This is an appeal from the denial of a petition for habeas corpus. Petitioner Kenneth Duane Roy challenges his state court convictions of robbery and first-degree murder for aiding and abetting a felony murder. He contends the state trial court erred by failing to instruct the jury on the specific intent that is a necessary element of aiding and abetting under California law. The district court agreed, but held the error harmless. A divided panel of this court affirmed. We granted en banc review, and now reverse.

Petitioner Kenneth Duane Roy and his friend Jesse McHargue met Archie Mannix and James Clark outside a liquor store in Gridley, California. The four began drinking beer. Several hours later, a Gridley police officer saw Mannix's truck narrowly miss a utility pole as it backed up in the store's parking lot. The officer stopped the truck and called another officer to assist. Mannix and Clark were both intoxicated. McHargue and Roy appeared to be sober, but neither had a driver's license. The officers told the men not to drive the truck and left. Two hours later, the truck was gone. The officers found it nose down in a ditch, with the bodies of Clark and Mannix nearby. Both had been stabbed. Mannix, whose body was partially submerged in the ditch, had drowned. Both were partially stripped and their pockets turned out. Mannix's wallet and papers were scattered on the ground. The officers located Roy and McHargue at a nearby restaurant, their clothes wet and muddy. Each was carrying a buck knife. Each had some of Mannix's property in his possession. Roy told police the killings occurred after McHargue lost control of the truck while making a turn, the truck went into the ditch, and Clark became angry and struck Roy. According to Roy, a fight ensued, Roy against Clark and McHargue against Mannix. Roy stabbed Clark and killed him.

Roy was charged with two counts of murder and two counts of robbery. At trial, Marie Smart testified she was driving home when she saw the truck in the ditch and stopped to offer assistance. Two men were standing over Mannix, who was lying on the ground and appeared to be hurt. McHargue told her help had been summoned. A pathologist testified that Mannix's fatal stab wound could have been made by either McHargue's or Roy's knife. Roy's knife bore traces of blood that could have come from either Roy or Mannix but not from Clark.

William Hudspeth, a jailhouse informant, testified Roy told Hudspeth that Roy and McHargue planned to take Clark and Mannix to the country, rob and kill them, and steal the pickup truck. McHargue had trouble subduing Mannix and Roy came to McHargue's aid, pulling Mannix away, stabbing Mannix and holding his head under water until he was dead. According to Hudspeth, Roy and McHargue then took the truck and drove back toward Gridley.

Another jailhouse informant, Sidney Hall, testified Roy told him that after the truck went into the ditch, Clark hit Roy with a stick. A fight followed, and Roy stabbed Clark. Roy saw McHargue was "getting the worst of it" in his fight with Mannix, and "went over to help" McHargue.

The state took the murder case to the jury on two theories, arguing Roy was guilty of first-degree murder (1) because the killings were premeditated and (2) because they were committed during the course of a felony, the robbery of Clark and Mannix. The jury found Roy guilty of second-degree murder of Clark and made a "special circumstance" finding, for purposes of sentencing, that Roy had used a knife to kill Clark. The jury acquitted Roy of robbing Clark. The jury found Roy guilty of robbery and first-degree murder of Mannix, with a "special circumstance" finding that Roy had not used a knife to kill Mannix. Roy challenges his convictions of robbery and first-degree murder of Mannix.

The jury's decision to convict Roy of second-degree murder of Clark indicates the jury rejected the state's theory that the defendants planned the crime. The jury also rejected the state's contention that Roy stabbed Mannix by finding Roy did not use a knife against Mannix. Thus Roy's conviction of first-degree murder of Mannix necessarily reflected a conclusion by the jury that Roy was guilty of felony murder of Mannix in the course of aiding and abetting the robbery of Mannix by McHargue.

The jury was given an aiding and abetting instruction which stated that "[a] person aids and abets the commission of a crime if, with knowledge of the unlawful purpose of the perpetrator of the crime, he aids, promotes, encourages or instigates by act or advice the commission of such crime." This instruction allowed the jury to convict Roy if he provided "knowing aid"-that is, if he knew of McHargue's intention to rob Mannix and took some action that had the effect of furthering the robbery. After Roy's case was tried, the California Supreme Court held in People v. Beeman, 35 Cal. 3d 547, 199 Cal. Rptr. 60, 674 P.2d 1318 (1984), that an instruction identical to the one given in Roy's case was flawed because an aiding and abetting conviction requires proof not merely of "knowing aid" but also that the defendant intended to encourage or facilitate the offense with which the principal was charged.1

On direct appeal, Roy contended the state trial court erred by failing to instruct the jury on the specific intent element of aiding and abetting identified in Beeman. The California court of appeal concluded error had occurred but was harmless beyond a reasonable doubt. The California Supreme Court denied relief on collateral review, and Roy then filed this federal habeas petition raising the Beeman issue. In denying the petition, the district court held the omission from the instruction of the specific intent requirement was error, but agreed with the state courts that the error was harmless beyond a reasonable doubt because "[n]o rational juror could find that Roy aided McHargue, knowing what McHargue's purpose was, without also finding that Roy intended to aid McHargue in his purpose." A divided panel of this court affirmed. Roy v. Gomez, 55 F.3d 1483 (9th Cir.1995).

We have held that omission of the specific intent element from jury instructions in the trial of a charge of aiding and abetting under California law deprives the defendant of his constitutional right to have a jury find the existence of each element of the charged offense beyond a reasonable doubt. Martinez v. Borg, 937 F.2d 422, 423 (9th Cir.1991); see In re Winship, 397 U.S. 358, 364, 90 S. Ct. 1068, 1072, 25 L. Ed. 2d 368 (1970) (due process requires proof beyond a reasonable doubt of all elements of the charged offense). We also held in Martinez that Beeman error is subject to harmless-error analysis. Martinez, 937 F.2d at 425. The panel agreed on both points. Roy, 55 F.3d at 1485-86. The panel divided, however, as to whether the error was harmless.

To determine whether the Beeman error was harmless, we apply the analysis developed by Justice Scalia in his concurring opinion in Carella v. California, 491 U.S. 263, 109 S. Ct. 2419, 105 L. Ed. 2d 218 (1989). See Martinez, 937 F.2d at 425.2

Carella involved a conclusive presumption that relieved the state of its burden of proof with regard to the intent element of embezzlement. Justice Scalia explained that use of such a presumption could be harmless only in the "rare situations" when the reviewing court could be confident that the error played no part in the jury's verdict. Carella, 491 U.S. at 270, 109 S.Ct. at 2423 (Scalia, J., concurring) (quoting Connecticut v. Johnson, 460 U.S. 73, 87, 103 S. Ct. 969, 977, 74 L. Ed. 2d 823 (1983)). Stated shortly, such an error is harmless under Carella only if no rational jury could find the predicate facts forming the basis for the presumption without also finding the presumed fact. Carella, 491 U.S. at 271, 109 S.Ct. at 2423 (Scalia, J., concurring). In applying Carella to an instruction omitting an element of the offense, we have treated the omitted element as the "presumed fact" and considered whether a rational jury could have found the remaining elements of the offense without also finding the omitted element. Martinez, 937 F.2d at 424; see also United States v. Parmelee, 42 F.3d 387, 393 (7th Cir.1994).

Pointing to our recent en banc decision in United States v. Gaudin, Roy argues we may no longer apply harmless error analysis to Beeman error. In Gaudin, the district court instructed the jury that an element of the crime was established as a matter of law. We held that "such an error cannot be harmless." United States v. Gaudin, 28 F.3d 943, 951 (9th Cir.1994) (en banc) ("When proof of an element has been completely removed from the jury's determination, there can be no inquiry into what evidence the jury considered to establish that element because the jury was precluded from considering whether the element existed at all."), aff'd, 515 U.S. 506, 115 S. Ct. 2310, 132 L. Ed. 2d 444 (1995). Relying on this language, some subsequent panel decisions have held omission of an element of a crime from jury instructions requires automatic reversal. See United States v. Hove, 52 F.3d 233, 235-36 (9th Cir.1995); United States v. Stein, 37 F.3d 1407, 1410 (9th Cir.1994).

The error in Gaudin differs in a crucial respect from omission of an element of the crime from jury instructions. When a court instructs the jury that an element of the crime has been established as a matter of law, proof of that element of the crime is removed from the jury's purview. Failure to mention an element of the crime, in contrast, does not "completely remove[ ]" from the jury's consideration the evidence relating to that element; it simply fails to alert the jurors they must consider it. See Gaudin, 28 F.3d at 951; United States v. Whitmore, 24 F.3d 32, 36 (9th Cir.1994) (omission of knowledge element did not bar jury from considering defendant's mental state); People v. Dyer, 45 Cal. 3d 26, 64, 246 Cal. Rptr. 209, 753 P.2d 1 (1988) (Beeman error "is not the type of instructional error that wholly prevents the jury from considering" the defendant's intent).3 Even though an element of the offense is not specifically mentioned, it remains possible the jury made the necessary finding. Review for harmless error is appropriate, but it is the type of review discussed in Carella.4 That is, the omission is harmless only if review of the facts found by the jury establishes that the jury necessarily found the omitted element.5

On the record in this case, we cannot be certain the jury necessarily found beyond a reasonable doubt that Roy intended to facilitate McHargue's robbery of Mannix, as required under Carella and Martinez before the instructional error can be treated as harmless. See Carella, 491 U.S. at 271, 109 S.Ct. at 2423; Martinez, 937 F.2d at 425 ("The error is harmless if no rational jury would have made these findings without also finding that appellant had the specific intent to aid" the principal's crime). Although there was evidence from which a jury could have found that Roy intended to facilitate Mannix's robbery, there were no findings from which we could conclude the jury actually did so.

The testimony of the two jail informers, Hudspeth and Hall, indicates Roy realized McHargue was "getting the worst of it" in his fight with Mannix, and went to McHargue's assistance. From this evidence, the jury could have found Roy's intent was not to help McHargue rob Mannix but to prevent Mannix from defeating McHargue. Alternately, the jury could have found that although Roy realized McHargue was trying to rob Mannix, and in fact aided McHargue by keeping Clark from going to Mannix's aid, Roy did not intend that result.6

It was for the jury, not the judges who have reviewed the case, to determine which interpretation of the evidence was correct. We are not free to evaluate the evidence and postulate what the jury would have found had it been properly instructed. "[T]he question is not whether guilt may be spelt out of a record, but whether guilt has been found by a jury according to the procedure and standards appropriate for criminal trials." Carella, 491 U.S. at 269, 109 S.Ct. at 2422 (Scalia, J., concurring) (emphasis added) (quoting Bollenbach v. United States, 326 U.S. 607, 614, 66 S. Ct. 402, 406, 90 L. Ed. 350 (1946)). Because we cannot say the jury necessarily found Roy acted with the intention of assisting McHargue in the robbery of Mannix, the error in the aiding and abetting instruction cannot be deemed harmless.7

The state argues relief is not warranted because Roy has not shown the error had a substantial or injurious effect on the jury's verdict, as required when the error is raised in collateral proceedings. See Brecht v. Abrahamson, 507 U.S. 619, 113 S. Ct. 1710, 123 L. Ed. 2d 353 (1993). We disagree.

On direct appeal, relief is granted for constitutional error unless the state demonstrates the error was harmless beyond a reasonable doubt. See Chapman v. California, 386 U.S. 18, 24, 87 S. Ct. 824, 828, 17 L. Ed. 2d 705 (1967). As Justice Scalia said in Carella, the Chapman test can be met only if the reviewing court can tell what the jury actually found, since only then can the court conclude " 'beyond a reasonable doubt,' Chapman v. California, 386 U.S. 18, 24 [87 S. Ct. 824, 828, 17 L. Ed. 2d 705] (1967), that the jury found the facts necessary to support the conviction." Carella, 491 U.S. at 271, 109 S.Ct. at 2424. The Chapman standard is inapplicable on collateral review, however. In Brecht, the Court adopted a stricter standard for harmless error in habeas cases, holding relief is warranted on collateral attack only if the error " 'had substantial and injurious effect or influence in determining the jury's verdict.' " Brecht, 507 U.S. at 623, 113 S.Ct. at 1714 (quoting Kotteakos v. United States, 328 U.S. 750, 776, 66 S. Ct. 1239, 1253, 90 L. Ed. 1557 (1946)); see Hegler v. Borg, 50 F.3d 1472, 1477 (9th Cir.1995). More recently, the Supreme Court has held relief is also appropriate if the record on collateral review leaves the judge in "grave doubt" as to the effect of the constitutional error. See O'Neal v. McAninch, 513 U.S. 432, ---- - ----, 115 S. Ct. 992, 994-95, 130 L. Ed. 2d 947 (1995). Relief was granted in O'Neal because the record was "so evenly balanced that a conscientious judge is in grave doubt as to the harmlessness of the error." O'Neal, 513 U.S. at ----, 115 S. Ct. at 995. In such circumstances, "the uncertain judge should treat the error, not as if it were harmless, but as if it affected the verdict (i.e., as if it had a 'substantial and injurious effect or influence in determining the jury's verdict')." Id. at ----, 115 S.Ct. at 994.

We are unable to conclude under Carella that the jury necessarily found the missing element; if this case were before us on direct review, the error would not be harmless beyond a reasonable doubt, our analysis would be at an end, and we would be required to reverse the conviction. Because this case reaches us on habeas, however, we must determine whether reversal is required under the Brecht/ O'Neal line of cases. We believe it is. When the reviewing court is unable to conclude the jury necessarily found an element that was omitted from the instructions, it is unable to gauge the effect of the error on the jury's verdict. In this situation, a conscientious judge can only be "in grave doubt as to the harmlessness of the error," O'Neal, 513 U.S. at ----, 115 S. Ct. at 995, and relief must be granted.

Roy's due process rights were violated when he was convicted under an aiding and abetting instruction that omitted California's requirement that the defendant have the specific intent to assist in the commission of the crime. Because a rational jury could have found Roy's actions had the effect of assisting McHargue in the robbery of Mannix, but Roy did not intend his actions to have that effect, we are unable to say the jury necessarily found the required intent. Under Carella, Brecht, and O'Neal, we cannot say the violation of Roy's due process rights was harmless error.

REVERSED and REMANDED.

WALLACE, Circuit Judge, with whom Circuit Judges CYNTHIA HOLCOMB HALL and RYMER join, concurring and dissenting:

I agree with the majority that omitting or misdescribing an element of an offense is subject to harmless-error review. However, I cannot agree either with the method of harmless-error analysis the majority employs or with its conclusion that the error was harmful. I respectfully dissent.

* The majority holds that whenever a jury instruction contains an element that has been misdescribed or omitted and the jury did not actually find the facts supporting the missing element, a judge can never know whether the error had a "substantial and injurious effect or influence in determining the jury's verdict" as required by Brecht v. Abrahamson, 507 U.S. 619, 623, 113 S. Ct. 1710, 1714, 123 L. Ed. 2d 353 (1993) (Brecht ) (internal quotation omitted). See maj. op. at 867. By equating error under Carella v. California, 491 U.S. 263, 271, 109 S. Ct. 2419, 2423, 105 L. Ed. 2d 218 (1989) (Scalia, J., concurring) (Carella concurrence), with error under Brecht, the majority pays lip-service to the exclusive, less onerous, standard that the Court in Brecht said we should apply to trial errors when our review is collateral. I believe the majority erroneously looks to the Carella concurrence for the standard of review in habeas corpus cases, and that its error is compounded by its misapplication of Brecht and O'Neal v. McAninch, 513 U.S. 432, ----, 115 S. Ct. 992, 995, 130 L. Ed. 2d 947 (1995) (O'Neal ).

I agree with the majority that Brecht 's less onerous standard completely supplants Chapman 's harmless-error test when we review collaterally. Maj. op. at 868; Brecht, 507 U.S. at 635-39, 113 S.Ct. at 1721-22. I also agree that the Carella concurrence explained how the Chapman harmless-error test is to be applied on direct appeal when the error involves a jury instruction. Maj. op. at 867-868. But from these two premises I conclude that because Carella was derived from Chapman, Brecht must supplant Carella when we review a jury instruction error collaterally.

Brecht explicitly requires us to apply its standard of review to determine "whether habeas relief must be granted because of constitutional error of the trial type." Brecht, 507 U.S. at 638, 113 S.Ct. at 1722; see also id. at 651, 113 S.Ct. at 1729 (O'Connor, J., dissenting) (Court's holding applies "to any trial error asserted on habeas"); O'Neal, 513 U.S. at ----, 115 S. Ct. at 994 (Brecht "sets forth the standard normally applied by a federal habeas court in deciding whether or not ... constitutional 'trial' error is harmless"). Misdescription of an element of an offense is trial error. See Arizona v. Fulminante, 499 U.S. 279, 306-07, 111 S. Ct. 1246, 1262-63, 113 L. Ed. 2d 302 (1991) (jury instruction containing an unconstitutional conclusive presumption is trial error). We should therefore apply Brecht--and Brecht alone--to determine whether the Beeman error had a "substantial and injurious effect or influence in determining the jury's verdict." Brecht, 507 U.S. at 623, 113 S.Ct. at 1714 (internal quotation omitted).

We must also apply Brecht rather than the Carella concurrence because doing so would be faithful to the Court's insistence that a "less onerous harmless-error standard on habeas promotes the considerations underlying our habeas jurisprudence." Id. at 623, 113 S. Ct. at 1714. Yet by effectively choosing to apply the Carella concurrence instead of Brecht, the majority implicitly engages in a balancing that the Court has already done. Brecht already balanced the stricter Chapman harmless-error standard against policies discouraging habeas corpus relief and concluded that a standard less onerous than Chapman was more appropriate in all habeas corpus cases involving trial error. The result of Brecht 's balance is clear: providing "habeas relief merely because there is a reasonable possibility that trial error contributed to the verdict is at odds with the historic meaning of habeas corpus--to afford relief to those whom society has grievously wronged." Id. at 637, 113 S.Ct. at 1721 (citations and internal quotations omitted). At the very least, Brecht requires a reviewing court to ask whether a petitioner suffered "actual prejudice." Id. at 637, 113 S.Ct. at 1722.

The strict Carella concurrence standard, of course, asks not whether there is a "reasonable probability that trial error contributed to the verdict," but whether a rational jury necessarily found certain facts. The Carella concurrence explained that Chapman 's "harmless beyond a reasonable doubt" standard may not substitute a judge's findings for a rational jury's findings. Under the Carella concurrence, an error may be harmful on direct review even if there is a reasonable probability, or a strong probability, or a near-certain probability that the error had absolutely no effect on the outcome. But the Supreme Court has told us that in these situations we may not disturb state convictions collaterally. Simply put, I would not hold, as the majority does, that "society has grievously wronged" every habeas corpus petitioner whose trial contains error that is harmful under the Carella concurrence.

I have even more difficulty following the majority's attempt to blend Carella, Brecht, and O'Neal. See maj. op. at 867-68. By first applying the stricter approach in Carella, the opinion eliminates the effect of Brecht 's "less onerous standard" of review. Brecht, 507 U.S. at 635-639, 113 S.Ct. at 1721-22. Brecht requires the reviewing court to determine whether the error had a substantial or injurious effect on the outcome. Brecht also requires such a court to review the record in order to determine an error's effect. The Carella concurrence, of course, does not permit this thorough review of the record. Thus, if the majority actually were "to determine whether reversal is required under the Brecht/ O'Neal " line of cases, maj. op. at 868, it must thoroughly review the record. Instead, the majority applies only the Carella concurrence, limiting inquiry to whether the jury necessarily found the uninstructed element. The majority insists that in every case in which a jury has not actually found the missing element, a reviewing court is never able to determine whether the instruction error had a substantial and injurious effect. Maj. op. at 868. Clearly, a judge's ability to determine an error's effect is foreclosed unless the judge thoroughly reviews each case's record with its individual circumstances. This is what Brecht requires. This is what the majority rejects.

I cannot understand how we can know that the "unusual" and "narrow" circumstance in O'Neal, 513 U.S. at ----, ----, 115 S. Ct. at 994, 995--which occurs when a judge is in grave doubt as to whether an error had a substantial and injurious effect--will always exist when jury instructions are harmful under the Carella concurrence. A judge need only reach the "grave doubt" issue in O'Neal after being unable to decide whether an error had a substantial and injurious effect. Id. at ----, 115 S.Ct. at 994. The majority holds that in every case involving a jury instruction not satisfying the Carella concurrence, "in the judge's mind, the matter is so evenly balanced that he feels himself in virtual equipoise as to the harmlessness of the error." Id. But O'Neal requires each individual judge reviewing a habeas petition to ask, "Do I, the judge, think that the error substantially influenced the jury's decision?" Id. at ----, 115 S. Ct. at 995. The majority mandates the answer to that question for every judge in our circuit. Maj. op. at 868.

In addition, the majority creates inter-circuit conflict by refusing to apply Brecht instead of the Carella concurrence when reviewing collaterally jury instruction errors. See Libby v. Duval, 19 F.3d 733, 739-40 (1st Cir.) (holding that Brecht 's record-review applies to instructional errors reviewed on habeas corpus and that Carella was inapplicable in the habeas corpus context), cert. denied, 513 U.S. 927, 115 S. Ct. 314, 130 L. Ed. 2d 277 (1994); Cuevas v. Washington, 36 F.3d 612, 620 & n. 17 (7th Cir.1994) (applying Brecht to instruction error on habeas corpus review).

I would therefore apply the standard set forth in Brecht and ask whether the district court's failure to include specific intent in the aiding and abetting robbery jury instruction had a substantial and injurious effect on the jury's verdict. The majority searched the record and contends that a rational jury could have found that Roy assisted McHargue in the robbery of Mannix, but also could have found that Roy did not intend his actions to have that effect. Maj. op. at 867, 868-69. However, given the evidence that the jury actually heard, this latter possibility does not establish the substantial and injurious effect that Brecht requires.

The jury found Roy guilty of second degree murder of Clark, guilty of aiding and abetting the robbery of Mannix by McHargue, and guilty of the felony murder of Mannix, with robbery as the underlying felony. The aiding and abetting instruction actually given required the jury to find that Roy aided in the commission of the robbery offense by McHargue. The instruction also required the jury to find that when Roy provided this aid, he did so with the actual knowledge of McHargue's unlawful purpose. Given what the jury actually found and the evidence in the record supporting Roy's specific intent to further the robbery, I would hold the Beeman error harmless under Brecht. The error did not have a substantial and injurious effect on the jury's verdict.

I certainly do not have the grave doubt that the majority holds I necessarily must. Roy admitted stabbing Clark. Mannix's shirtless body was found submerged in water under his truck. Mannix died either from being stabbed or drowned. Mannix's wallet was found with one dollar in it. Upon his arrest, Roy's pants were wet from the calf down. Police found $170 and Mannix's wristwatch among Roy's possessions. Roy told Hall that he "helped" McHargue when McHargue and Mannix were fighting. Hudspeth testified that Roy admitted a plan to rob Clark and Mannix and admitted robbing both. Hall testified that Roy admitted helping McHargue with Mannix. From this I would conclude that there is not even a reasonable probability that Roy did not assist McHargue with the intent to further the robbery of Mannix. Therefore, Roy's habeas corpus petition should be denied.