Tr. Day 5 at 28 (urging the witness to answer a question voluntarily withdrawn by the prosecutor). The following exchange is illustrative:
THE COURT: Is there any reason why you don’t ask [the witness] what [another witness, Miss Reiland,] said to her?
MR. WENZEL[the prosecutor]: Because technically it is hearsay. THE
COURT: It is admissible.
MR. HART: Judge, with all due respect, I would rather just fight Mr. Wenzel and not—
THE COURT: You know what, you’re not acting like[ ] a lawyer. We are talking about — at least it has been established that this is an exciting event, and it makes a whole lot more sense if the witness tells us what was said to her. Now, don’t object anymore, Mr. Hart, when, things are so obvious. Now, would you please ask her what Miss Reiland said.
Tr. Day 2 at 68.
During Hart’s cross-examination of Ni-meth, the man stabbed by Lyell, the court became particularly active, interrupting the cross-examination — without a prior objection from the prosecution — 18 times. Id. at 130 (2 interruptions), 132 (2 interruptions), 133, 134, 137, 140, 141, 142, 143, 146, 162, 169, 170, 189; Tr. Day 3 at 22, 38, 41, 43. At least 14 of these interruptions occurred in less than an hour. See Tr. Day 2 at 178. The interruptions often contained implications that Hart’s attempts to discredit Nimeth’s character for truthfulness were not relevant to the case, because Nimeth’s proclivity for lying to the police was “not the issue in this case.” Id. at 170; id. (“[N]o issues have been raised as far as I have been able to determine ....”); see id. at 130, 131, 132, 133, 142, 143. When Hart persisted in this line of questioning, the trial judge interjected: ‘What does that have to do with this? I don’t understand the point you’re making.” Id. at 149. Hart explained that he intended to use the questions to discredit the prosecution’s theory of motive, to which the judge responded, “I guess I just don’t get it.” Id. at 150.
The judge’s repeated interruptions of Hart’s questioning often came in the form of insults directed at Hart. For instance, she told him, “You want to be an actor. Be a lawyer.” Id. at 131. Shortly thereafter, she added: “Don’t act like a child, Mr. Hart. You’re a lawyer,” id. at 133, and “Would you please position yourself and act like [a lawyer],” id. at 134. She also accused him of being “a smart aleck,” id. at 141, of being “silly,” id. at 147, and of “trying to create a furor,” id. at 141. When Hart appeared (to the judge’s mind) to be investigating a forbidden line of questioning, the following exchange resulted:
THE COURT: Mr. Hart, you know you’re exhausting all of us. Mr. Hart, do you have any more questions for this witness before he is excused?
MR. HART: Yes, I do, Judge.
THE COURT: I don’t know why you keep doing these things over, and over again. That was a terrible thing, terrible thing for you to do.
MR. HART: I disagree.
*1181THE COURT: Doesn’t make any difference whether you agree or not.
Tr. Day 3 at 22.
The conflict between the judge and Hart culminated when — once again in the presence of the jury — the judge held Hart in contempt, fined him $250 and commented that “[tjhere are some of these people who have never heard lawyers, who have never been in a courtroom before, it is embarrassing to all of us to have you act in this fashion.” Id. at 27. Six days after the cross-examination of Nimeth and the contempt ruling, the judge instructed the jury that she and Hart did not “bear each other any animus,” that the verdict must be based solely on the evidence, that Hart was not “a fact in [the] case” and that finding Hart in contempt should not be considered by the jury when “making a determination about the facts in [the] case.” Tr. Day 6 at 19-20.
The jury returned a guilty verdict. After the jury announced its verdict, Lyell’s counsel asked the court to poll the 14 jurors. During the polling, the first 11 jurors all concurred in the verdict. But the 12th juror apparently changed her mind, refusing to concur in the verdict:
THE CLERK: [Juror] was that and is this your verdict?
JUROR NO. 12: No. I am sorry, Judge.
THE COURT: Don’t talk anymore. Let me just say this to you. May I ask the remaining two ... jurors, was that and is that your verdict?
JUROR NO. 13: Yes.
JUROR NO. 14: Yes.
THE COURT: It is not possible for me to talk to you any further. But I really would ask you to go back and, you know, discuss with each other where you are and what processes you’re involved in to see if you can arrive at a verdict. I don’t urge anyone to give up their ideas or their thoughts, but I do think it is very important to, you know, talk with each other and to see what it is that you disagree upon. If you would be kind enough to do that, I would be appreciative.
Id. at 38-39. Lyell’s counsel did not object when the court polled the last two jurors. But after the court gave this charge, Lyell moved for a mistrial. Id. at 39-40. The court did not rule on the motion, and when the jurors returned about an hour later they announced a guilty verdict. Id. at 40-41.
On direct review, a divided panel of the Michigan Court of Appeals, relying on state law, concluded that “the circumstances surrounding the [jury] polling did not tend to coerce the jurors to reach a particular verdict.” People v. Lyell, No. 214100, slip op. at 3. As to the issue of judicial bias, a divided panel (again relying on state law) concluded that “the judge acted within her power and discretion to control the trial” and that her “conduct did not demonstrate partiality that influenced the jury’s verdict.” Id. at 4. In reviewing Lyell’s habeas petition, the district court rejected both claims.
II.
Under the Antiterrorism and Effective Death Penalty Act of 1996 (AED-PA), we may grant the writ as a general rule only if the state court’s decision on the merits was either “contrary to, or involved an unreasonable application of, clearly established Federal law, as determined by the Supreme Court of the United States.” 28 U.S.C. § 2254(d)(1); see Williams v. Taylor, 529 U.S. 362, 405-06, 120 S.Ct. 1495, 146 L.Ed.2d 389 (2000). The statute, however, acknowledges an exception. “By its very language, 28 U.S.C. § 2254(d) is applicable only to habeas claims that were adjudicated on the merits in state court.” *1182 Newton v. Million, 349 F.3d 873, 878 (6th Cir.2003) (internal quotation marks omitted); see 28 U.S.C. § 2254(d) (providing that heightened deference applies “to any claim that was adjudicated on the merits in State court proceedings”). If “the state court did not assess the merits of a claim properly raised in a habeas petition, the deference due under AEDPA does not apply,” Maples v. Stegall, 340 F.3d 433, 436 (6th Cir.2003), and we “conduct [our] review de novo,” McAdoo v. Elo, 365 F.3d 487, 498 (6th Cir.2004).
In this instance, no one disputes that Lyell presented federal polling and fair-trial claims to the state court of appeals. Perhaps because Lyell also raised state-law challenges to his conviction, however, the state court of appeals addressed Lyell’s claims only in state-law terms in its decision. In the absence of a ruling on the merits of these federal-law claims, we must give fresh review to Lyell’s polling and fair-trial claims under the United States Constitution. See Danner v. Motley, 448 F.3d 372, 376 (6th Cir.2006).
A.
The due-process “principle that jurors may not be coerced into surrendering views conscientiously held is so clear as to require no elaboration.” Jenkins v. United States, 380 U.S. 445, 446, 85 S.Ct. 1059, 13 L.Ed.2d 957 (1965) (per curiam) (internal quotation marks omitted). The most frequent setting in which the risk of juror coercion arises is when a court inquires into the numerical division of a deadlocked jury. In Brasfield v. United States, 272 U.S. 448, 47 S.Ct. 135, 71 L.Ed. 345 (1926), the trial judge recalled the jury after it had deliberated for several hours and asked the jury what its numerical division was. The Supreme Court reversed the federal conviction that followed this question, noting that the practice was “harmful” and “not to be sanctioned.” Id. at 450, 47 S.Ct. 135. The procedure, the Court reasoned, “serves no useful purpose that cannot be attained by questions not requiring the jury to reveal the nature or extent of its division. Its effect upon a divided jury will often depend upon circumstances which cannot properly be known to the trial judge or to the appellate courts and may vary widely in different situations, but in general its tendency is coercive.” Id. at 450, 47 S.Ct. 135.
The issue under Brasfield is whether the trial court inquired into, and learned, the numerical division of a deadlocked jury. See Lowenfield v. Phelps, 484 U.S. 231, 239, 108 S.Ct. 546, 98 L.Ed.2d 568 (1988) (“[T]he trial court had, after deliberations stalled, inquired as to how the jury was divided, and was informed simply that the jury stood nine to three.”). Once it is established that this question was asked and answered, the courts agree that per se error has occurred. Id. (“This Court concluded that the inquiry into the jury’s numerical division necessitated reversal because it was generally coercive and almost always brought to bear in some degree, serious although not measurable, an improper influence upon the jury.”) (internal quotation marks omitted); Williams v. Parke, 741 F.2d 847, 851 (6th Cir.1984) (acknowledging the existence of “this per se rule”); see United States v. Lash, 937 F.2d 1077, 1085 (6th Cir.1991) (“In Brasfield v. United States, the Supreme Court held that a judge’s inquiry into how the jury was divided was coercive and required reversal.”); Camel v. Sowders, No. 90-6083, 1991 WL 225563, at *2-3, 1991 U.S.App. LEXIS 26537, at *7 (6th Cir. Nov. 4, 1991) (“[I]t is reversible error for a federal trial judge to give an Allen charge after having inquired as to the division of the jury.”).
*1183Lyell faces four problems in claiming that Brasfield requires us to grant him habeas relief. First, since Brasfield, the Supreme Court has recognized that the decision involved an exercise of its supervisory power over the federal courts, not an exercise of its authority to construe the Due Process Clause. Lowenfield, 484 U.S. at 239-40, 108 S.Ct. 546 & n. 3. Every court of appeals to address the issue, including ours, has declined to extend Bras-field to habeas claims premised on a violation of the federal Constitution. See Williams, 741 F.2d at 851; see also Montoya v. Scott, 65 F.3d 405, 412 (5th Cir.1995); Locks v. Sumner, 703 F.2d 403, 406 (9th Cir.1983); United States ex rel. Kirk v. Director, Dep’t of Corr., 678 F.2d 723, 727 (7th Cir.1982); Cornell v. Iowa, 628 F.2d 1044, 1048 (8th Cir.1980); Ellis v. Reed, 596 F.2d 1195, 1200 (4th Cir.1979).
Lowenfield, it is true, leaves open the possibility that some combination of an improper inquiry into the numerical division of a deadlocked jury and of an improper Allen charge “might” violate due process. See Lowenfield, 484 U.S. at 241, 108 S.Ct. 546. But no court to our knowledge has taken the Court up on this possibility, and Lyell has offered no good reason to apply Lowenfield’s might-be-a-violation dictum to this ease.
Second, Brasfield and all of the other cases upon which Lyell relies stem from judicial inquiries into the numerical division of a deadlocked jury. Yet there is a world of difference between juror-coercion claims arising from deadlocked juries and those arising from post-verdict juror polling. In the former situation, there is never any reason to expose the numerical division of the jurors. The trial court may decide to give a seemingly deadlocked jury an Allen charge to urge it to continue its deliberations in good faith, but the court has no reason to ask, or find out, which jurors stand where on the charges.
The same is not true with juror polling. There, it is not only necessary but desired (at least from the defendant’s perspective) for the public to learn that at least one juror has opted to take a stand against conviction. Consider what happened in this case and what normally happens in this setting. The jury announced to the court that it had reached a verdict, which required unanimity among the jurors. The jurors returned to the court room, and the foreman read the verdict, stating that the defendant was guilty as charged. At that point, defendant’s counsel, not the court or the prosecution, asked the court to poll the jurors. If there was any tendency to coerce at this point, it was the tendency for jurors to say that what they had just agreed to unanimously in the deliberation room was still their vote in the case. And unlike a deadlocked jury, it was utterly unavoidable that if a juror did have a change of heart, he or she would have to go public with it. That is what happened here when juror 12 explained that she did not support the verdict. At this point, we acknowledge, there seemed to be little point to continuing to poll the last two jurors (save with defense counsel’s consent) because one holdout suffices to send the jury back to deliberate. But there is nothing about the judge’s actions that suggests that this was anything more than a slip in inertia after polling the first 12 jurors. Nor, it bears adding, did defense counsel do anything to stop the judge, which itself suggests a non-coercive environment.
The possibility of unconstitutional coercion under these circumstances seems quite unlikely. Any coercion could not stem from juror 12’s public statement of her position, which was prompted by defense counsel’s request for a poll. And *1184coercion could not flow from the juxtaposition of juror 12’s public statement with the first 11 jurors’ public statements of their position, which also was a necessary consequence of the poll. The risk of coercion, if any, stems from the unsurprising fact that the last two jurors wanted to convict— unsurprising because most jurors presumably do not change their minds minutes after making a decision. The prospect that juror 12, who retained the courage of her convictions to tell the public that she now wanted to change her vote, would be brow-beaten into submitting to the majority simply by the fact that two other jurors (and eleven before that) did not change their minds seems quite slim.
That is all the more true given that the trial judge, immediately after hearing the results of the poll, asked the jury to continue its deliberations on any disputed issues and permissibly instructed them in doing so. She urged each of them not to “give up their ideas or thoughts,” and she told them that the jurors retained the right to disagree with one another. Tr. Day 6 at 38-39. At no point did the judge require the jury to reach a verdict. Cf. Jenkins, 380 U.S. at 446, 85 S.Ct. 1059. After the court sent the jury back to deliberate, it took an hour for the jury ultimately to reach a unanimous verdict, further suggesting that when juror 12 changed her mind (yet again), coercion was not the cause.
Third, Rule 31(d) of the Federal Rules of Criminal Procedure confirms that jury polling differs considerably from inquiries into the numerical division of a deadlocked jury. Rule 31(d) provides:
Jury Poll. After a verdict is returned but before the jury is discharged, the court must on a party’s request, or may on its own, poll the jurors individually. If the poll reveals a lack of unanimity, the court may direct the jury to deliberate further or may declare a mistrial and discharge the jury.
Unlike an inquiry into the numerical division of a deadlocked jury, sanctioned by no rule or practice, Rule 31(d) allows jury polling with no express limitation and thus permits an inquiry into at least an initial numerical breakdown of the jury. Indeed, the whole premise of the rule would seem to be to permit parties, particularly criminal defendants, to ferret out juror coercion and prevent it. On top of that, because Rule 31(d) permits a trial court to direct the jury to continue deliberations or to discharge the jury when the poll shows an absence of unanimous concurrence, it is not self-evident that polling of the whole jury, even after one juror dissents, dis-serves the defendant’s interest. Presumably the odds that a trial court will discharge a jury go up when there are two dissenters rather than one.
Fourth, the federal case law involving jury polling that continues after one juror dissents from the verdict strongly supports the conclusion that a due process violation did not occur here. To our knowledge, five federal circuit courts have considered this question. Four courts of appeals upheld the verdict. See United States v. Gambino, 951 F.2d 498 (2d Cir.1991); United States v. Fiorilla, 850 F.2d 172 (3d Cir.1988); Amos v. United States, 496 F.2d 1269 (8th Cir.1974); United States v. Brooks, 420 F.2d 1350 (D.C.Cir.1969). One court of appeals did not. See United States v. Spitz, 696 F.2d 916 (11th Cir.1983) (per curiam).
The problem with Spitz, the decision that found a due process violation, is that it applied Brasfield’s deadlocked-jury rule without considering the differences between the two scenarios. Id. at 917-18. As a result, with little explanation, it applied Brasfield’s “per se error” rule to jury *1185polling that continued after one juror dissented. Id. at 917.
The other appellate decisions, by contrast, acknowledge the Brasfield rule, then explain why its inflexibility does not apply to polling claims. See Gambino, 951 F.2d at 501 (noting that “the weight of authority suggests that when the trial judge continues to poll the jury after one juror disagrees with the verdict, reversible error occurs only when it is apparent that the judge coerced the jurors into prematurely rendering a decision, and not merely because the judge continued to poll the jury”); Fiorilla, 850 F.2d at 175-76 (declining to apply Brasfield’s “per se reversal” rule and favoring “a rule vesting discretion in the trial court”); Amos, 496 F.2d at 1273 (noting a “distinction” between Brasfield, where the court’s inquiry into the numerical division “will be deemed coercive,” and jury polling, where “Rule 31(d) gives the court some latitude in polling of the jury to clear up an apparent confusion”); Brooks, 420 F.2d at 1354 (noting that “a jury poll ... stands on quite a different footing” from Brasfield).
All four courts left it to the discretion of the trial court under the facts and circumstances of the case to continue the poll once it became apparent that there was a recalcitrant juror. See Gambino, 951 F.2d at 502 (noting that whether a jury poll is “coercive must be evaluated on the facts and circumstances of the particular case and not simply because the trial judge continued to poll the jury after a juror dissented from the reported verdict” and holding that the district court did not abuse its “discretion” in continuing the poll); Fiorilla, 850 F.2d at 174 (finding “a lack of coercion ... demonstrated by this record” and accordingly “finding] no abuse of discretion”); Amos, 496 F.2d at 1273 (“pay[ing] due deference” to district court); Brooks, 420 F.2d at 1354 (noting that Rule 31(d) “invests the trial judge with a measure of discretion in assessing the impact of a dissenting vote during a jury poll, and the reasonable exercise of this discretion should be accorded proper deference by a reviewing court”).
The same factors that these four courts considered in determining that no coercion occurred prompt a similar conclusion here. Defense counsel in this case did not object to the continued polling. See Gambino, 951 F.2d at 501-02; Fiorilla, 850 F.2d at 176; Amos, 496 F.2d at 1273 (“This lack of objection by counsel permits an inference that the procedures utilized did not appear coercive at the time.”); Brooks, 420 F.2d at 1354. The trial judge gave an appropriate cautionary instruction before asking the jury to continue deliberating. See Gambino, 951 F.2d at 502; Fiorilla, 850 F.2d at 177. And the jury did not return a verdict immediately after renewing deliberations but took an hour to reach a verdict. Gambino, 951 F.2d at 502 (noting that Amos and Brooks took 20 minutes to return a verdict after the renewed deliberations). As in these cases, no unconstitutional coercion of juror 12 occurred.
Judge Clay’s concurrence warrants two brief responses. The contention that the jury polling in this case violated state law — either Michigan Court Rules, see M.C.R. 6.420(D), or Michigan case law — is of no moment because we may not grant federal habeas relief based on violations of state law. Estelle v. McGuire, 502 U.S. 62, 67, 112 S.Ct. 475, 116 L.Ed.2d 385 (1991). Even if we were to consider state-law violations, moreover, we would have to contend with the Michigan courts’ rulings, which considered — and rejected — these state-law arguments on direct review in determining that no reversible error occurred. See People v. Lyell, No. 214100, 2001 WL 671474, at * 1 (Mich.App. May 4, 2001).
*1186Nor may we shore up the deficiencies in Lyell’s jury-polling argument by looking to the strength of his second claim — that the trial judge’s interaction with defense counsel deprived Lyell of a fair trial. For one, neither Lyell in his pro se brief nor his counsel raised this argument. They both maintained that the Brasfield per se rule applies whenever there is an inquiry into the numerical division of the jury, regardless of the circumstances. That contention, as we have shown, is simply mistaken. For another, the trial court’s mistreatment of defense counsel did not establish a claim of unconstitutional juror coercion. The jury polling occurred soon after the trial court’s instruction that any differences between her and defense counsel had no place in the jury’s deliberations and occurred well after her contempt ruling six days earlier. The polling occurred because defense counsel asked for it and received it, not because he sought something and the trial court rejected it — the unfortunate pattern that had occurred earlier in the trial. The circumstances of the jury polling thus represent a far cry from the judge’s intemperate interactions with defense counsel when she obstructed efforts of defense counsel to try his case. That defense counsel, who was hardly reticent when it came to objections during the trial, did not object to the polling of the thirteenth and fourteenth jurors also suggests that the polling of the last two jurors was at worst inadvertent and at best done with the acquiescence of defense counsel. Notably, when defense counsel moved for a mistrial immediately after the jury returned to its deliberations, he did not make the motion on the ground that the court should not have polled the last two jurors — which is understandable, given that he had just watched the court do that very thing without raising an objection.
B.
Lyell next argues that the trial judge violated another guarantee of the Due Process Clause of the Fourteenth Amendment — the right to a “fair trial in a fair tribunal before a judge with no actual bias against the defendant.” Bracy v. Gramley, 520 U.S. 899, 904-05, 117 S.Ct. 1793, 138 L.Ed.2d 97 (1997) (internal quotation marks omitted); see In re Murchison, 349 U.S. 133, 136, 75 S.Ct. 623, 99 L.Ed. 942 (1955) (“A fair trial in a fair tribunal is a basic requirement of due process.”). While Liteky v. United States, 510 U.S. 540, 114 S.Ct. 1147, 127 L.Ed.2d 474 (1994), addresses the statutory recusal standards for federal judges, see 28 U.S.C. § 455, this court has relied on the decision in assessing judicial-bias claims under the Due Process Clause, see Alley v. Bell, 307 F.3d 380, 386 (6th Cir.2002) (stating that “[tjhis court has looked to the Supreme Court’s decision in Liteky v. United States to provide the standard for deciding judicial bias claims” under the Due Process Clause). Even though Liteky involved statutory, as opposed to constitutional, interpretation, the parties have not objected to the district court’s application of that standard here, have not objected to our circuit’s prior application of that standard in this setting and have not offered any suggestion that the Liteky standard differs from the Court’s other due-process precedents in this area.
Under Liteky, a judge’s misconduct at trial may be “characterized as bias or prejudice” only if “it is so extreme as to display clear inability to render fair judgment,” Liteky, 510 U.S. at 551, 114 S.Ct. 1147 (internal quotation marks omitted), so extreme in other words that it “display[s] a deep-seated favoritism or antagonism that would make fair judgment impossible,” id. at 555, 114 S.Ct. 1147. “[JJudicial remarks during the course of a trial that are critical *1187or disapproving of, or even hostile to, counsel, the parties, or their cases ordinarily do not support a bias or partiality challenge____[But] they will do so if they reveal such a high degree of favoritism or antagonism as to make fair judgment impossible.” Id. (emphasis omitted). “[Expressions of impatience, dissatisfaction, annoyance, and even anger, that are within the bounds of what imperfect men and women ... sometimes display,” by contrast, do not establish such bias or partiality. Id. at 555-56, 114 S.Ct. 1147; see also In re Murchison, 349 U.S. at 136, 75 S.Ct. 623; Offutt v. United States, 348 U.S. 11, 17, 75 S.Ct. 11, 99 L.Ed. 11 (1954).
Difficult as this standard may be to reach, the trial court seemingly made every effort to satisfy it. As our prior discussion of the trial reveals, the trial judge took over the cross-examination of the central witness in the case (Nimeth) and elicited information not revealed on direct examination. Tr. Day 2 at 152-55. The trial judge rarely waited for the prosecution to object before limiting questioning; she instead chose to limit questioning on her own throughout the trial. The concentration of interruptions during Nimeth’s cross-examination, the derogatory tone and content of many of the interruptions (throughout the trial) and the implicit disapproval of defense counsel’s theory of the case through these interruptions all put Lyell at a unique disadvantage in trying to encourage the jury to see the case through his eyes.
Making matters worse, the trial judge’s interruptions ran in one direction. While the trial judge frequently interrupted Lyell’s presentation of his case in an unhelpful way, she rarely interrupted the prosecution’s presentation of the case, save when doing so helped the government. At one point the judge urged the prosecutor to ask a question even after the prosecutor explained that it would elicit inadmissible hearsay, and at another point she sought an answer to a question that the prosecutor had voluntarily withdrawn.
Altogether, then, we have a case in which the judge sua sponte interrupted the prosecution to assist it, sua sponte interrupted Hart’s questioning in a way that undermined his presentation of the case (frequently during the cross-examination of the central witness in the case), failed to interrupt in a like manner during the prosecution’s questioning (at least in a way that undermined its case), stated or implied her disapproval of Lyell’s theory of the case (evidenced by her statements to the effect that Nimeth’s proclivity for lying to police was not an issue in the case or that she “didn’t get” the point of Hart’s motive-questioning) and made clear her disapproval of Lyell’s defense counsel (calling him an actor, a child, silly and a smart aleck). Capping all of this off was the trial judge’s inexplicable decision to issue a contempt order against Lyell’s counsel in front of the jury. Cf. United States v. Kelley, 314 F.2d 461, 463 (6th Cir.1963). On this record, the trial judge’s actions— considered in the context of the entire trial — made a fair trial impossible.
The State’s principal response to all of this is not to defend the trial judge; it is to say that however poorly she presided over the trial, Lyell still received a constitutionally fair trial. Two decisions from our circuit suggest otherwise, one arising in a criminal context, one arising in a civil context. In the criminal case, United States v. Hickman, 592 F.2d 931 (6th Cir.1979), we concluded that the trial judge’s “constant interruptions[,] which frustrated the defense at every turn,” denied the defendant a fair trial, id. at 936. After reviewing the record, “[t]he only impression which could have been left in the mind of the jury was that the trial judge was a *1188surrogate prosecutor.” Id. During the trial, the judge interjected himself into the proceedings more than 250 times. Id. at 932. The judge “[s]ua sponte interrupt[ed][ ] witnesses] or counsel ... to state why the witness’ particular testimony was in some way objectionable.” Id. at 934. At one point the trial judge took over the cross-examination of a witness and elicited information not addressed during direct examination. Id. at 935. The judge also interrupted during counsel’s opening statement and later expressed his “contemptuous disbelief’ about the testimony of the defense’s final witness. Id. at 936. This behavior led us to caution that a trial court’s “interference with the presentations of counsel has the potential of making a mockery of a defendant’s right to a fair trial, even in the absence of open hostility,” and to recognize “that undue interference with cross examination rights will result if a judge takes over examination by counsel.” Id. at 934.
In the civil case, Nationwide Mutual Fire Insurance Co. v. Ford Motor Co., 174 F.3d 801 (6th Cir.1999), we held that the trial court’s “constant interruption of ... direct examination ... and the intemperate tone and content of the court’s often argumentative questions” denied the plaintiff a fair trial, id. at 808. The judge, we noted, interrupted plaintiffs counsel’s opening statement six times, took over questioning of witnesses, admonished witnesses for not being brief enough in their responses and suggested to defense counsel when to make objections. Id. at 805-08.
Lyell’s trial judge, it seems to us, borrowed a page from the play book of the trial judges in both cases. As in Hickman and Nationwide, the judge sustained objections before they were made; the judge constantly interrupted the examinations of witnesses and did so almost exclusively with respect to one side of the case; the trial involved fairly straightforward issues that did not demand this kind of supervision; and the witnesses at the time were not being unresponsive to the prosecution’s questioning.
Brandt v. Curtis, 138 Fed.Appx. 734 (6th Cir.2005), an unpublished opinion that (unlike this case) involved deferential AEDPA review, does not suggest a different outcome. It involved two alleged sources of judicial bias: some “imprecise!, but] clearly not prejudicial,” pretrial comments made by the judge and the judge’s interruption of defense counsel “at least eight times during ... questioning of various witnesses.” Id. at 742 (internal quotation marks omitted). Although the judge did not make all of the interruptions sua sponte, we nonetheless characterized them as “troubling,” particularly in view of the negative and harsh tone used by the judge. Id. When juxtaposed with the 18 interruptions during one cross-examination (and roughly 40 interruptions overall) in this case, the numerous comments implying disapproval of Hart’s theory of the defense and the insults directed at Hart, the trial court’s conduct in Brandt offers a poor analogy.
Nor do McBee v. Grant, 763 F.2d 811 (6th Cir.1985), and Cox v. Treadway, 75 F.3d 230 (6th Cir.1996), both of which declined to find cognizable judicial bias, lead to a different conclusion. McBee involved a “single comment by the trial judge” and did not involve “a pattern of pervasive interference in the proceedings.” 763 F.2d at 818. And Cox involved just three interruptions of plaintiffs counsel’s opening statement. 75 F.3d at 237.
Other decisions from our court do not offer a handhold for the prosecution’s argument either. While United States v. Tilton, 714 F.2d 642 (6th Cir.1983) (per curiam), involved 28 interruptions as well *1189as active questioning of witnesses by the trial judge, “the interruptions appeared to go in both directions, affecting the defendant and the prosecution,” id. at 644. United States v. Smith, 561 F.2d 8 (6th Cir.1977), reached a similar conclusion for a similar reason, noting that “there was no strategic focusing of judicial questioning to favor the Government,” id. at 14. And in Maiurino v. Johnson, 210 F.3d 638 (6th Cir.2000), “[m]ost, if not all, of the disagreements between defense counsel and the court occurred outside the presence of the jury,” id. at 645.
The State also places considerable weight on the curative instruction issued by the judge and on the well-accepted presumption that juries may be presumed to follow all jury instructions. For one, the critical point of the instruction was to respond to her intemperate decision to issue a contempt order against Lyell’s counsel in front of the jury; it did little, if anything, to respond to, or account for, her other one-sided interruptions in the case. For another, the instruction came six days after the contempt order, when the damage from the contempt order in combination with her decision to play prosecutor had already been done. See McMillan v. Castro, 405 F.3d 405, 412 (6th Cir.2005) (noting the importance of issuing curative instructions as soon as possible after the occurrence of “conduct that could be viewed as ... biased”). Under these circumstances, the instruction cannot carry the weight that the State places on it. “[S]ome comments may be so highly prejudicial that even a strong admonition by the judge to the jury, that they are not bound by the judge’s views, will not cure the error.” Bursten v. United States, 395 F.2d 976, 983 (5th Cir.1968); see Quercia v. United States, 289 U.S. 466, 471-72, 53 S.Ct. 698, 77 L.Ed. 1321 (1933) (holding that an instruction to the jury that the judge’s opinion of the evidence was not binding on them was not sufficient to cure the prejudice caused by the judge’s statement that people who wipe their hands while testifying are liars); Frantz v. United States, 62 F.2d 737, 740 (6th Cir.1933) (instruction to the jury “was ineffective to correct the mischief already done” by the trial judge’s behavior); United States v. Chibbaro, 361 F.2d 365, 378-79 (3d Cir.1966) (holding that because the identity of the scarred perpetrator was contested, the judge’s remark upon seeing the unscarred defendant that “[s]cars can be mended” warranted a new trial notwithstanding the judge’s charge to the jurors that they were the sole judges of the facts); Zebouni v. United States, 226 F.2d 826, 828-29 (5th Cir.1955) (holding that based on the entire record the defendant’s right to a fair and impartial trial was violated despite the issuance of a curative instruction almost immediately after one prejudicial episode between the judge and counsel).
III.
For these reasons, we affirm in part, reverse in part and conditionally grant the writ.