concurring in part and dissenting in part:
I respectfully dissent from Parts I, IV, V, and VI of the majority opinion, and concur in the remainder.1 Because the relevant state court decisions addressing Hedlund’s Eddings,2 ineffective assistance of counsel, and shackling claims are unreasonable applications of clearly established Supreme Court precedent and include unreasonable factual determinations, I would reverse and remand this appeal.
Charles Michael Hedlund’s constitutional rights were violated at every stage of his proceedings, from his plea negotiations to his sentencing, yet the Arizona courts affirmed on direct appeal and denied his petition for postconviction relief. First, Hedlund’s trial counsel allowed a plea offer to expire that would have likely saved Hedlund from a death sentence. Then, during the trial that resulted from that *825failure, Hedlund was shackled, in view of the jury and without an individualized determination that he, rather than his co-defendant James McKinney, posed a threat to courtroom security or an escape risk. After Hedlund was convicted, his counsel failed to present an expert who could have testified definitively that Hedlund suffered from brain damage. Finally, the sentencing court erroneously refused to consider Hedlund’s mitigation evidence because Hedlund could not show .that any of his mitigating circumstances bore a causal nexus to his crimes.
In examining the reasonableness of the state courts’ decisions under 28 U.S.C. § 2254(d), we look to the last reasoned state court opinion on the claim. Milke v. Ryan, 711 F.3d 998, 1005 (9th Cir.2013) (citing Ylst v. Nunnemaker, 501 U.S. 797, 805, 111 S.Ct. 2590, 115 L.Ed.2d 706 (1991)). The relevant state courts’ decisions were “contrary to, or involved an unreasonable application of, clearly established Federal law, as determined by the Supreme Court of the United States” and were “based on an unreasonable determination of the facts in light of the evidence presented in the State court proceeding.” 28 U.S.C. § 2254(d). These constitutional violations had a “substantial and injurious effect or influence in determining the jury’s verdict,” Brecht v. Abrahamson, 507 U.S. 619, 623, 113 S.Ct. 1710, 123 L.Ed.2d 353 (1993) (quoting Kotteakos v. United States, 328 U.S. 750, 776, 66 S.Ct. 1239, 90 L.Ed. 1557 (1946)), thus entitling Hedlund to habeas relief. Therefore, I would reverse the district court’s decision and remand with instructions to grant a conditional writ of habeas corpus setting aside Hedlund’s conviction and sentence.
I.
Although Hedlund raises a number of constitutional claims, not all of them are legally cognizable. A complete picture of Hedlund’s trial and sentencing, however, is important to appreciate the context in which Hedlund’s constitutional rights were violated. Hedlund’s case should never have proceeded to trial. First, the trial court judge, Judge Sheldon, rejected a plea agreement that the government and Hedlund had agreed upon after reading letters from one of the victims’ families asking him to do so, but indicated he would accept an agreement with greater accountability for McClain’s death. Second, Hedlund’s trial counsel allowed the deadline to expire for a second plea agreement, which would have spared Hedlund from a death sentence while providing for the greater accountability that Judge Sheldon demanded.
Hedlund was tried with his half-brother and co-defendant James McKinney. Judge Sheldon initially granted Hedlund’s unopposed motion for a severance, but then sua sponte decided to empanel dual juries instead, over both the defendants’ and State’s objections and contrary to then Arizona law. See State v. Lambright, 138 Ariz. 63, 673 P.2d 1, 7 (1983). On a petition for special action, the Arizona Court of Appeals vacated the trial court’s order pursuant to Lambright. The Arizona Supreme Court reversed, and overruled its own precedent, by overturning Lambright and holding that Judge Sheldon’s decision to empanel a dual jury was properly within his discretion. Hedlund v. Sheldon, 173 Ariz. 143, 840 P.2d 1008, 1011 (1992) (en banc). Judge Sheldon also decided, based on ex parte, triple-hearsay inmate information that did not even identify Hedlund, that Hedlund should be shackled, and then conducted a hearing on the matter only so he could make a record of his decision.
Hedlund’s counsel moved to recuse Judge Sheldon based on the appearance of impropriety in his sua sponte and onesided *826pretrial rulings. On the same day that Judge Sheldon conducted a hearing on the shackling issue, he was defending himself against Hedlund’s recusal motion before one of his colleagues. Hedlund’s recusal motion was denied. Judge Sheldon then refused to direct that a curtain or other barrier be placed around Hedlund’s table so that the jury would not see the shackles. Judge Sheldon proceeded to conduct the trial in a manner unfair to Hedlund, including conducting the voir dire of a key prosecution witness — voir dire that Hedlund’s counsel had requested — outside of Hedlund’s counsel’s presence. Because Judge Sheldon concluded that McKinney and Hedlund had an “identity of interests,” he determined that McKinney’s counsel’s presence was sufficient representation of Hedlund.
At sentencing, Judge Sheldon found Hedlund’s expert mental health evidence not credible, dismissed his mitigating evidence of child abuse as not related to the crime, and sentenced Hedlund to death.
Finally, when Judge Sheldon “became” the state postconviction relief (“PCR”) court,3 he acted as a witness in the case, opining that Hedlund’s counsel had been very diligent and effective, and denied Hedlund’s request to develop the record through an evidentiary hearing where his experts could testify. The Arizona Supreme Court summarily affirmed the PCR court’s denial of relief. It is clear that Hedlund did not receive a fair trial, and has not yet had access to a meaningful review of his claims. Under AEDPA’s deferential standard, Hedlund’s Eddings, ineffective assistance of counsel, and shackling claims entitle him to habeas relief.
II.
Under Eddings v. Oklahoma, 455 U.S. 104, 102 S.Ct. 869, 71 L.Ed.2d 1 (1982), a state court may not treat mitigating evidence of a defendant’s background as “irrelevant or non-mitigating as a matter of law” merely because it lacks a causal connection to the crime. Towery v. Ryan, 673 F.3d 933, 946 (9th Cir.2012) (per curiam), cert. denied, — U.S. -, 132 S.Ct. 1738, 182 L.Ed.2d 271 (2012). Here, the state courts did precisely what Eddings prohibits: they found mitigating evidence of Hedlund’s abusive childhood as a matter of fact, but treated it as non-mitigating as a matter of law because it lacked a causal connection to the crime.
A.
Words fail to adequately describe the horrors that Hedlund suffered as a child. Indeed, Hedlund’s mental health expert, Dr. Holler, described Hedlund’s childhood as “probably as gruesome as anything that I have come across in 25-plus years in this business.” The testimony of Hedlund’s half-sisters and aunt paint a ghastly picture of extreme physical and emotional abuse. Hedlund and his three half-siblings were raised in the outskirts of Chandler, Arizona, by Shirley and James McKinney, Sr. Shirley and James Senior, however, were not Hedlund’s biological parents. Hedlund was conceived during an extramarital affair that Hedlund’s mother, who was James Senior’s first wife, had with a man named Charles Hedlund. Growing up with the McKinneys, Hedlund was reminded daily that he was not a McKinney.4
*827Hedlund lacked even the basic necessities as a child. He grew up living in filth and was rarely clothed. His home was covered in animal feces, bile, urine, used female hygiene products, and dirty diapers. Hedlund and his siblings were forced to share a very small room with a variety of domesticated and exotic animals, including, at times, a goat, a calf, a spider monkey, and snakes. The children were terrified of snakes, but Hedlund’s stepparents nevertheless housed the snakes in the children’s closet. Hedlund and his siblings were also not allowed to eat anything unless they received permission from Shirley McKinney. At least once a month, she would lock the children in the house while she went shopping and threaten that if they ate anything while she was gone, she would beat them. During the "summer, Shirley would lock Hedlund and his siblings out of the house in heat exceeding 100 degrees. Because they were usually wearing only underwear, Hedlund and his siblings were frequently sunburned while locked out of their home. On one occasion, Hedlund’s aunt saw the children locked outside in the heat, while Shirley and her own biological daughter, who was favored and spared from Shirley’s abuse, sat inside and told Hedlund and his siblings that they were not allowed to drink water from the hose outside. In effect, Shirley and James Senior treated Hedlund and his siblings like the animals they housed in the children’s bedroom.
Shirley McKinney tormented her stepchildren, especially Hedlund, who was the oldest and the “bastard.” She hit them with belts, old wooden boards, skillets, wire hangers, shovels, garden hoses, or anything else that was within reach when she grew angry. She struck indiscriminately and was often helped by her own biological daughter, who held down her step-siblings as they were beaten. Hedlund’s half-sister Donna testified that she and her half-siblings were punched in the face at least once a day. Hedlund’s half-sisters and aunt testified to a number of particular episodes where Hedlund was beaten. For instance, James Senior kept a vicious dog that once attacked Hedlund, who was a small child at the time, and injured his face so badly that he had to receive over 200 stitches. The day after Hedlund received medical treatment, Shirley and her daughter woke Hedlund up in the early morning and beat him for over an hour because his medical treatment had cost them money. Hedlund’s aunt also testified to another incident where she saw Shirley violently beat Hedlund and his half-brother James McKinney. James had been kicked off the school bus for fighting and as Shirley was marching him to the house, she clipped off a foot-long piece of a water hose and began beating James with it. James was small enough at the time that Shirley was holding him in the air by his arm while she beat him with the water hose. Hedlund tried to stop Shirley’s strikes by jumping on her arm and yelling, “Momma, stop it. Momma, stop it.” In response, Shirley pushed Hedlund off and struck him across the face with the hose. Hedlund fell to the ground, hitting the back of his head against the concrete sidewalk.
By any measure, Hedlund’s savage childhood was a mitigating factor that the Arizona courts should have considered. But because they applied the prohibited causal nexus test, Hedlund has not yet received the constitutionally-required review that he is due.
*828B.
It is weli established that when Hedlund was sentenced in 1993, Eddings and its progeny required that the sentencer give “independent mitigating weight” to all relevant mitigating evidence. See Eddings, 455 U.S. at 110, 102 S.Ct. 869 (internal quotation marks omitted). At that time, the Supreme Court had already clarified that the Eighth and Fourteenth Amendments specifically require the sentencer to fully consider all mitigating evidence, regardless of the lack of a causal connection between the evidence and the defendant’s crime of conviction:
There is no disputing that this Court’s decision in Eddings requires that in capital cases the sentencer ... not be precluded from considering, as a mitigating factor, any aspect of a defendant’s character or record and any of the circumstances of the offense that the defendant proffers as a basis for a sentence less than death. Equally clear is the corollary rule that the sentencer may not refuse to consider or be precluded from considering any relevant mitigating evidence. These rules are now well established ____
Skipper v. South Carolina, 476 U.S. 1, 4, 106 S.Ct. 1669, 90 L.Ed.2d 1 (1986) (emphasis in original) (internal quotation marks and citations omitted); see also McKinney v. Ryan, 730 F.3d 903, 921-22 (9th Cir.2013) (Wardlaw, J., dissenting), vacated and reh’g en banc granted, 745 F.3d 963 (9th Cir.2014).
The last reasoned state court decision addressing Hedlund’s Eddings claim is the Arizona Supreme Court’s decision on direct appeal. The Arizona Supreme Court reviewed the sentencing court’s mitigation findings for abuse of discretion and held that the sentencing court “did not err in concluding that Hedlund’s family background was not sufficiently mitigating to require a life sentence.” State v. McKinney, 185 Ariz. 567, 917 P.2d 1214, 1227 (1996) (en banc), superseded by statute on other grounds as stated in State v. Martinez, 196 Ariz. 451, 999 P.2d 795, 806 (2000) (en banc). Then, after concluding that the trial judge erred in finding one of the two aggravating factors it applied in Hedlund’s case, the Arizona Supreme Court reweighed the aggravating and mitigating circumstances. However, the Arizona Supreme Court applied no “reasoned and individualized” independent analysis of the evidence of childhood abuse that Hedlund presented in mitigation. Cf Towery, 673 F.3d at 945 (“The [Arizona] supreme court also independently weighed the mitigating evidence against the aggravating circumstances to determine whether leniency was called for. As part of that review, the court considered whether evidence of Towery’s difficult childhood should be given substantial weight.”) (internal alterations, quotation marks, and citations omitted). Instead, the Arizona Supreme Court adopted the sentencing court’s analysis of the mitigation evidence, noting only that “the judge did not improperly exclude mitigating evidence at sentencing and the mitigating evidence is not of great weight.” McKinney, 917 P.2d at 1231. Because the Arizona Supreme Court adopted the sentencing court’s mitigation analysis, reviewing it for abuse of discretion and conducting no reasoned and individualized independent analysis of its own, we must “look at both decisions to fully ascertain the reasoning of the last decision.” Id. at 944 n. 3 (quoting Barker v. Fleming, 423 F.3d 1085, 1093 (9th Cir. 2005)).
The sentencing court found evidence of Hedlund’s tortured childhood “compelling and credible” and then presented its analysis of the mitigation evidence:
*829I have considered [the childhood abuse]. I think it is the Court’s obligation to consider it, whether or not it complies with the requirements in (G)(1) [a statutory mitigating factor].
I have also considered all of the other mitigating factors which were set forth in three separate pleadings submitted by defense counsel in this case. I have reviewed all of them again as recently as yesterday and some of those factors this morning. The Court, after carefully considering and weighing all of the aggravating or mitigating factors presented in this case, and not limited to the personality traits discussed by Dr. Holler, past drug and alcohol use discussed about [sic] Dr. Shaw, Dr. Holler and the other witnesses who testified, and the child abuse which the Court finds is a fact, that none of those mitigating factors considered separately or cumulatively indicates to the Court that these factors affected the defendant’s ability to control his physical behavior at the time of the offense or to appreciate the wrongfulness of his conduct, that the defendant was aware at all times while these offenses were occurring that what he was doing was wrong, that he continued to participate in them and that he had the intelligence and the ability to refuse continued participation.
Sentencing Hr’g Tr. at 23-24, July 30,1993 (emphasis added). The sentencing court made explicit that it considered and weighed Hedlund’s mitigating evidence insofar as that evidence affected Hedlund’s capacity at the time of the crime. The court evaluated Hedlund’s abusive childhood only to the extent it may have “affected the defendant’s ability to control his physical behavior at the time of the-offense or to appreciate the wrongfulness of his conduct.” Id. at 24. This refusal to consider and give effect to significant mitigating evidence that the court found credible because it was not tied to his behavior in committing the crime is contrary to Ed-dings. The sentencing court never indicated that its analysis quoted above, clearly requiring a nexus between the child abuse and the crime, went to the weight it was giving the evidence, rather than its relevance.
The majority’s conclusion that the Arizona Supreme Court permissibly weighed the mitigating evidence of Hedlund’s tormented childhood, and did not improperly apply a causal nexus test, is contrary to the Arizona Supreme Court’s plain reasoning. The majority supports its reading of the record based on a single sentence from the Arizona Supreme Court’s decision: “A difficult family background, including childhood abuse, does not necessarily have substantial mitigating weight absent a showing that it significantly affected or impacted a defendant’s ability to perceive, to comprehend, or to control his actions.” McKinney, 917 P.2d at 1227. But the majority ignores the legal authority that the court cited to support this rule. The Arizona Supreme Court cited to a portion of its decision in State v. Ross that states unambiguously, a “difficult family background is not a relevant mitigating circumstance unless a defendant can show that something in that background had an effect or impact on his behavior that was beyond the defendant’s control.” 180 Ariz. 598, 886 P.2d 1354, 1363 (1994) (internal quotation marks omitted). If he cannot show this, “Defendant’s background therefore is not a mitigating circumstance.” Id. The court in Ross did not say that family background unconnected to criminal conduct is of “little or no weight or value.” Cf. Towery, 673 F.3d at 945 (“[T]he [Arizona] court’s reasoned and individualized decision to give [defendant’s] evidence little or no weight was not contrary to Supreme Court precedent.”) (emphasis in *830original). Instead, Ross held, as a matter of law, that a causally unrelated family background was not relevant. By deeming the unconnected background evidence not relevant, the Ross court applied the impermissible causal nexus test that the U.S. Supreme Court has unequivocally rejected. See Smith v. Texas, 543 U.S. 37, 45, 125 S.Ct. 400, 160 L.Ed.2d 303 (2004) (per curiam) (stating that the causal nexus test is one “we never countenanced and now have unequivocally rejected”); Tennard v. Dretke, 542 U.S. 274, 284-86, 124 S.Ct. 2562, 159 L.Ed.2d 384 (2004) (explaining that the causal nexus test “has no foundation in the decisions of this Court”). In the same way it did in Ross, the Arizona Supreme Court impermissibly refused to consider Hedlund’s childhood abuse as a mitigating factor. Although the Arizona Supreme Court cloaked this exclusion in the language of weighing and sufficiency, it plainly and improperly applied a causal nexus requirement.
The majority avoids reaching this conclusion by requiring Hedlund to meet a heightened standard of proof that is not the law of our circuit. The majority stretches the language of Schad v. Ryan, 671 F.3d 708 (9th Cir.2009) (per curiam), cert. denied, — U.S. -, 133 S.Ct. 432, 184 L.Ed.2d 264 (2012), to invent a new legal test which requires Hedlund to show a “clear indication” that the Arizona courts applied an unconstitutional causal nexus test.5 Our use of the phrase “clear indication” in Schad, however, was very different than the majority’s use here. In Schad, we concluded that because there was “no indication” that the state courts applied a causal nexus test, “either as a method of assessing the weight of the mitigating evidence, or as an unconstitutional screening mechanism,” we would not presume constitutional error absent a “clear indication in the record.” 671 F.3d at 724. Thus, we would not presume error from silence. Here, in contrast, the record unquestionably shows that the Arizona courts applied a causal nexus test. The only question is whether they applied the unconstitutional version of the causal nexus test. The “clear indication” test from Schad — if it even is a test — helps courts determine whether an Eddings violation exists when a record contains no indication that any causal nexus test was used. It was not intended as a tool for determining what variety of causal nexus test a state court applied. Our controlling precedent for determining that question is whether the state court’s reasoning “appears to have” imposed the impermissible test. Styers v. Schriro, 547 F.3d 1026, 1035 (9th Cir. 2008); see also Poyson, 743 F.3d at 1188 (Kozinski, C.J., dissenting from the denial of rehearing en banc). To conclude otherwise is to disavow controlling Supreme Court authority, and blatantly break from our circuit’s precedent, as the same majority did in McKinney v. Ryan, 730 F.3d at 919-21, vacated and reh’g en banc granted, 745 F.3d 963 (9th Cir.2014); see also Poyson, 743 F.3d at 1187-88 (Kozinski, C.J., joined by eleven other judges, dissenting from the denial of rehearing en banc because of a similar Eddings issue).
Lastly, the Arizona Supreme Court did not remedy its Eddings violation by reweighing the aggravating circumstance of pecuniary gain against the aggregate mitigating evidence. Because the Court had already determined that evidence of Hedlund’s abusive childhood was not a mitigat*831ing factor (because it was not causally connected to his criminal conduct and, therefore in accordance with Ross, not relevant), that evidence was not considered in the “minimal mitigating evidence” that the court reweighed against the aggravating circumstance of pecuniary gain. McKinney, 917 P.2d at 1231. The Arizona Supreme Court unconstitutionally applied the causal nexus test to screen out Hedlund’s evidence of childhood abuse and family background.6
III.
During sentencing, defense counsel presented expert testimony describing Hedlund’s alcoholism and the psychological effects of his abusive childhood. In particular, Dr. Holler testified that Hedlund suffered from post-traumatic stress disorder (PTSD), alcohol dependence, and depression. Dr. Holler also testified that because Hedlund was constantly seeking acceptance, he was susceptible to engaging in criminal activity to demonstrate loyalty to his family. However, Judge Sheldon did not find the evidence presented by Dr. Holler and Dr. Shaw, Hedlund’s other expert witness, credible. Dr. Holler’s evidence was not credited because: (1) he did not raise PTSD in his initial report, instead announcing it for the first time while testifying; (2) some of the information was self-reported by Hedlund; and (3) some of the conclusions were based on an erroneous presentence report. Dr. Shaw’s evidence was not credited because his information came from Hedlund, who had a motive to lie about the extent of his alcohol consumption, and his statements conflicted with those of Hedlund’s sisters as well as a presentence report from an earlier conviction.
Hedlund claimed before the state PCR court that his trial counsel’s assistance was constitutionally deficient during the penalty phase because counsel failed to present competent expert testimony that Hedlund suffered from brain damage that affected his behavior at the time of the offenses. Hedlund offered a report from Dr. Walter, a neuropsychologist, and requested a hearing at which Dr. Walter could testify. The PCR court denied Hedlund’s request for an evidentiary hearing, concluding that the opinions of Hedlund’s sentencing phase experts were “substantially similar to those of Dr. Walter’s” and that “it was adequately informed of this information through Dr. Holler and Dr. Shaw’s opinions.” State v. Hedlund, No. CR 1991-090926, Minute Entry at 6 (Ariz.Super.Ct. June 1, 2001).
The majority incorrectly frames the issue as whether the state PCR court’s finding that the evidence was “substantially similar” is an unreasonable determination of fact. If the sentencing court had credited the evidence, but found it insufficiently *832mitigating, this would be the correct inquiry. However, here, the sentencing court wholly discredited the expert mental health evidence because the sources underlying each item were questionable. Because Hedlund’s counsel had presented unreliable expert testimony, it was as if he had presented none at all. Thus, the only relevant inquiry for the state PCR court was whether Dr. Walter’s report was also so unreliable that it (or a similar report procured by competent counsel) would •have been rejected by the sentencing court. The state PCR court did not engage in this analysis.
“An unreasonable application of Supreme Court precedent occurs when a ‘state-court decision ... correctly identifies the governing legal rule but applies it unreasonably to the facts-of a particular prisoner’s case.’ ” Libberton v. Ryan, 583 F.3d 1147, 1161 (9th Cir.2009) (quoting Williams v. Taylor, 529 U.S. 362, 407-08, 120 S.Ct. 1495, 146 L.Ed.2d 389 (2000)). Because the state court unreasonably applied Strickland by failing to address the claim of ineffective assistance of counsel presented, and rejected Hedlund’s request to develop the record as to this issue, Hedlund is entitled to an evidentiary hearing on this claim. See Martinez v. Ryan, - U.S. -, 132 S.Ct. 1309, 1317, 182 L.Ed.2d 272 (2012) (ineffective assistance of counsel claims “often require investigative work” and “often turn[] on evidence outside the trial record”); see also Cullen v. Pinholster, — U.S. -, 131 S.Ct. 1388, 1400-01, 179 L.Ed.2d 557 (2011) (citing Schriro v. Landrigan, 550 U.S. 465, 473, 127 S.Ct. 1933, 167 L.Ed.2d 836 (2007), for the proposition that “district courts, under AEDPA, generally retain the discretion to grant an evidentiary hearing” and noting that “[sjection 2254(e)(2) continues to have force where § 2254(d)(1) does not bar federal habeas relief’); Siripongs v. Calderon, 35 F.3d 1308, 1310 (9th Cir.1994) (“In a capital case, a habeas petitioner who asserts a colorable claim to relief, and who has never been given the opportunity to develop a factual record on that claim, is entitled to an evidentiary hearing in federal court.”).
IV.
A.
Before Hedlund’s trial began, his trial counsel and the prosecutor were engaged in plea negotiations, which resulted in two proposed plea agreements. The first plea agreement would have allowed Hedlund to plead guilty to second-degree murder for his involvement in the death of Mertens and to theft with a prior for his involvement in the death of McClain. This plea agreement was presented to and rejected by Judge Sheldon on September 18, 1992. Judge Sheldon rejected the agreement because he wanted the offenses to which Hedlund would plead guilty to reflect greater accountability for the McClain homicide.7 To give the parties an opportunity to reach'a second plea that addressed his concerns, Judge Sheldon granted a two-week continuance. The week before the trial was scheduled to begin, the parties were in the, final stages of a second plea agreement that addressed Judge Sheldon’s concerns. Under the second proposed plea agreement, Hedlund would have pled guilty to the second-degree murder of Mertens, and to two felonies — theft *833and burglary — in connection with McClain’s death. At this point in the proceedings, Hedlund’s primary goal was to avoid the death penalty, which the second plea agreement accomplished.
On Tuesday, October 6, 1992, a week before trial was to begin, Hedlund’s counsel and the prosecutor called Judge Sheldon to clarify what might be appropriate for the second plea agreement. Judge Sheldon’s deadline for the parties to appear to discuss the second plea agreement with the court was two days later, on Thursday, October 8. The prosecutor set a deadline for Hedlund to accept the second plea agreement by Friday, October 9. But instead of presenting this second plea agreement to Judge Sheldon or accepting the prosecutor’s offer, Hedlund’s trial counsel called the court at the end of the day on October 8 to ask Judge Sheldon to recuse himself, because he believed that the judge was biased against Hedlund. Judge Sheldon refused defense counsel’s request, and the deadline to discuss the second plea agreement expired without any further communication from Hedlund’s counsel. The next day, Hedlund’s counsel filed a motion to recuse Judge Sheldon, and the prosecutor’s second plea offer expired. Another judge held a hearing on the recusal motion four days later and denied the motion. Voir dire for Hedlund’s trial began that same day.
Hedlund claims that his trial counsel was ineffective in the plea process. The last reasoned court decision addressing this issue is the Arizona Superior Court’s denial of Hedlund’s petition for post-conviction relief. The state PCR court’s denial of Hedlund’s ineffective assistance during plea negotiations claim was an unreasonable application of clearly established federal law as determined by the Supreme Court. See 28 U.S.C. § 2254(d)(1).
B.
“[C]riminal justice today is for the most part a system of pleas, not a system of trials.... [T]he right to adequate assistance of counsel cannot be defined or enforced without taking account of the central role plea bargaining plays in securing convictions and determining sentences.” Lafler v. Cooper, — U.S. -, 132 S.Ct. 1376, 1388, 182 L.Ed.2d 398 (2012). “Ninety-seven percent of federal convictions and ninety-four percent of state convictions are the result of guilty pleas.” Id. The majority correctly identifies the two-part Strickland test applicable to ineffective assistance claims in plea negotiations and properly emphasizes that under AED-PA, review of the state court’s application of Strickland is “doubly deferential.” See also Harrington v. Richter, — U.S. -, 131 S.Ct. 770, 788, 178 L.Ed.2d 624 (2011). Where the majority fails, however, is in its examination of the state PCR court’s actual reasons for rejecting Hedlund’s claim.
After describing Arizona’s version of Strickland, the state PCR court rejected Hedlund’s ineffective assistance claim in four sentences. It said:
Counsel claimed that Defendant’s trial attorney was ineffective in handling his plea negotiations. A hearing was conducted on this matter in the trial court and Defendant’s attorney properly, aggressively and professionally pursued the issue. There was no ineffective assistance of counsel in the court’s rejection of the plea offer presented to it. As the Court noted to counsel below, any plea which had lacked accountability for the McClain homicide would have been rejected by the Court.
State v. Hedlund, No. CR 1991-090926, Minute Entry at 3-4 (Ariz.Super.Ct. June 1, 2001) (emphasis added). This ruling is *834an unreasonable application of clearly established Supreme Court precedent.
Contrary to the state PCR court’s statement, a hearing was not conducted by the trial court on either the issue of ineffective assistance of counsel or the plea negotiations and second plea agreement. There was, however, a hearing on the motions for recusal and for change of judge. By incorrectly stating that the hearing regarding the motion to recuse Judge Sheldon was a hearing on the ineffective assistance claim, the state PCR court applied the correct law to the wrong facts. Based on the recusal hearing, the state PCR court concluded that defense counsel “properly, aggressively and professionally pursued the issue.” But defense counsel’s efforts to transfer Hedlund’s case to another judge were irrelevant to evaluating whether counsel performed deficiently by abandoning the second plea agreement.
Also, even if the state PCR court had applied Strickland to the correct facts, its conclusion that Hedlund’s trial counsel’s performance was not constitutionally deficient was unreasonable. To prove ineffective assistance during the plea phase of a prosecution, Hedlund “must demonstrate gross error on the part of counsel.” Turner v. Calderon, 281 F.3d 851, 880 (2002) (quoting McMann v. Richardson, 397 U.S. 759, 772, 90 S.Ct. 1441, 25 L.Ed.2d 763 (1970)). Hedlund’s trial counsel committed gross error when he allowed the second offer of a plea agreement, which would have saved Hedlund from the death penalty, to expire even though Hedlund had told him he was willing to accept the plea. The majority’s cursory dismissal of this issue and focus on trial counsel’s “tactical” decision to move for a recusal are misguided.
The majority, like the state PCR court, misreads the record. The majority concludes that the record is unclear as to whether a second plea offer was still on the table on Thursday, October 8 — the date that defense counsel first asked Judge Sheldon to recuse himself and allowed the court’s plea deadline to expire. Because of this purported lack of clarity, the majority dismisses Hedlund’s contention that the second plea offer was still available and that his counsel’s decision to let the second plea offer expire and expose Hedlund to the death penalty was ineffective assistance. The record does reflect that on Tuesday, October 13, when the parties were arguing defense counsel’s motion to recuse, the prosecutor said, “There is no plea offer to Mr. Hedlund as of today.... We were talking Tuesday afternoon [October 6] and [defense counsel] advised me that Mr. Hedlund turned down the latest offer which is in his pleading today. I said, fine. If we don’t plead something by Friday, that is it.” Evidentiary Hr’g Tr. at 56-57, Oct. 13, 1992. But it is wrong to conclude from this statement alone that the second plea agreement had been unconditionally rejected by the time defense counsel moved to recuse Judge Sheldon.
When he moved to recuse Judge Sheldon, defense counsel had a plea offer in hand that would have saved Hedlund from the death penalty.8 The record, however, is ambiguous as to whether the “latest offer” the prosecutor referred to during the recusal hearing was the second plea offer or a different offer. At other times during the hearing, the prosecutor, Judge *835Sheldon, and Judge Sheldon’s assistant each said that they believed the parties were going to come in on Thursday, October 8 to discuss the latest plea agreement with the judge. Thus, it is unlikely that the “latest plea” had been rejected on Tuesday, October 6 if the parties, as of Wednesday, were planning to meet with Judge Sheldon to discuss that very plea proffer. Also, as of October 9, defense counsel stated in his motion to recuse Judge Sheldon that “Hedlund and the State have come to a[sic] agreement which is within the range suggested by the court.” Lastly, on the same day that defense counsel told the prosecutor that the “latest offer” had been rejected, the parties phoned the court to clarify what would be a more appropriate plea. The parties were supposed to come in on Thursday, October 8 to further discuss the second plea agreement with Judge Sheldon. Defense counsel’s failure to accept the proposed second plea offer, or, at the very least, meet with Judge Sheldon and the prosecutor to further discuss the offer on October 9 as they had agreed, was deficient performance. To conclude otherwise is an unreasonable application of Strickland.
The test for deficient performance is “whether counsel’s assistance was reasonable considering all the circumstances.” Strickland, 466 U.S. at 688,104 S.Ct. 2052. Contrary to the majority’s position, defense counsel’s tactical decision to move for recusal does not excuse his abandonment of the second plea offer negotiations.9 Defense counsel knew that Hedlund’s primary goal in pleading guilty to lesser charges was to avoid a capital sentence, yet when presented with a plea that would have resulted in a sentence of a term of years, defense counsel irresponsibly shifted all of his efforts to the recusal motion and allowed the second plea offer to expire. The majority portrays the decision to abandon the plea negotiations and pursue the recusal motion as one strategic choice subject to the doubly-deferential standard under Strickland and AEDPA. But the majority provides no reason as to why it was strategic for Hedlund’s counsel to drop one and pursue the other, rather than pursue both concurrently. The record contains no articulable reason as to why abandoning the plea negotiations, after Hedlund told counsel he was willing to accept the latest offer, was strategic.
The majority’s explanation is unconvincing. If defense counsel strategically ended plea negotiations in Judge Sheldon’s court, why did he cease all negotiations with Judge Sheldon and the prosecutor, thus allowing the offer — which according to the majority he sought to preserve — to expire? The only explanation is error, amounting to deficient performance. Defense counsel’s failure to attend the plea discussions with the prosecutor scheduled before Judge Sheldon suggests that negligence, rather than strategy, was at play. Defense counsel’s failures were especially egregious because trial was to begin only a few days later and the chances of another plea offer were slim to nil.
“During all critical stages of a prosecution, which must include the plea bar*836gaining process, it is counsel’s ‘dut[y] to consult with the defendant on important decisions and to keep the defendant informed of important developments in the course of the prosecution.’ ” Nunes v. Mueller, 350 F.3d 1045, 1053 (9th Cir. 2003) (quoting Strickland, 466 U.S. at 688, 104 S.Ct. 2052). It is unknown if Hedlund knew that defense counsel had abandoned the plea bargaining process in favor of trying to recuse Judge Sheldon. By abandoning plea negotiations merely as a consequence of pursuing the recusal motion, and not as the result of an informed and separate decision, defense counsel failed to provide constitutionally effective counsel to Hedlund.
The state PCR court’s conclusion that Hedlund did not suffer prejudice because “any plea which had lacked accountability for the McClain homicide would have been rejected by the Court,” was also an unreasonable application of established Supreme Court precedent.10 To succeed on his ineffective assistance claim, Hedlund must also show that:
but for the ineffective advice of counsel there is a reasonable probability that the plea offer would have been presented to the court (ie., that the defendant would have accepted the plea and the prosecution would not have withdrawn it in light of intervening circumstances), that the court would have accepted its terms, and that the conviction or sentence, or both, under the offer’s terms would have been less severe than under the judgment and sentence that in fact were imposed.
Lafler, 132 S.Ct. at 1385. “[I]t is necessary to show a reasonable probability that the end result of the criminal process would have been more favorable by reason of a plea to a lesser charge or a sentence of less prison time.” Missouri v. Frye, - U.S. -, 132 S.Ct. 1399, 1409, 182 L.Ed.2d 379 (2012). “A reasonable probability is a probability sufficient to undermine confidence in the outcome.” Strickland, 466 U.S. at 694, 104 S.Ct. 2052. If the second plea offer had been accepted, Hedlund would have been sentenced to a term of years and avoided the death penalty. The majority’s explanation to support its view that Judge Sheldon would not have accepted the second plea agreement is flawed.11
First, the majority concludes that it is doubtful whether Judge Sheldon would have found that the second plea agreement “provided sufficient accountability for the McClain homicide.” But the excerpts from the hearing on defense counsel’s motion to recuse that the majority *837cites do not support this conclusion. Instead, they demonstrate Judge Sheldon’s concern regarding a second-degree murder charge under the felony murder rule. In the first plea offer, Hedlund would plead to second-degree murder for the homicide of Mertens, not McClain. Judge Sheldon’s concern with the second-degree murder plea was not related to Hedlund’s culpability in the McClain homicide. Instead, he was concerned with the legal feasibility of Hedlund pleading guilty to second-degree murder under Arizona’s felony murder rule. See Ariz.Rev.Stat. § 13-1105(A)(2). Judge Sheldon was “very surprised” by the first plea agreement because he had doubts that the parties would be able “to establish a factual basis for Second Degree Murder to a Felony Murder charge ... [considering that] there are no lesser included offenses to Felony Murder.” Evidentiary Hr’g Tr. at 37-38, Oct. 13, 1992. If the second plea agreement had been presented to the court, however, there is a reasonable probability that Judge Sheldon would have accepted it even with the second-degree murder charge for the Mertens homicide. In fact, Judge Sheldon sentenced Hedlund based, in part, on a second-degree murder conviction under Arizona law. Thus, because the facts permitted the jury to convict Hedlund of second-degree murder, Judge Sheldon’s concern over the legal feasibility of the felony murder rule would not have barred his approval of the second plea agreement.
Second, the record shows that any accountability concerns that Judge Sheldon had regarding Hedlund’s culpability in the McClain homicide would have been likely addressed with the addition of the burglary charge in the second plea offer. Although Judge Sheldon had “not made up [his] own mind,” at the time of the recusal hearing as to whether burglary would be sufficient to address his culpability concerns, he did “believe[ ] at the very least, what [counsel] should be looking at would be Burglary in the First Degree which would be entry in the home armed with a weapon either as a dangerous offense or non-dangerous.” Id. at 32. While both sides can speculate about whether the burglary charge would have definitively satisfied Judge Sheldon’s accountability concerns, the record contains enough evidence to raise doubt, and undermine confidence, as to whether Hedlund would have received a capital sentence if defense counsel had presented the second plea agreement to the court.
Third, the majority’s conclusion that Judge Sheldon’s due process concerns would have doomed the second plea agreement is incorrect. Judge Sheldon’s concern regarding disparate sentences was a reason that he rejected the first plea agreement, not a reason that he may have rejected the second plea agreement. Unlike the first plea agreement, the second plea agreement included a first-degree burglary charge which, according to petitioner’s counsel at oral argument, would have exposed Hedlund to approximately fourteen more years of prison time than the first plea agreement. Also, disparate sentences between McKinney and Hedlund were appropriate as their culpability, in at least the Mertens homicide, differed. In fact, the jury convicted McKinney of first-degree murder, and Hedlund of second-degree murder, for the Mertens homicide.
As to Hedlund’s willingness to accept the second plea offer, there is no evidence in the record that Hedlund himself did not want to enter a plea in front of Judge Sheldon. The only evidence of Hedlund’s input into defense counsel’s decision to abandon the second plea offer and pursue the motion to recuse is that Hedlund “fe[lt] that [the] court ha[d] become biased against him.” But Hedlund’s feelings *838about Judge Sheldon were insufficient to justify defense counsel’s decision to abandon the plea negotiations altogether and instead move for recusal.
Finally, the ultimate question is whether there is a “reasonable probability” that Hedlund would have been spared the death penalty had defense counsel presented the second plea agreement to Judge Sheldon. See Strickland, 466 U.S. at 694, 104 S.Ct. 2052. Speculating about whether Judge Sheldon would have accepted the second plea agreement is an intermediate step. And in truth, speculating about whether Judge Sheldon would have accepted the second plea agreement is just that: speculation. At the time of the recusal hearing, even Judge Sheldon admitted that he “had not made up [his] own mind whether or not that [the burglary charge] would be an appropriate disposition.” Evidentiary Hr’g Tr. at 32, Oct. 13, 1992. What is known is that Judge Sheldon was not committed to sentencing Hedlund to death, and was open to the State pursuing a term of years instead of a death sentence. When presented with the first plea agreement, he “was very surprised there had not been a plea to First Degree Murder with the State stipulating it would not seek the death penalty.” Id. at 37. Thus, there is a reasonable probability that if Judge Sheldon had been presented with a plea agreement that satisfied his other concerns and did not include the possibility of a capital sentence, he would have accepted it.
y.
The last reasoned state court decision on Hedlund’s shackling claim is the Arizona Supreme Court’s decision on direct appeal. See McKinney, 917 P.2d at 1223. Because that decision is both “contrary to ... clearly established Federal law,” 28 U.S.C. § 2254(d)(1), and “based on an unreasonable determination of the facts in light of the evidence presented in the State court proceeding,” id. at § 2254(d)(2), we “must then resolve the claim without the deference AEDPA otherwise requires,” Panetti v. Quarterman, 551 U.S. 930, 953, 127 S.Ct. 2842, 168 L.Ed.2d 662 (2007); accord Frantz v. Hazey, 533 F.3d 724, 735-37 (9th Cir.2008) (en banc).
A.
The trial court decided to shackle Hedlund before it even heard testimony about a possible flight risk. Then, to “put [the] matter on the record,” the trial court heard testimony from Deputy Sheriff Jack Roger Lane, who said that an inmate told a subordinate officer that he had overheard James McKinney and another prisoner who was charged with murder discussing an escape plan. The inmate could not identify McKinney’s co-plotter. Deputy Sheriff Lane testified that Hedlund’s jail card contained a narrative about the escape plan, which read, “Warning, take keys and clothing per class A1920. McKinney planning escape by jumping guard per information, 300120, Per request CPD 2525.” Deputy Sheriff Lane had asked that the information about McKinney’s escape attempt be placed on Hedlund’s jail card. This is the only record evidence that Hedlund was an escape risk.
Under AEDPA, a challenge to a state court’s factual determination may be based on “the claim that the finding is unsupported by sufficient evidence, that the process employed by the state court is defective, or that no finding was made by the state court at all.” Taylor v. Maddox, 366 F.3d 992, 999 (9th Cir.2004) (citations omitted). “[W]e must more than merely doubt whether the process operated properly. Rather, we must be satisfied that any appellate court to whom the defect is pointed out would be unreasonable in holding that *839the state court’s fact-finding process was adequate.” Id. at 1000. While this “is a daunting standard ... [it] is not impossible to meet.” Id.
The Arizona Supreme Court rejected Hedlund’s shackling claim because “the trial judge specifically made a record to document his security concerns: Hedlund attempted an escape during the summer of 1991 and also made plans with another capital defendant to escape by attacking a guard and taking his uniform and gun.” McKinney, 917 P.2d at 1223. But it was McKinney, not Hedlund, who attempted to escape in 1991. And it was McKinney, not Hedlund, who had been overheard making an escape plan in 1992. Therefore, the Arizona Supreme Court’s factual determination as to Hedlund was not only incorrect, it was an unreasonable determination given that the only fact possibly supporting the notion that Hedlund was a security risk was the jailhouse informant’s hearsay statement that McKinney was planning to escape with another capital defendant. But none of the state’s witnesses knew the identity of McKinney’s co-plotter. By misattributing McKinney’s 1991 escape attempt to Hedlund, the trial court unreasonably inferred that Hedlund’s past attempt made it probable that he was planning another escape with McKinney. Absent incorrectly believing that Hedlund had tried to escape in 1991, the Arizona Supreme Court gave no indication that it would have reached the same result, and thus the finding that Hedlund was a security risk was “unsupported by sufficient evidence,” Taylor, 366 F.3d at 999.
Moreover, the Arizona Supreme Court did not just err as to the 1991 escape attempt. When discussing the 1992 escape plan, it again erroneously attributed evidence against McKinney to Hedlund. It said that the trial judge’s assessment of Hedlund’s security threat was supported because Hedlund had made plans with another capital defendant to escape. But the only evidence of an escape plan was from Deputy Sheriff Lane, who said that McKinney, known by name to the inmate who overheard the plot, made escape plans with another defendant charged with murder, whose name the inmate did not know. While the inference that the other defendant was Hedlund may or may not be permissible, the Arizona Supreme Court did not make that inference. Instead, it erroneously attributed the exact record evidence of McKinney’s flight risk to Hedlund, resulting in an unreasonable determination of fact to support its conclusion that the trial court had well-founded security concerns as to Hedlund.
The Arizona Supreme Court did not make a finding that Hedlund was McKinney’s unnamed co-plotter. But even if it had, that finding would also be an unreasonable determination of fact. The only evidence that Hedlund was involved in the 1992 plan is that he was charged with murder (thus fitting the description of the inmate with whom McKinney allegedly spoke) and had McKinney’s escape plan denoted on his jail card. The majority places great weight on the jail’s “security-based decision” to identify Hedlund as a threat on his jail card. However, Deputy Sheriff Lane testified that he asked that information about McKinney’s 1992 plot be placed on Hedlund’s jail card. Thus, the jail card reflected nothing more than Deputy Sheriff Lane’s opinion that McKinney’s co-plotter was likely Hedlund. But Deputy Sheriff Lane’s testimony demonstrates that the only basis for his opinion was the third-hand information that McKinney had plotted with another defendant charged with murder. Deputy Sheriff Lane’s “opinion,” then, was nothing more than a hunch. And hunches should not justify the use of highly prejudicial restraints during trial. *840Cf. Gonzalez v. Pliler, 341 F.3d 897, 902 (9th Cir.2003) (“[W]hen the imposition of restraints is to be based upon conduct of the defendant that occurred outside the presence of the court, sufficient evidence of that conduct must be presented on the record so that the court may make its own determination of the nature and seriousness of the conduct and whether there is a manifest need for such restraints; the court may not simply rely upon the judgment of law enforcement or court security officers or the unsubstantiated comments of others.”) (quoting People v. Mar, 28 Cal.4th 1201, 124 Cal.Rptr.2d 161, 52 P.3d 95, 107 (2002)). Without the misattributed 1991 escape attempt and 1992 escape plot, the trial court’s only basis for shackling Hedlund in view of the jury was Deputy Sheriff Lane’s speculation that Hedlund was McKinney’s co-plotter, which itself was based on an inmate’s recollection of what he overheard.
In the absence of any individualized, evidence tying Hedlund to McKinney’s 1992 escape plot, the Arizona Supreme Court’s factual determination that Hedlund was involved in the 1992 plot was unreasonable.
B.
It was clearly established by 1996, when the Arizona Supreme Court issued its decision rejecting Hedlund’s shackling claim, that shackling a criminal defendant during his jury trial is “the sort of inherently prejudicial practice that ... should be permitted only where justified by an essential state interest specific to each trial.” Holbrook v. Flynn, 475 U.S. 560, 568-69, 106 S.Ct. 1340, 89 L.Ed.2d 525 (1986). Indeed, “no person should be tried while shackled and gagged except as a last resort.” Illinois v. Allen, 397 U.S. 337, 344, 90 S.Ct. 1057, 25 L.Ed.2d 353 (1970) (emphasis added); see Wilson v. McCarthy, 770 F.2d 1482, 1484-85 (9th Cir.1985) (“Shackling may be justified as a last resort, in cases of extreme need, or in cases urgently demanding that action.”) (citations and internal quotation marks omitted). The Arizona Supreme Court’s rejection of Hedlund’s shackling claim was an unreasonable application of clearly established Supreme Court case law in two ways.
First, even if we assume that the trial court had an “essential state interest” in courtroom security, Holbrook, 475 U.S. at 569, 106 S.Ct. 1340, shackling Hedlund in view of the jury was not possibly “the fairest and most reasonable way,” Allen, 397 U.S. at 344, 90 S.Ct. 1057, to alleviate that concern. Although the trial judge replaced the original table, behind which Hedlund and McKinney sat shackled, with a table that covered a greater portion of their legs, he unreasonably refused defense counsel’s request that a curtain or other barrier be placed around the table so that the jury could not see the shackles. As a result, the jury saw Hedlund shackled and the trial court failed in its duty to “carefully guard against dilution of the principle that guilt is to be established by probative evidence and beyond a reasonable doubt.” Estelle v. Williams, 425 U.S. 501, 503, 96 S.Ct. 1691, 48 L.Ed.2d 126 (1976).12
Second, it was an unreasonable application of clearly established federal law to conclude that the trial court had sufficient *841individualized evidence to support shackling Hedlund. Because the Arizona Supreme Court’s legal analysis of this claim was dependent on an antecedent unreasonable factual determination, we “must then resolve the claim without the deference AEDPA otherwise requires.” Panetti 551 U.S. at 953, 127 S.Ct. 2842. Thus, the Arizona Supreme Court’s decision is an unreasonable application of clearly established Supreme Court case law if, after correcting the unreasonable factual determinations, Hedlund’s shackling was constitutionally impermissible. Without misattributing McKinney’s 1991 escape attempt and Deputy Sheriff Lane’s testimony describing McKinney’s 1992 escape plot to Hedlund, the only individualized analysis supporting “the judge’s well-founded security concerns” is Deputy Sheriff Lane’s speculation that Hedlund was McKinney’s co-plotter in the 1992 escape plan. If the safeguards embodied in Allen, Estelle, and Holbrook mean anything, they must require more than a single security officer’s hunch to justify the highly prejudicial act of shackling a criminal defendant in view of the jury charged with determining his guilt.
Even if it was not unreasonable for the trial court to infer that Hedlund was McKinney’s co-plotter, it was an unreasonable application of Supreme Court precedent to shackle Hedlund in view of the jury solely because another inmate may have overheard Hedlund discussing the possibility of escape. To support its conclusion that the trial court was justified in shackling Hedlund, the majority cites Hamilton v. Vasquez, 882 F.2d 1469 (9th Cir.1989), and Crittenden v. Ayers, 624 F.3d 943 (9th Cir.2010). But these cases actually support Hedlund’s position. In Hamilton, we did not address whether “the threat of violence or disturbance was serious enough to justify the shackling of Hamilton throughout his lengthy trial.” 882 F.2d at 1473. However, we did emphasize that “[s]hackling is proper where there is a serious threat of escape or danger.” 13 Id. at 1471 (emphasis added). In quoting the same language, the majority ignores the directive that the threat of escape or danger must be “serious” and fails to analyze this with respect to Hedlund. In Crittenden, shackling the defendant during trial was justified because the defendant had already once escaped from jail, kidnapping a man and commandeering his truck in the process, and twice attempted to escape, assaulting a prison guard during one of the attempts. 624 F.3d at 948. Furthermore, even with this “state interest specific to Crittenden’s trial,” the trial court in that case allowed shackles and handcuffs only outside the presence of the jury. Id. at 971. Thus, not only was the state’s interest in preventing Crittenden’s escape buttressed by much stronger evidence than that in Hedlund’s case, but the means used to protect that interest were much less likely to infringe Crittenden’s constitutional right to a fair trial.
Here, by forcing Hedlund to wear visible leg braces during his trial, the trial judge *842jeopardized the “principle that there is a presumption of innocence in favor of the accused ... axiomatic and elementary, and [whose] enforcement lies at the foundation of the administration of our criminal law.” Coffin v. United States, 156 U.S. 432, 453, 15 S.Ct. 394, 39 L.Ed. 481 (1895); see Spain v. Rushen, 883 F.2d 712, 721 (9th Cir.1989). And he threatened Hedlund’s most fundamental right solely because Deputy Sheriff Lane had a hunch that Hedlund was the unidentified co-plotter in McKinney’s escape plan. In other words, the trial judge decided that one security officer’s suspicion of Hedlund’s involvement in McKinney’s escape plot, based wholly on a conversation that was relayed to that security officer via another inmate and a subordinate officer, was sufficient to extinguish Hedlund’s right to a fair trial. This was an egregious constitutional error that more than satisfies AEDPA’s stringent demands.
C.
Even if Hedlund was unconstitutionally restrained, the violation must have had a “substantial and injurious effect or influence” on the verdict to warrant habeas relief. Brecht, 507 U.S. at 623, 113 S.Ct. 1710 (internal quotation marks omitted). We evaluate the effect keeping in mind that “in cases of grave doubt as to harmlessness the petitioner must win.” O’Neal v. McAninch, 513 U.S. 432, 437, 115 S.Ct. 992, 130 L.Ed.2d 947 (1995). “To determine whether the imposition of physical restraints constitutes prejudicial error, we have considered the appearance and visibility of the restraining device, the nature of the crime with which the defendant was charged and the strength of the state’s evidence against the defendant.” Larson v. Palmateer, 515 F.3d 1057, 1064 (9th Cir.2008). Fundamentally, we ask “whether what [the jurors] saw was so inherently prejudicial as to pose an unacceptable threat to defendant’s right to a fair trial.” Holbrook, 475 U.S. at 572, 106 S.Ct. 1340.
Hedlund was prejudiced when he was shackled in view of the jury. Here, as in Rhoden v. Rowland, 172 F.3d 633 (9th Cir.1999), “the unjustified shackles were not obtrusive, but were visible and actually seen by some of the jurors,” supporting a finding of prejudice. Id. at 637. In Dyas v. Poole, 317 F.3d 934 (9th Cir.2003) (per curiam), we noted that “shackling during trial carries a high risk of prejudice because it indicates that the court believes there is a ‘need to separate the defendant from the community at large, creating an inherent danger that a jury may form the impression that the defendant is dangerous or untrustworthy.’ ” Id. at 937 (quoting Rhoden, 172 F.3d at 636). Prejudice was “particularly likely” in Dyas because at least one juror saw the petitioner’s leg restraints, which were unobtrusive enough that the trial judge had thought they would not have been very visible to jurors. Id. at 936-37. Here, Hedlund’s restraints were visible to jurors throughout the trial. And even though the trial judge allowed Hedlund to sit in front of a table that partially covered his legs, the restraints were still visible. Their visibility supports a finding of prejudice. See Holbrook, 475 U.S. at 568-69, 106 S.Ct. 1340; Dyas, 317 F.3d at 937; Rhoden, 172 F.3d at 637.
The nature of the criminal charges filed against Hedlund also suggests prejudicial error. When “the defendant is charged with a violent crime ... shackling essentially brands him as having a violent nature.” Larson, 515 F.3d at 1064 (internal alterations and quotation marks omitted). Hedlund was charged with the murders of Mertens and McClain, both violent crimes. Thus, shackling Hedlund was especially prejudicial because it risked causing the *843jury to infer that Hedlund was a dangerous naan with a propensity to act violently.
Lastly, because the state’s evidence against Hedlund was not overwhelming, the risk that shackling Hedlund prejudiced the jury is real. In Rhoden:
the basic issue at his trial concerned whether there was consent or whether Rhoden used force or fear to overcome his accuser’s will. The evidence on this issue was disputed and the jurors deliberated for over nine hours over three days, which suggests that they did not find the case to be clear-cut.
172 F.3d at 637. Evidence of Hedlund’s involvement in the crimes may be “overwhelming.” Cox v. Ayers, 613 F.3d 883, 891 (9th Cir.2010) (quoting Dyas, 317 F.3d at 937). However, as in Rhoden, one of the critical issues at trial, Hedlund’s mental state, was hotly disputed. Hedlund’s defense theory was that McKinney was the mastermind and that Hedlund was a reluctant participant. While the jury ultimately convicted Hedlund of the first-degree murder of McClain and the second-degree murder of Mertens, the evidence supporting Hedlund’s first-degree murder conviction was not overwhelming. Because the jury convicted McKinney of first-degree murder for the deaths of both Mertens and McClain, it is reasonable to conclude that the jury found McKinney more culpable than Hedlund. Thus, on the crucial issue of whether Hedlund’s culpability warranted a first-degree murder conviction, which triggered a possible death sentence, visibly shackling him prejudiced the jury against him.
Because the restraints were visible, because this case involves violent crimes, and because the evidence of Hedlund’s mental state was disputed and not overwhelming on the critical issue that would trigger a death sentence, the trial court’s error substantially influenced the jury’s verdict and warrants habeas relief. See Rhoden, 172 F.3d at 637.
“Because sentences of death are qualitatively different from prison sentences,” the U.S. Supreme Court has gone to “extraordinary measures to ensure that the prisoner sentenced to be executed is afforded process that will guarantee, as much as is humanly possible, that the sentence was not imposed out of whim, passion, prejudice, or mistake.” Eddings, 455 U.S. at 117-18, 102 S.Ct. 869 (O’Connor, J., concurring) (citations and internal quotation marks omitted). The process used to try, convict, and sentence Hedlund to death was unconstitutional because it increased, rather than decreased, the chances that Hedlund was sentenced to death out of “whim, passion, prejudice, or mistake.” Id. The Arizona courts unreasonably concluded otherwise. Even under the Supreme Court’s restrictive interpretation of AEDPA provisions, such constitutional violations entitle Hedlund to habeas relief.
. I agree that the remaining uncertified issues are not "debatable amongst jurists of reason,” Miller-El v. Cockrell, 537 U.S. 322, 336, 123 S.Ct. 1029, 154 L.Ed.2d 931 (2003), and therefore need not be discussed.
. As Chief Judge Kozinski recently predicted, Hedlund’s petition is yet another example of the state courts applying an unconstitutional causal nexus test before imposing and affirming a sentence of death. See Poyson v. Ryan, 743 F.3d 1185, 1188 (9th Cir.2013) (Kozinski, C.J., dissenting from the denial of rehearing en banc).
. As a consequence of Arizona’s state habeas system, see Ariz. R.Crim. P. 32.3, Hedlund’s petition for post-conviction relief was reviewed in superior court by the same judge who presided over his underlying criminal case.
. Hedlund lived with Shirley and James McKinney, Sr. from ages six to fourteen. *827When he was twelve years old, Shirley told him that he had been conceived during his biological mother’s affair with another man,
. Dissenting in Poyson v. Ryan, Judge Thomas accused the majority there of similarly distorting Schad's "clear indication" language "to create a new, more stringent test for determining whether a state court applied an unconstitutional causal nexus analysis." 743 F.3d at 1207 (Thomas, J., dissenting).
. Whether Hedlund must also show actual prejudice for the writ to issue is an unsettled question in the Ninth Circuit. Compare Stokley v. Ryan, 705 F.3d 401, 404 (9th Cir.2012) (applying harmless error review to the Arizona Supreme Court's alleged Eddings violation), and Landrigan v. Stewart, 272 F.3d 1221, 1230 & n. 9 (9th Cir.2001) (applying harmless error review to the state court’s failure to consider the defendant’s alleged intoxication and past history of drug use as a nonstatutory mitigating factor), adopted by Landrigan v. Schriro, 501 F.3d 1147 (9th Cir. 2007) (en banc) (order), with Williams v. Ryan, 623 F.3d 1258, 1270-72 (9th Cir.2010) (granting habeas relief for an Eddings violation without conducting harmless error analysis), and Styers, 547 F.3d at 1035-36 (same). The issue is also the subject of a split among other circuits. Compare Bryson v. Ward, 187 F.3d 1193, 1205 (10th Cir.1999) (collecting cases from multiple circuits applying harmless error review), with Nelson v. Quarterman, 472 F.3d 287, 314-15 (5th Cir.2006) (enbanc) (declining to apply harmless error review).
. Under Arizona law, the court “may, in its sole discretion, participate in settlement discussions” by directing counsel to "participate in a good faith discussion with the court” to reach a resolution that “conforms to the interests of justice.” Ariz. R.Crim. P. 17.4. By contrast, Federal Rule of Criminal Procedure 11(c)(1) expressly mandates that the district court "must not participate” in plea agreement discussions.
. The record is not ambiguous as to whether Hedlund had an offer in hand when his attorney moved to recuse Judge Sheldon. While insisting that the record is unclear, the majority accepts that defense counsel had an offer in hand to justify its conclusion that defense counsel’s actions were strategic. Therefore, there is no explanation, whether reasonable or not, that supports defense counsel’s abandonment of plea negotiations.
. While counsel’s motion to have Hedlund’s case moved before a different judge was a tactical decision, his failure to timely respond to the prosecutor’s second plea offer was not. Counsel’s decision to bring a recusal motion is not at issue here. Thus, the majority’s thorough analysis of why counsel’s decision to move for recusal was within the "wide range of reasonable professional assistance” is off point. Strickland, 466 U.S. at 689, 104 S.Ct. 2052. The relevant inquiry, instead, is whether defense counsel’s abandonment of ongoing plea negotiations with the prosecutor and the judge, when a second plea agreement was nearly complete, constituted deficient performance.
. The state PCR court's decision was also contrary to Arizona law because "a petitioner is not required to prove that the trial court would have accepted the plea agreement in order to establish prejudice resulting from counsel's deficient advice.” State v. Donald, 198 Ariz. 406, 10 P.3d 1193, 1202 (Ct.App. 2000) (holding that "such a requirement ‘presents inherent problems of proof,’ and that it would be 'unfair and unwise to require litigants to speculate as to how a particular judge would have acted under particular circumstances' ”) (citations omitted).
. The majority also relies on arguments not found in the state PCR court's decision to conclude that counsel’s performance was not deficient. Although we are required to evaluate all reasonable arguments that a state court could have relied upon when assessing a state high court’s summary denial of original habeas petitions, see Richter, 131 S.Ct. at 784, we should not evaluate such hypothetical arguments when assessing a state court's denial of a habeas petition that actually relied on a lower court’s reasoning. Instead, we "must 'look through' that judgment to the last reasoned state-court decision on the merits.” Cannedy v. Adams, 706 F.3d 1148, 1157 (9th Cir.2013). Thus, we should look only to the state PCR court's actual reasoning to determine whether it was an unreasonable application of clearly established Supreme Court case law.
. The majority insinuates that it was Hedlund's fault that the jurors saw the shackles. But Hedlund’s constitutional right to a fair trial rests in the Fourteenth Amendment, which the courts, not Hedlund, are required to protect. Estelle, 425 U.S. at 503, 96 S.Ct. 1691. Therefore, Hedlund’s efforts, or alleged lack thereof, to keep his restraints concealed are irrelevant.
. The two decisions that we relied on in Hamilton involved much graver threats of violence and escape. See Loux v. United States, 389 F.2d 911, 919 (9th Cir.1968) (holding that shackling defendants is permissible where one of the defendants had twice successfully escaped from prison and lost an arm in an unsuccessful attempt, another had escaped from prison five times, and the third defendant had escaped from prison three times and was then serving a sentence for second-degree murder); see also Stewart v. Corbin, 850 F.2d 492, 497-98 (9th Cir.1988) (holding that shackling defendant is permissible where defendant had a prior felony conviction for escape while in police custody, had an outstanding arrest warrant for another escape in custody, had physically assaulted officers in the courtroom, and had threatened a judge and an attorney).