CHIEF JUDGE HUG, JUDGE PREGERSON, JUDGE REINHARDT concurring; JUDGE BROWNING, JUDGE TASHIMA, JUDGE THOMAS concurring in parts I, II and IV, and in the judgment.
Thomas Martin Thompson’s execution is scheduled for August 5, 1997, for the rape and murder of Ginger Fleischli. We recall the mandate that issued on the March 6, 1997, amended opinion in Thompson v. Calderon, 109 F.3d 1358 (9th Cir.), cert. denied, — U.S. -, 117 S.Ct. 2426, 138 L.Ed.2d 188 (1997), grant Thompson’s petition for a writ of habeas corpus in part, remand in part, and direct that the order scheduling his execution for August 5, 1997, be vacated.
On March 29, 1995, the district court granted in part Thompson’s first federal ha-beas petition. It found that Thompson’s trial counsel provided constitutionally deficient representation in failing to refute the rape evidence and in failing to impeach the testimony of a notoriously unreliable jailhouse informant. The district court ordered that Thompson receive a new trial on the rape conviction and the rape special circumstance finding and vacated his death sentence.
The panel reversed the district court’s grant of habeas relief. It concluded that although trial counsel may have been ineffective, any deficient performance did not prejudice Thompson’s defense to the rape and rape special circumstance charges. It affirmed the district court’s denial of relief on Thompson’s other claims, including his claim of prosecutorial misconduct. Due to procedural misunderstandings within our court, no en banc call was made. The United States Supreme Court denied Thompson’s petition *1048for a writ of certiorari, and on June 11,1997, our court’s mandate issued.
On July 22,1997, Thompson filed an emergency motion with this court requesting that the panel recall its mandate and reconsider its decision.1 The panel denied the motion. A majority of the active, nonreeused judges of this circuit then voted to reconsider en banc whether to recall the mandate. Because of the exceptional circumstances surrounding our handling of the appeal from Thompson’s first habeas petition and because we are convinced that the panel committed fundamental errors of law that would result in manifest injustice, we now sua sponte recall the mandate and reverse the panel’s decision. We affirm the part of the district court order granting Thompson’s writ with respect to the rape conviction and rape special circumstance charge, and vacating his death sentence. We remand for further review with respect to the first-degree murder conviction, and vacate the State’s order of execution.
I. MOTION TO RECALL THE MANDATE
A.
We have the power to recall the mandate of a final decision of our court, and to do so sua sponte. See Malik v. Brown, 65 F.3d 148, 149 (9th Cir.1995). We have long recognized the power to recall the mandate as a means of protecting the integrity of our processes and decisions. Perkins v. Standard Oil Co., 487 F.2d 672, 674 (9th Cir. 1973). The decision whether to recall the mandate “is entirely discretionary with this court,” Feldman v. Henman, 815 F.2d 1318, 1322 (9th Cir.1987), and the Supreme Court will review that decision, if at all, only for an abuse of discretion. See Hawaii Housing Auth. v. Midkiff, 463 U.S. 1323, 1324, 104 S.Ct. 7, 8, 77 L.Ed.2d 1426 (1983) (Rehnquist, J., in chambers).
Recalling a mandate is an extraordinary remedy and we will exercise our authority to do so only in exceptional circumstances, such as when it is necessary in order to prevent injustice. Zipfel v. Halliburton Co., 861 F.2d 565, 567-68 (9th Cir.1988). We have not defined or limited the exceptional circumstances that may warrant recalling the mandate. We have recalled the mandate most frequently when an intervening statutory change or Supreme Court decision has undermined the basis of our decision. See, e.g., Malik, 65 F.3d at 149 (statute); United States v. Davis, 36 F.3d 1424, 1429-30 (9th Cir.1994) (Supreme Court decision); Bryant v. Ford Motor Co., 886 F.2d 1526, 1529-30 (9th Cir.1989) (statute); Zipfel, 861 F.2d at 567-68 (Supreme Court decision). Exceptional circumstances warranting recall of the mandate also exist when recall is necessary to prevent an erroneous ruling from working an injustice. See Nevius v. Sumner, 105 F.3d 453, 461 (9th Cir.1996).
B.
Our interest both in protecting the integrity of our processes and in preventing injustice are implicated in the case before us. The circumstances here are exceptional for a number of reasons, individually and collectively. First, our normal en banc process did not function in the intended manner. But for procedural misunderstandings by some judges of this court, an en banc call would have been made and voted upon at the ordinary time. Second, in reversing the district court, the panel appears to have made fundamental errors of law that, if not corrected, would lead to a miscarriage of justice. The consequence of our failure to act would be the execution of a person as to whom a grave question exists whether he is innocent of the death-qualifying offense, the alleged rape, and whose conviction on the first-degree murder charge may be fundamentally flawed. This is a person who has never before been convicted of a crime. Under these circumstances, we have an obligation to recall the mandate in order to preserve the integrity of the judicial process.
*1049We must make it clear at the outset that we act to review the panel’s decision not because of any new evidence offered by Thompson nor because of any new argument or ground for relief he suggests; nor do we act because Thompson presents any second or successive claim for relief. Although, in the interests of comity, we delayed acting until Thompson’s state court proceedings had concluded, we act now because we have sua sponte concluded that the interests of justice require us to afford Thompson the en banc process we should have initiated immediately after the panel’s decision.
Our normal procedure when we, as a court, believe a three-judge panel has erred, is to call for a vote on whether to rehear the case en banc. No active judge requested an en banc vote when the panel’s initial opinion was filed in September 1996. However, when the panel filed its amended opinion on March 6, 1997, two judges of this court advised the panel that they wished to call for an en banc vote. A series of misunderstandings led those judges to believe that the time for an en banc vote had passed, and that they were prohibited from making an en banc call at that time.2
Because of misunderstandings or errors by members of the court, no en bane vote was taken when it ordinarily would have occurred. Instead, the panel stayed issuance of its mandate while Thompson sought a writ of certiorari from the Supreme Court, and upon denial of his certiorari petition, the mandate issued. A sua sponte request to consider en banc whether to recall the mandate was made shortly thereafter, even before the mandate was spread in the district court. Through a consultative process, the court decided to postpone action on whether sua sponte to recall the mandate until after Thompson concluded his state court habeas proceedings and filed a further request for review in this court.3
Among the most important and consistent themes in ... death penalty jurisprudence is the need for special care and deliberation in decisions that may lead to the imposition of that sanction. The [Supreme] Court has accordingly imposed a series of unique substantive and procedural restrictions designed to ensure that capital punishment is not imposed without the serious and calm reflection that ought to precede any decision of such gravity and finality.
Thompson v. Oklahoma, 487 U.S. 815, 856, 108 S.Ct. 2687, 2710, 101 L.Ed.2d 702 (1988) (O’Connor, J., concurring). We now provide for the first time en banc review of Thompson’s first habeas petition — a review which became available to him only a brief period ago. The review we now provide Thompson is available to every habeas petitioner on his first petition upon majority vote of this court. It is a part of the full and fair procedure our court affords in connection with initial habeas petitions.
We reiterate that Thompson is not responsible for the lateness of the hour. When we recognized our own procedural error with respect to this first habeas petition, we did not act immediately, but determined, in the interests of comity, to wait until the state court proceedings were concluded and Thompson sought to return to federal court. Thompson returned at the earliest possible time, and we then initiated en banc rehearing on his first habeas petition. Thus neither the Anti-Terrorism and Effective Death Penalty Act of 1996, 28 U.S.C. § 2244(a) & (b) (“AEDPA”), nor McCleskey v. Zant, 499 U.S. 467, 111 S.Ct. 1454, 113 L.Ed.2d 517 (1991), is implicated here. Judge Hall's suggestion that we are opening the floodgates to frivolous last-minute petitions is clearly in error. We are acting not upon the basis of Thompson’s petition, but upon the basis of our sua sponte determination to remedy our own errors. AEDPA and McCleskey are aimed at abusive efforts by petitioners to file second or successive petitions, not at a court’s sua sponte determination that it must act to “prevent injustice [and] to protect the integ*1050rity of its process.” Bryant, 886 F.2d at 1529.
We may not use our recall power “ ‘simply as a device for granting late rehearing.’” Moran v. McDaniel, 80 F.3d 1261, 1267 (9th Cir.1996) (quoting Johnson v. Bechtel Assocs., 801 F.2d 412, 416 (D.C.Cir.1986)). But that is not what happened here. The failure to call for en banc at the regular time in a capital case involving serious constitutional issues is in itself exceptional in our court. In fact, it is unprecedented. Indeed, there has been no occasion in the past several decades in which this court has made a similar error in any type of case. Yet our mistake should not prevent a capital habeas petitioner from receiving full and fair consideration of his claims by an en banc court, particularly in a ease in which it appears that the panel decision may result in a miscarriage of justice.
The district court granted Thompson partial habeas relief on the ground that his counsel was unconstitutionally ineffective and that the ineffective assistance prejudiced the outcome of Thompson’s trial. The panel reversed. An erroneous reversal alone may not be exceptional. But here, Thompson’s trial counsel’s ineffectiveness, as found by the district court, directly affects the issue of Thompson’s possible innocence of the rape, and raises the question whether, but for the ineffectiveness, the jury would have acquitted Thompson of the rape charge and special circumstance upon which his death sentence is predicated.
The district court, after an extensive evidentiary hearing, considered Thompson’s habeas petition for a prolonged period before rendering a decision granting partial relief. Its thorough and well-reasoned opinion discussed in detail trial counsel’s failure to present an effective defense against the State’s charge that Thompson raped Fleischli, or even to question whether a rape occurred. The district court found that trial counsel provided an ineffective defense to the State’s forensic evidence of rape and that counsel’s deficient performance in that respect prejudiced the verdict. See Strickland v. Washington, 466 U.S. 668, 693-94, 104 S.Ct. 2052, 2067-68, 80 L.Ed.2d 674 (1984). The district court also determined that Thompson’s counsel was ineffective in fading to investigate the backgrounds of two jailhouse informants who testified that Thompson confessed to the rape, and in failing to present readily available evidence undermining the informants’ testimony. The district court concluded that “the evidence of rape was not substantial” and that the failure to impeach at least one informant also prejudiced the verdict.
The panel reversed, relying primarily on the prejudice prong of the Strickland test. It concluded that counsel’s assumed incompetence in failing to rebut the State’s forensic evidence and to investigate and impeach witnesses whose history gravely undermines the reliability of their testimony did not affect the outcome of Thompson’s trial. It reached this conclusion although the district court had explained carefully and persuasively, on the basis of its credibility determinations and assessment of the physical evidence, why the result probably would have been different with respect to the rape charge and the rape special circumstance finding. Moreover, both areas of ineffectiveness relate to the alleged rape, and Thompson’s death sentence rests exclusively upon the existence of the rape conviction and the rape circumstance. At the very least, questions of the most serious nature exist as to the panel’s reversal of the district court’s finding that counsel’s ineffectiveness raised a reasonable probability that the jury convicted Thompson of a rape he did not commit and then erroneously sentenced him to die for a capital offense.
In addition, a serious question exists as to whether Thompson was deprived of due process of law by the prosecutor’s presentation of flagrantly inconsistent theories, facts, and arguments to the two juries that separately heard Thompson’s case and that of his co-defendant, David Leitch. While the district court concluded that no constitutional violation occurred, our review of the record persuades us that the prosecutor’s tactics may well have resulted in Thompson’s receiving a fundamentally unfair trial. These tactics, including the use of the two highly dubious jailhouse informants, appear to compound the constitutional violations that flow from defense counsel’s ineffective performance.
*1051When we examine the totality of the circumstances (both substantive and procedural), the grave questions that exist regarding Thompson’s innocence of any capital offense, the likelihood that the panel’s decision is erroneous, and the consequences that would flow from allowing an erroneous decision to go unreviewed, we are persuaded that the circumstances involved are extraordinary.
C.
Before we may recall the mandate, we confront a chicken and egg dilemma. In most cases, to justify recalling the mandate, we must conclude that there is substantial doubt as to whether the panel erred and that recall is necessary to prevent manifest injustice. By the same token, we cannot reexamine the panel’s decision and rule on the merits without having recalled the mandate. This process is not unlike determinations in respect to standing; the court must peek under the corner of the rug at the merits to determine what is at stake before knowing whether the parties have standing enabling the court to exercise its jurisdiction to determine the merits. Our inquiry here thus has a two-fold objective: First, we must determine whether the panel may have been wrong, and whether if we fail to act we may be perpetrating a manifest injustice; if so, we must recall the mandate. Second, once we do, we must determine on the merits whether habeas relief should be granted.
We recognize, as Judge Hall’s dissent notes, that our court recently has rejected a petitioner’s motion to recall the mandate which was clearly an effort to circumvent the procedural bars embodied in the AEDPA. See Nevius, 105 F.3d at 461-62. But the dissent overlooks that in Nevius, we explicitly stated that “Nevius wants us to recall our mandate ... not to nullify an erroneous decision, but to reopen the proceeding so that he may present new claims that now cannot be addressed in a subsequent petition.” Id. at 461 (emphasis added). This statement recognizes the principles embodied in the cases we cite here; i.e., that the mandate may be recalled to correct an erroneous decision which would result in a manifest injustice.
Moreover, our rejection of the motion to recall the mandate in Nevius, and in other death penalty appeals, see, e.g., Greenawalt v. Stewart, 105 F.3d 1268, 1278 (9th Cir. 1997), merely substantiates that here, exceptional circumstances caused a majority of the active judges of this court to vote in favor of allowing an en banc court to determine whether to recall the mandate.
Our initial examination of this ease leaves us with “a firm belief that this is the exceptional case requiring recall of the mandate in order to prevent an injustice.” Verrilli v. City of Concord, 557 F.2d 664, 665 (9th Cir. 1977). We believe just as strongly that it is necessary to do so in order to protect the integrity of our process. For all of the reasons we have heretofore discussed, we order the mandate recalled.
Only an en banc panel court may overrule an earlier decision by a panel of this court. See Fed. R.App. P. 35(a)(1); Ninth Cir. R. 35-1. A majority of the active judges of this circuit has voted to rehear en banc the panel’s refusal to recall the mandate, and this en bane court has now determined that it shall be recalled. Implicit in our power to do so is our authority to reconsider the merits of the case as an en banc court. See Songer v. Wainwright, 769 F.2d 1497 (11th Cir.1985) (after panel denied motion to recall the mandate, a majority of the court voted to recall the mandate and reconsider the appeal en banc); Uzzell v. Friday, 625 F.2d 1117, 1122 (4th Cir.1980) (Widener, J., dissenting) (court on its own motion recalled the mandate of the panel decision and ordered rehearing en banc on the merits). Thompson’s appeal from the district court’s decision, and the State’s cross-appeal, are now properly before this en banc court, and we turn to the merits of those appeals.
II. INEFFECTIVE ASSISTANCE OF COUNSEL
We now review the district court’s partial grant of Thompson’s habeas petition. Sanders v. Ratelle, 21 F.3d 1446, 1451 (9th Cir. 1994). We must consider the district judge’s determinations on both the performance and prejudice prongs of the ineffective assistance *1052of counsel test. See Strickland, 466 U.S. at 698, 104 S.Ct. at 2070. Although Thompson must rebut the “strong presumption” that his trial counsel provided adequate assistance and exercised reasonable professional judgment in representing him, id. at 690, 104 S.Ct. at 2066, the district court’s factual findings are entitled to the usual deference. Hendricks v. Calderon, 70 F.3d 1032, 1036 (9th Cir.1995) (noting that although an ineffectiveness claim presents a mixed question of law and fact, reviewed de novo, the district court’s factual findings should be accepted unless clearly erroneous). We find ineffective assistance only when counsel’s performance falls below a reasonable standard of professional representation and when counsel’s deficient representation undermines our confidence in the outcome of the trial. Strickland, 466 U.S. at 694, 104 S.Ct. at 2068; Clabourne v. Lems, 64 F.3d 1373, 1378 (9th Cir.1995).
A. Failure to rebut forensic evidence of rape
To support the rape case against Thompson, the State called the coroner, Dr. Richards, to testify. Richards testified that in his autopsy of Fleischli he found bruises on her ankles, palms, wrists, right elbow, and right knee, as well as a “crushing injury” to her right wrist. Richards testified that except for the right knee, the injuries had all occurred at or near the time of death and were the result of heavy handling. Richards also testified that he found no vaginal tearing or bruising, but that such tearing or bruising is not necessary to indicate rape.
Thompson’s trial counsel, Ronald Brower, cross-examined Richards only on the question of whether there was vaginal tearing. He did not cross-examine Richards at all regarding the bruises, although on redirect examination Richards stated that evidence of bruising to the body could indicate rape. Richards also testified that he found no “anatomical” evidence of rape, but that sperm and semen analysis might provide such evidence. Brower did not cross-examine Richards to clarify these statements.
At the evidentiary hearing before the district court, Thompson presented the testimony of Dr. Root, an independent pathology expert in whom Brower had consulted, but had not called as a witness. Root testified that he had reviewed the autopsy report and that, if called, he would have testified that the bruises described in the report were several weeks old, with the exception of the wrist injury. Root opined that the bruises on Fleisehli’s hands were the result of livor-mortis, not injury, and that the time of the right wrist injury could not be estimated accurately because of the lack of cellular response.
Root further testified that the coroner’s report showed that there was no semen drainage in the crotch of Fleischh’s tight jeans, which were zipped and belted and which she wore without underpants. Fleischli’s vaginal swab revealed recent semen with “infrequent” sperm. Both the lack of drainage and the infrequent sperm suggested that Fleischli had douched or washed after sex, consistent with consensual sex but not with rape fohowed by murder.
Brower testified at the evidentiary hearing that he had not pursued further consultations with Root because he feared that if he decided not to use Root as a witness, the prosecutor would find out and use that fact against him. The district court soundly rejected this specious argument, reminding Brower that the prosecutor could not have introduced evidence that Brower had consulted with but not called an independent pathologist. Brower also testified that he cross-examined Richards only on the vaginal tearing issue, and not on the bruising, because Richards earlier had informed Brower that there was “no anatomical evidence” of rape, and Brow-er wanted to make sure Richards testified consistently with that opinion.
Brower admitted that he knew the prosecution would argue that the bruising evidenced rape, but he testified that he planned to argue to the jury that David Leitch, not Thompson, inflicted those bruises. Brower testified that he did not think it was “necessary” to rebut Richards’ evidence about the bruises. Finally, Brower testified that he did not recall learning of the lack of drainage in Fleischh’s jeans or of the significance of that evidence. Brower somehow found it *1053unnecessary to pursue readily available evidence that would have undermined the State’s rape ease because his theory was that Leiteh raped Fleischli, not that no rape occurred.
The district court found, and the panel assumed, that Brower’s failure to investigate, develop, and present evidence rebutting the State’s forensic evidence of rape fell below a reasonable standard of professional representation. As we have noted, the district court found that defense counsel “had no valid tactical reason for not attacking the rape evidence.” Like the district court, we cannot find that Brower’s purported choice was professionally reasonable.
Brower’s theory that Leiteh raped Fleisch-li was inconsistent with his reliance on the coroner’s statement that there was “no anatomical evidence of rape.” Brower irrationally devoted much of his defense to this theory, rather than advancing the theory that no rape had taken place. The existence of a rape conviction exposed Thompson to the death penalty on two possible grounds: felony murder and murder with rape as a special circumstance. By not rebutting the prosecution’s rape evidence, Brower unnecessarily risked Thompson’s life, ostensibly in order to pursue an alternative theory of the case.4 We agree with the district court that Brower’s performance in this respect was constitutionally deficient.
We also conclude, as did the district court, that Brower’s failure to attack the rape evidence was prejudicial. Richards’ testimony provided the State’s strongest evidence on the rape charge. As the district court found, if Brower had “negated the rape, he would have cast considerable doubt on Thompson’s role in the murder____ Without the rape, theory, Thompson had no obvious motive to kill Fleischli.” Had Brower impeached the coroner’s testimony, the State would have been left mainly with evidence that Fleisch-li’s vaginal swab contained semen consistent with Thompson’s blood type and Thompson’s testimony that he had consensual sex with Fleischli on the night of her death.5 This evidence could not have sustained a rape conviction.6 Thus, absent Brower’s ineffectiveness, the outcome of Thompson’s trial would in all probability have been different— he most likely would not have been convicted of the rape or the rape special circumstance, and therefore could not have been sentenced to die. Had Brower adequately challenged the State’s rape charge, he would have “put the whole ease in such a different light as to undermine confidence in the verdict.” Kyles v. Whitley, 514 U.S. 419, 435, 115 S.Ct. 1555, 1566, 131 L.Ed.2d 490 (1995). We can think of no error more prejudicial than one that is the precipitating cause of an erroneous death sentence.
B. Failure to impeach Fink
At Thompson’s trial, the State called Edward Fink, a jailhouse informant, who testified that Thompson had admitted raping and murdering Fleischli. Fink testified that he had come forward with Thompson’s confession on his own initiative and that in exchange for his statement he asked only that he be allowed to serve his parole violation sentence in Chino. Fink also admitted to four prior felony convictions.
Brower cross-examined Fink and elicited that Fink had spoken twice with law enforcement officers about Thompson’s case. Brow-er also produced Fink’s rap sheet, which listed his additional felonies and numerous aliases. Brower also inquired into Fink’s history of drug abuse, and got Fink to admit that he was known as an informant.
*1054There was, however, far more to the story of perennial informant Fink. At the eviden-tiary hearing, Thompson produced Fink’s extensive history as an informant who frequently claimed that fellow inmates confessed their crimes to him while awaiting trial and who frequently received favors in exchange for cooperating with the police. Joel Baruch, a defense attorney, testified that in another capital case, in which he represented the defendant, he had discovered Fink’s informant history, including statements in Fink’s file by law enforcement officers that Fink was an unreliable informant, a con man, and a heroin addict. Baruch testified that after he repeatedly advised the prosecutor in his case that he was well aware of Fink’s background, the case was dismissed. Thompson also offered a transcript of Fink’s testimony in People v. Goldstein, a 1980 California murder trial. Goldstein’s counsel cross-examined Fink at length, and he undermined Fink’s credibility by showing that Fink regularly informed to law enforcement officers in exchange for future favors and probably had fabricated Goldstein’s confession.
Brower testified at the evidentiary hearing that he had stopped investigating Fink’s background before trial because he believed he had enough material to cross-examine Fink, and that he had stopped cross-examining him because he thought the judge was getting restless and the jury had “gotten the message.” Like Brower’s excuse for not calling an independent forensic expert, this feeble explanation underscores Brower’s ineffectiveness. Brower readily could have obtained evidence about Fink’s background as an informant and the favors he received in exchange for his frequent testimony. Of particular relevance to Thompson’s case was the fact that Fink was released from his parole hold soon after he provided the information that incriminated Thompson. Rather than investigating Fink’s background or requesting discovery from the prosecutor regarding Fink, Brower focused on Fink’s drug use and on whether Fink in fact had been in sufficient proximity to Thompson to obtain the alleged confession. Given the crucial import of Fink’s testimony, Brower’s conduct in failing adequately to impeach him fell well below a reasonable standard of professional representation.
Brower’s failure to investigate and impeach Fink severely prejudiced Thompson. Brower could have destroyed Fink’s credibility. Without Fink’s testimony, the State’s rape ease against Thompson would have been dramatically weaker. We are in accord with the district court; Brower’s ineffective assistance in defending against the forensic evidence and against Fink’s testimony fatally undermines our confidence in the outcome of Thompson’s rape conviction. See Kyles, 514 U.S. at 434-35, 115 S.Ct. at 1566. Without the rape finding, Thompson could not have been sentenced to death. Brower’s ineffective performance therefore prejudiced not only the rape conviction and the rape special circumstance, but the ultimate outcome: the finding of death eligibility and the imposition of a capital sentence.
C. Failure to impeach Del Frate
A second informant, John Del Frate, also testified at Thompson’s trial. The district court concluded that Brower readily could have discovered and presented substantially more impeaching evidence against Del Frate. In particular, Brower failed to show that Del Frate’s testimony that Thompson said he stabbed Fleisehli in the torso and buried her in a “shallow grave” was both false as to what happened to Fleisehli and, even more important, parroted almost verbatim inaccurate news reports. Brower also failed to discover that Del Frate had served as an informant since the age of fourteen, that two police agencies for whom Del Frate informed considered him unreliable, that Del Frate’s family considered him to be a pathological liar, and that Del Frate had shared a cell with Leitch for several weeks before coming into contact with Thompson. The panel’s conclusion that “Brower could hardly have impeached Del Frate more than he did” is patently wrong.
Brower’s failure to investigate and impeach Del Frate, whose testimony he admitted was important to the prosecution’s case, was unconstitutionally deficient. It also was prejudicial. As the prosecutor himself stated *1055in his closing argument, Fink and Del Frate’s testimony provided the dispositive evidence that Thompson had raped Fleischli and murdered her to cover up the rape. Had Brower impeached Del Frate, as well as Fink, the prosecution’s rape case against Thompson would have been substantially weakened. Brower’s cumulative errors in failing to investigate and impeach the jailhouse informants and to rebut the forensic evidence of rape cast grave doubt on the reliability of the rape conviction and the rape special circumstance finding, and thus of the death sentence itself.
For all of the reasons we have previously discussed, we grant Thompson’s writ of habe-as corpus with respect to the rape count, the rape special circumstance allegation, and the capital sentence.
III. PROSECUTORIAL MISCONDUCT
There is an alternate ground that also requires the relief we order. In his habeas petition before the district court, Thompson claimed that the prosecutor’s use of fundamentally inconsistent theories at his and Leitch’s trials violated due process. The district court rejected this argument, and the panel affirmed that conclusion. Having carefully reviewed the record, and having given due deference to the district court’s analysis, we conclude that we must reverse on this ground as well.
A. The pretrial phase
At the preliminary hearing, when Thompson and Leiteh were still joined as co-defendants, Deputy District Attorney Daniel Brice asserted that Leiteh wanted Fleischli dead, and that he enlisted Thompson to join in the killing. In support of this theory, the prosecution presented the testimony of four jailhouse “informants” who claimed to have obtained confessions from Thompson: David Vogel, David Wright, Timothy Gravelle, and Robert Evans. Each informant told essentially the same story: Leiteh wanted Fleisch-li dead for interfering with his attempts to reconcile his relationship with his ex-wife, Tracy Leiteh, and, therefore, had recruited Thompson to help him kill her. According to Vogel’s testimony, Thompson had told him that on the night of the murder, he engaged in consensual sex with Fleischli.7 Then, when Leiteh returned home, the two executed Leitch’s “plan,” and killed Fleischli. The State charged both men with capital murder based on rape and murder of Fleischli.
At the close of the preliminary hearing, the magistrate found that there was insufficient evidence to hold Thompson and Leiteh on the rape charge and the rape special circumstance allegation. Nonetheless, the State filed an information charging Thompson and Leiteh, inter alia, with murder, rape, and rape as a special circumstance.
After the information was filed, Deputy District Attorney Michael Jacobs replaced Brice as lead counsel for the prosecution. Jacobs then reviewed all of the tapes, police reports, forensic reports, and court transcripts in the ease. Those records, among other things, described Leitch’s violent history, Leitch’s past threats toward Fleischli, Leitch’s desire to reconcile with his ex-wife8 and his perception that Fleischli was an impediment to that goal, physical evidence that tied Leiteh to the crime, and Leiteh’s inconsistent statements to police following the crime.
Upon reviewing the evidence, Jacobs concluded that he would continue to pursue murder-with-special-eircumstanee convictions against both defendants, on the theory that Leiteh had the motive for killing Fleischli, and that both men were equally culpable for raping and murdering her. On his own admission, Jacobs never altered his view of the motive and the crime, either before or after he won Thompson’s conviction.
Some months later, Thompson and Leiteh moved under California Penal Code § 995 to set aside the rape charge and the rape special circumstance allegation as unsupported *1056by probable cause. At the hearing, at which the record shows Jacobs to have been present, one of his assistants represented to the court that “[Leitch] is the only person throughout the record who has a motive____ According to [Tracy Leitch,] David Leitch was negotiating a reconciliation with Tracy Leitch and Ginger Fleischli was not reacting very cooperatively to that circumstance, and there is the motive — she was in the way.” The trial court denied the defendants’ motion without comment.
Leitch then moved successfully to sever his trial from Thompson’s. The prosecutor elected to try Thompson first.9
B. The trials of Thompson and Leitch
The prosecutor’s pursuit of fundamentally inconsistent theories is evident from the transcripts of the two trials. In the second trial, the prosecutor presented essentially the same theory used at the pretrial proceedings. Yet, in between those court proceedings, the prosecutor convicted Thompson under an entirely different theory, and argued critical facts to Thompson’s jury that were at odds with those presented both in the preliminary hearing and Leitch’s trial.
The prosecutor presented markedly different and conflicting evidence at the two trials. At Thompson’s trial, the prosecutor did not call any of the four jailhouse informants who had testified at the preliminary hearing and claimed that Thompson had confessed to them. Instead, he called two new jailhouse informants, Fink and Del Frate, who also claimed that they were the recipients of Thompson’s confession, but who told an entirely different and factually inconsistent story. Contrary to the testimony of the earlier informants, the two new informants testified that Thompson had told them that he had raped Fleischli and had killed her before Leitch returned home, and that he killed her in order to prevent her from reporting the rape.10 Fink further testified on direct examination that Leitch found out about the murder only when he returned home, and that he merely assisted Thompson in disposing of the body because Thompson convinced him that they should go to Mexico together. The testimony of Fink and Del Frate provided the only direct evidence that (1) Thompson killed Fleischli, (2) Fleischli was raped, and (3) Thompson raped her. The prosecutor termed the two informants’ testimony “dispositive” and “very, very damaging” because they had “no reason whatsoever [to] he.” The latter statement is patently untrue, as demonstrated by the fact that shortly after Fink incriminated Thompson, his parole hold was dropped and he was released from jail on the basis of favorable information provided by law enforcement officials.
The prosecutor did not call either Fink or Del Frate at Leitch’s trial. Instead, he called defense witnesses from Thompson’s trial (most of whom he had subpoenaed for Leitch’s trial immediately after objecting to their testimony at Thompson’s trial). These witnesses testified about Leiteh’s violent disposition, Leitch’s threats toward Fleischli, and Leiteh’s motive to kill Fleischli. The prosecutor relied heavily on their testimony to establish Leitch’s motive for the murder.
C. Closing arguments
The glaring inconsistency between the prosecutor’s theories, arguments, and factual representations at the two trials is apparent when one juxtaposes his closing arguments.
At Leitch’s trial, the prosecutor argued that:
[Leitch is] the only one, before the victim’s death, who expressed any hatred for her and the only one with any motive for her death.
[Leiteh’s desire to reunite with his ex-wife is] really the only motive we have in this case, and people have killed for less.
*1057[Leitch’s] was the only motive or reason for her demise.
(Emphasis added).
Yet at Thompson’s trial, the prosecutor advanced a contrary motive for Fleischli’s death. He asserted that the sole motive was Thompson’s desire to cover-up the alleged rape:
Why did Ginger Fleischli die? Because she said she was going to tell for what he [Thompson] did to her. So he killed her.
What did both Mr. Fink and Del Frate say? What do both of their statements have in common? [Thompson] didn’t want any witnesses.
The defendant said he raped the victim and killed her to prevent being caught for rape.
(Emphasis added).
At Leitch’s trial, the prosecutor adhered to the State’s original theory that Leitch was an active participant in the murder:
And I want to make it very clear that the State has charged Mr. Leitch with first degree murder, and with special circumstances. And as far as we are concerned, he’s guilty of that or he’s guilty of nothing.
The problem is, all of the evidence we have incriminates Mr. Leitch, at best, equally, and more so than Mr. Thompson.
... Both men were together inside that apartment with Ginger Fleischli.
So we have to ask ourselves, why would Mr. Thompson murder Miss Fleischli alone in an apartment where he lived, with no transportation, no means to move the body and wait for Mr. Leitch to come home to be an A-l witness for the murder of his ex-girlfriend? Is that reasonable or logical? Do you think that’s what happened?
You think Mr. Thompson did this all by himself and waiting for this good guy to come home so he could see him standing over his ex-girlfriend, who he lived with ten days before? No, it didn’t happen that way.
(Emphasis added).
The prosecutor, however, asserted as the truth before Thompson’s jury the story he subsequently labeled absurd and incredible in Leitch’s trial:
... [Thompson] was the only person in that apartment with Miss Fleischli the night — at the time she was killed.
We have the evidence that establishes Mr. Thompson alone in an apartment with a girl who is raped and murdered.
The David Leitch involvement — we have all this stuff on the board about him saying bad things about his girlfriend. What evidence do we really have that he did anything, had any part except that his car was used to move the body and that his shoe print was at the scene? There is no evidence we have putting him in the apartment that night.
(Emphasis added).
D. Post-trial theory
Not surprisingly, the prosecution’s original theory continued to prevail after the trials were concluded. The prosecutor submitted the following statement for use in Leitch’s future parole proceedings: “Fleischli was bound and gagged ... by [Leitch and Thompson] ... and was then stabbed ... by one or both of the defendants.” He made no mention of any rape.
E. Due process analysis
We are left with a picture of the prosecutor’s temporary abandonment during Thompson’s trial of the theory he presented, and supported with evidence, at the preliminary hearing, in the pretrial motions, and again at and after Leitch’s trial. The prosecutor manipulated evidence and witnesses, argued inconsistent motives, and at Leitch’s trial essentially ridiculed the theory he had used to obtain a conviction and death sentence at Thompson’s trial. The question before us is whether this prosecutorial misconduct violat*1058ed Thompson’s right to due process and a fair trial.
The Supreme Court has long emphasized our Constitution’s “overriding concern with the justice of finding guilt.” United States v. Agurs, 427 U.S. 97, 112, 96 S.Ct. 2392, 2401, 49 L.Ed.2d 342 (1976). In particular, the Due Process Clause guarantees for every defendant the right to a trial that comports with basic tenets of fundamental fairness. Lassiter v. Department of Soc. Servs., 452 U.S. 18, 24-25, 101 S.Ct. 2153, 2158-59, 68 L.Ed.2d 640 (1981); see also Turner v. Louisiana, 379 U.S. 466, 471-72, 85 S.Ct. 546, 549-50, 13 L.Ed.2d 424 (1965); In re Murchison, 349 U.S. 133, 136, 75 S.Ct. 623, 625-26, 99 L.Ed. 942 (1955).
The prosecutor, as the agent of the people and the State, has the unique duty to ensure fundamentally fair trials by seeking not only to convict, but also to vindicate the truth and to administer justice. The Supreme Court recognized, over sixty years ago, that
[bjecause the prosecutor is in a peculiar and very definite sense the servant of the law, the twofold aim of which is that guilt shall not escape or innocence suffer ..., [i]t is as much his duty to refrain from improper methods calculated to produce a wrongful conviction as it is to use every legitimate method to bring about one.
Berger v. United States, 295 U.S. 78, 88, 55 S.Ct. 629, 633, 79 L.Ed. 1314 (1935), overruled on other grounds by Stirone v. United States, 361 U.S. 212, 80 S.Ct. 270, 4 L.Ed.2d 252 (1960).
This court has reaffirmed that this duty of furthering just convictions “is [the prosecutor’s] highest purpose.” Bruno v. Rushen, 721 F.2d 1193, 1195 (9th Cir.1983). In United States v. Kojayan, 8 F.3d 1315, 1323 (9th Cir.1993), we further stated: “While lawyers representing private parties may — indeed, must — do everything ethically permissible to advance their clients’ interests, lawyers representing the government in criminal cases serve truth and justice first. The prosecutor’s job isn’t just to win, but to win fairly, staying well within the rules.” See also United States v. Kattar, 840 F.2d 118, 127 (1st Cir.1988) (stating that the function of the prosecutor “is not merely to prosecute crimes, but also to make certain that the truth is honored to the fullest extent possible”); Bruno, 721 F.2d at 1195.11 This is so because “[s]ociety wins not only when the guilty are convicted but when criminal trials are fair; our system of justice suffers when any accused is treated unfairly.” Brady v. Maryland, 373 U.S. 83, 87, 83 S.Ct. 1194, 1197, 10 L.Ed.2d 215 (1963).
The prosecutor may not “[become] the architect of a proceeding that does not comport with the standards of justice.” Id. The prosecutor, therefore, violates the Due Process Clause if he knowingly presents false testimony — whether it goes to the merits of the case or solely to a witness’s credibility. Napue v. Illinois, 360 U.S. 264, 79 S.Ct. 1173, 3 L.Ed.2d 1217 (1959); Mooney v. Holohan, 294 U.S. 103, 55 S.Ct. 340, 79 L.Ed. 791 (1935).12 Moreover, the prosecutor has a constitutional duty to correct evidence he knows is false, even if he did not intentionally submit it. Giles v. Maryland, 386 U.S. 66, 87 S.Ct. 793, 17 L.Ed.2d 737 (1967).
From these bedrock principles, it is well established that when no new significant evidence comes to light a prosecutor cannot, in order to convict two defendants at separate trials, offer inconsistent theories and facts regarding the same crime. Then-judge Kennedy wrote for our court that when there are *1059claims of inconsistent prosecutorial conduct, reversal is not required where the underlying theory “remains consistent.” Haynes v. Cupp, 827 F.2d 435, 439 (9th Cir.1987). Here, little about the trials remained consistent other than the prosecutor’s desire to win at any cost.
Drake v. Kemp, 762 F.2d 1449 (11th Cir. 1985) (en banc), bears remarkable similarities to our case. There, the State charged two men with murder and armed robbery. The State prosecuted the first defendant, Campbell, on the theory that Campbell alone committed the murder. But in the second defendant’s trial, the State asserted that Campbell, due to illness, was physically incapable of killing the decedent. Rather, the State argued, the second man, Drake, must have “actually beat the victim,” and, in support of this theory, the State presented Campbell as its star witness. Id. at 1472 (Clark, J., concurring). On direct examination, Campbell recited the same testimony that the State had refuted at Campbell’s trial. Drake was convicted of murder and robbery.
The court granted habeas relief on other grounds, but Judge Clark wrote separately to emphasize that the prosecutor’s actions of advancing inconsistent theories constituted a “fundamental and egregious error” that violated the Due Process Clause. Id. at 1470 (Clark, J. concurring). Judge Clark reasoned:
[T]he prosecution’s theories of the same crime in the two different trials negate one another. They are totally inconsistent. This flip flopping of theories of the offense was inherently unfair. Under the peculiar facts of this case the actions by the prosecutor violate the fundamental fairness essential to the very concept of justice. Lisenba v. California, 314 U.S. 219, 62 S.Ct. 280, 86 L.Ed. 166 (1941). However, it makes no sense to say that only Campbell’s due process rights were violated by the inconsistent theories____ This is especially true in this case because the fingerprints, blood tests, and a witness linked Campbell and not Drake to the crime. That evidence is highlighted by Campbell’s subsequent recantation.
The state cannot divide and conquer in this manner. Such actions reduce criminal trials to mere gamesmanship and rob them of their supposed search for truth. In prosecuting Campbell and Drake for the murder of Mr. Eberhart, the prosecutor changed his theory of what happened to suit the state. This distortion rendered Henry Drake’s trial fundamentally unfair.
Id. at 1479 (Clark, J., concurring).
We believe that Judge Clark’s concurrence expresses well the principles that govern the case before us. We disagree with the panel’s statement that if any party suffered from the prosecutor’s conduct, it was Leitch, not Thompson. To the contrary, Thompson was prejudiced. From the beginning, the prosecutor’s theory was that the murder resulted from Leitch’s plot to eliminate his former girlfriend, who impeded his reunion with his ex-wife and her money, and that both Leitch and Thompson were involved in the actual killing. Only in Thompson’s trial did the prosecutor change the theory and the arguments, and offer facts that directly conflicted with the underlying premise of the charges he brought. Only in Thompson’s trial did the prosecutor assert that Thompson was alone in the apartment and killed Fleisehli to cover up a rape. Only in Thompson’s trial did the prosecutor use as witnesses Fink and Del Frate, who were known to law enforcement officers to be wholly unreliable. In Leiteh’s trial, the prosecutor returned to his original theory and discredited the very evidence he had previously offered in Thompson’s trial. Thus, Thompson, rather than Leitch, suffered from the due process deprivation that infected the conflicting prosecutions.
IV. MURDER CONVICTION
Finally, consideration must be given to the effect of our vacatur of the rape conviction on the first-degree murder conviction. Thompson contends, with some force, that the conviction cannot stand because, inter alia, the jury was instructed under the felony murder doctrine. However, the district judge did not address that issue, nor did he discuss the murder conviction in his opinion, except to say, “as to the murder charge, ... the writ of *1060habeas corpus is denied.” Accordingly, we remand the question of the murder conviction to the district court for its initial consideration in light of our vacatur of the rape conviction.
CONCLUSION
Having recalled the mandate, we vacate the panel’s decision and affirm the district court with respect to the rape conviction, the rape special circumstance, and the death penalty. We remand the first degree murder conviction to the district court for further consideration in light of this action. We grant Thompson’s writ of habeas corpus in part and order that the State vacate its order of execution. The district court shall enter the partial writ unless the State elects to retry Thompson within a reasonable time. The vacatur of the execution order shall be effective immediately. The mandate shall issue forthwith.
AFFIRMED IN PART, REVERSED IN PART, AND REMANDED; EXECUTION ORDER VACATED.
. Thompson also filed a Rule 60(b) motion in the district court that was denied by the district court. He has appealed the denial to our court. We dismiss that appeal, which is now moot, in a separate order.
. But see G.O. § 5.4.c.(3), which provides that any judge may sua sponte call for an en banc vote within 21 days after the filing of a disposition or amended disposition. (Emphasis added).
. As we have emphasized in the text, supra, our recall of the mandate is not predicated on any new evidence or claims Thompson raises in his motion to recall the mandate.
. The district court found Brower's explanations not credible and criticized both his performance at trial and his deposition testimony at the habe-as hearing.
. The Stale also called two jailhouse informants, who testified that Thompson had admitted to raping Fleischli. As we discuss infra, Brower also was ineffective for failing to impeach these informants, whose testimony was patently unreliable.
.There also was evidence that Fleischli's blouse and bra had been cut up the middle and pulled down to her elbows. Viewed with the evidence that Fleischli’s jeans were on, zipped, and belted, this more likely suggests that the restraint was connected to the murder than that it evidenced a rape.
. Vogel testified, for example, that Thompson had told him that Fleischli had been "a little tipsy,” but that "[h]e didn’t rape her.”
. According to the prosecution, Leitch’s desire to reconcile with his ex-wife was motivated by the fact that she had recently come into a substantial sum of money as the result of a personal injury settlement.
. Henceforth, when we refer to “the prosecutor,” we mean Jacobs.
. In fact, Fink had made his testimony available to the State before the joint preliminary hearing. The state elected not to call him at that stage, but rather chose to rely on other informants who told very different stories.
. The American Bar Association also recognizes the special place of prosecutors in our constitutional system. "The responsibility of a public prosecutor differs from that of the usual advocate; his duty is to seek justice, not merely to convict.” ABA Model Code of Prof. Responsibility EC 7-13 (1981); see also ABA Standards for Criminal Justice § 3-5.8(c)(d) (2d ed.1981) (prosecutor has responsibility to guard rights of accused and those of society).
. Given that a prosecutor may not “knowingly” introduce false testimony, we disagree with the district court’s decision to read, but to declare "irrelevant,” the prosecutor's deposition testimony taken for this federal habeas proceeding. We think that the prosecutor's knowledge and beliefs, before and after Thompson’s trial, regarding this crime are entirely probative of whether the prosecutor presented at the trial a theory and set of facts that he knew contradicted the theory and facts that he planned to advance, and eventually advanced, at Leitch’s trial.