Opinion by Judge KOZINSKI; Dissent by Judge GEORGE.
KOZINSKI, Circuit Judge.In this state habeas appeal, we consider whether petitioner was represented by constitutionally deficient counsel.
I
Sixteen years ago, Charles Tomlin was convicted of first degree murder for shooting *1237Daniel Stewart during a drug deal gone bad. Stewart and his girlfriend, Laura Leticia Mendez, had arranged to sell nine pounds of marijuana through an intermediary, Charles Tillman. On the night the sale was to take place, everything seemed to go according to plan until Tillman directed Stewart and Mendez to an alley and got out of their truck, promising to return with the money. Seconds later, an armed man got into the truck and directed Stewart, who was at the wheel, to drive. The assailant soon fired his gun in Stewart’s direction to get him to comply, which Stewart did, and then robbed them each of a few dollars and ordered Stewart to drive down an alley. When Stewart refused and appeared to reach for a weapon, the man shot Stewart in the head, grabbed the drugs and fled.
Mendez, who was unharmed, immediately provided a description of the assailant to police: a black man, approximately twenty-five years old; about five feet six to eight inches tall,1 150-160 .pounds, stocky and broad-shouldered; with a one and a half to two inch afro and, perhaps, a mustache; and wearing jeans and a Pendleton-type shirt. RT 222, 226.2 She also pointed the finger— first at Tomlin’s picture, and then directly at him, both in a live line-up and during trial. In fact, Mendez was the prosecution’s case.3
II
The live line-up, however, was illegal. Although Tomlin was represented, his lawyer— apparently through inadvertence on the part of the officers involved — was not notified and did not participate. The state, therefore, could not adduce any evidence of the line-up at his trial. Gilbert v. California, 388 U.S. 263, 273-74, 87 S.Ct. 1951, 1957, 18 L.Ed.2d 1178 (1967). And — if challenged — it would only have been entitled to present Mendez’s in-court identification by demonstrating through “clear and convincing evidence that the in-court identification [wa]s based upon observations of the suspect other than the lineup identification.” United States v. Wade, 388 U.S. 218, 240, 87 S.Ct. 1926, 1939, 18 L.Ed.2d 1149 (1967). Yet Tomlin’s counsel never challenged Mendez’s in-court identification and, in fact, himself elicited Mendez’s testimony that she’d identified Tomlin in a live line-up.
The only issue before us is whether Tomlin’s lawyer was constitutionally deficient. To show ineffective assistance of counsel, Tomlin must show that “counsel made errors so serious that [he] was not functioning as the ‘counsel’ guaranteed the defendant by the Sixth Amendment,” and that “the deficient performance prejudiced the defense.” Strickland v. Washington, 466 U.S. 668, 687, 104 S.Ct. 2052, 2064, 80 L.Ed.2d 674 (1984). There is a “strong presumption that counsel’s performance f[ell] within the ‘wide range of professional assistance.’” Kimmelman v. Morrison, 477 U.S. 365, 381, 106 S.Ct. 2574, 2586, 91 L.Ed.2d 305 (1986) (quoting Strickland, 466 U.S. at 689, 104 S.Ct. at 2065).
The district court held that counsel’s performance was constitutionally deficient, but that Tomlin was not prejudiced because Mendez’s in-court identification was derived from an independent source and would have been admissible under Wade4 A claim of ineffective assistance of counsel is a mixed question of law and fact, reviewed de novo. Mannhalt v. Reed, 847 F.2d 576, 579 (9th Cir.1988); see Adams v. Peterson, 968 F.2d 835, 843 (9th Cir.1992) (en banc).
Ill
When faced with a client who has been identified in an illegal line-up, most *1238defense attorneys would challenge the admission of any evidence related to it. After all, a defendant “arguably ... has ‘everything to gain and nothing to lose’ in filing a motion to suppress,” United States v. Molina, 934 F.2d 1440, 1447 (9th Cir.1991), especially one involving an identification by the sole eyewitness to the crime.
Of course, “it is not professionally unreasonable to decide not to file a motion ... clearly lacking in merit.” Id. The government, however, does not argue that the suppression motion here would have lacked merit, nor could it. Rather, it asserts that Tomlin’s counsel had a tactical reason for failing to challenge the admission of this evidence, namely that the state would call Charles Tillman as a witness. See Transcript of Evi-dentiary Hearing at 250. Tillman had confessed to functioning as an intermediary during the events leading up to Stewart’s death. It was he who led Stewart and Mendez to an alley and then disappeared, leaving Stewart and Mendez to the tender mercies of an armed assailant. It would seem a pretty safe bet he’d know the assailant’s identity. And he had, according to the police, provided them with Tomlin’s nickname, “Treetop,” when asked who’d committed the murder.5
In light of this scenario, the state argues, it made sense for Tomlin’s counsel to avoid challenging the line-up evidence — which might have led to suppression of Mendez’s in-court identification — in order not to force the state to call Tillman as a witness. As the theory goes, defense counsel believed that the state would make a deal with Tillman for his testimony only if it lost Mendez’s in-court identification,6 and he thought Mendez’s testimony would be easier to impeach than Tillman’s.
In a case that so hinges on an eyewitness’s testimony, however, it’s difficult to assume that a reasonable tactical decision was made not to challenge that testimony. We agree with the district court that the explanation presented by the State is simply too implausible to support a finding that counsel’s performance was objectively reasonable. We also have serious doubts whether a competent attorney in this position would have predicted the state could make a better case with Tillman’s testimony than with that of an innocent eyewitness. Tillman was, after all, the assailant’s apparent accomplice, something certain to affect his credibility in the eyes of the jury; in fact, the trial court would surely have instructed the jury to use caution in weighing his testimony. See CALJIC 3.18 (5th ed. 1988). And, in order to secure a conviction using Tillman’s testimony, the State would need corroborating evidence connecting Tomlin to the offense. Cal.Penal Code § 1111 (West 1985); People v. Price, 1 Cal.4th 324, 443-44, 3 Cal.Rptr.2d 106, 821 P.2d 610 (1991). If Mendez’s live line-up and in-court identifications were kept out, the only such evidence the state would have had was Mendez’s preline-up description of the assailant which, as discussed below, differs in material respects from Tomlin’s appearance, and her identification of him in a photo lineup which, as also addressed below, she later testified was in various ways suggestive.
Moreover, the state’s version of events does not coincide with counsel’s testimony at the state habeas evidentiary hearing. According to counsel, he “was almost certain that [Tillman] was going to be offered a deal on the thing.” Transcript of Evidentiary Hearing at 257. Not that this was contingent on whether Mendez would identify Tomlin — indeed, not that it was at all influenced by that fact. He thought it was going to happen. This does not reflect the type of deep strategic planning the state now attributes to Tomlin’s counsel.
*1239Nor is there any other basis for finding counsel’s actions to have been “the result of reasonable professional judgment.” Strickland, 466 U.S. at 690, 104 S.Ct. at 2066. Some seven years later, at the evidentiary hearing, counsel seemed to explain that he didn’t think the in-eourt identification would have been excluded because “there had been a previous photographic line-up that would not be a poisoned line-up and would have been admitted.” Transcript of Evidentiary Hearing at 258. But he did not indicate that that was the basis on which he chose not to object. In fact, when asked if he would “have any reason for not objecting to [a live line-up in the absence of counsel],” he responded, “No. I certainly — if this was a fact, in retrospect, I certainly should have objected.” Id. at 254. He called himself “derelict” for not making a motion on these grounds. Id. at 258-59. Nor could we deem such an approach reasonable, absent some indication the motion would have been lacking in merit. Molina, 934 F.2d at 1447. There is none. And without it, the failure to bring to the court’s attention a major constitutional error in the prosecution’s case is not the product of reasonable professional judgment.
We therefore agree with the district court that defense counsel’s performance fell below an objective standard of reasonableness.
IV
Before making a determination as to possible prejudice, we first take note of another troubling aspect of this case: Mendez didn’t just identify Tomlin in court; she also testified she’d identified him previously in a live line-up. That evidence is without a doubt excludable under Gilbert, 388 U.S. at 273-74, 87 S.Ct. at 1957. The twist is, Mendez first revealed this line-up evidence in response to questioning by Tomlin’s counsel on cross-examination (“And you later made an identification at a line-up in the Kern County Jail, is that correct? ... At that time the person you identified did not have an Afro, did he?” RT 216). The prosecution then immediately pounced on the issue on redirect, asking a number of questions about the illegal line-up, mostly designed to confirm Mendez had in fact selected Tomlin. Tearing wide open the narrow gap created by defense counsel’s hapless inquiry, the prosecutor later questioned the officer who had conducted the line-up about how it was conducted, and introduced into evidence photographs taken at the time. Finally, in his closing statement he stressed to the jury that Mendez had identified Tomlin in a live lineup and that the line-up had been fair, and reminded the jury to look at the photographs of it.
Although we’re concerned about defense counsel’s decision to bring evidence of the illegal line-up before the jury, and his failure to challenge the extent of the prosecutor’s subsequent exploitation of it, we’re also quite troubled by the prosecutor’s actions. This is certainly not the first case where a defense lawyer decided to elicit this kind of evidence on cross-examination of an eyewitness. See, e.g., Wade, 388 U.S. at 220, 87 S.Ct. at 1928-29; United States v. Patton, 721 F.2d 159, 162 (6th Cir.1983).7 But the government’s behavior here, in the face of its own initial culpability in conducting an illegal line-up, and doing nothing to later correct the error,8 pushes hard on the limits of acceptable rebuttal, cf. Clemons, 408 F.2d at 1246: Defense counsel essentially asked Mendez a couple of questions about Tomlin’s hairstyle at the time she identified him, RT at 216-17; the prosecutor used this opportunity to drive home to the jury that Mendez had identified *1240Tomlin in a live line-up, that the line-up had been fair and that Mendez had “no doubt” in her mind when she picked Tomlin, RT at 220.9
Tomlin, however, raised the issue of this illegal line-up testimony for the first time in a one paragraph supplemental letter a few days before argument in this court, submitting relevant pages from the trial transcript and directing us to our recent decision in United States v. LaPierre, 998 F.2d 1460 (9th Cir.1993), dealing with testimony regarding an illegal line-up. It’s not clear whether Tomlin is arguing that this testimony should not have been admitted at all. But, whatever his argument is, we cannot address it now. 28 U.S.C. § 2254(b), (c).
Nor can we examine the issue in the context of Tomlin’s ineffective assistance claim. All along, the basis for this claim has been that counsel failed to challenge the admission of evidence relating to the illegal line-up, not that he himself opened the door to such evidence at trial and thus opened the door for the state’s massive follow-up. Because Tomlin did not raise this as part of his ineffective assistance claim below, we’re unable to act on it.
Y
We look, then, to whether Mendez’s in-eourt identification would have been admitted under Wade. Tomlin can show prejudice by showing a reasonable probability that, had the government been called to the task, it would have been unable to satisfy Wade’s requirements: “to establish by clear and convincing evidence that the in-court identification[ ] [was] based upon observations of the suspect other than the lineup identification.” Wade, 388 U.S. at 240, 87 S.Ct. at 1939.
In making this determination, we are mindful of the Supreme Court’s concerns regarding pretrial line-ups set forth in Wade. There, the Court recognized “the dangers inherent in eyewitness identification and the suggestibility inherent in the context of the pretrial identification.” Id. at 235, 87 S.Ct. at 1936. The Court requires counsel to be present at in-person line-ups with good reason: “A major factor contributing to the high incidence of miscarriage of justice from mistaken identification has been the degree of suggestion inherent in the manner in which the prosecution presents the suspect to witnesses for pretrial identification.” Id. at 228, 87 S.Ct. at 1933. Although a witness may later deny there was any suggestiveness — as Mendez did at trial here — that witness is “[un]likely to be schooled in the detection of suggestive influences.” Id. at 230, 87 S.Ct. at 1934. The same goes for the defendant. Id. at 231, 87 S.Ct. at 1934. And, the reality of the situation is, once a witness has picked out a suspect — whether influenced by the police in doing so or not — she will not easily go back on her word. Id. at 229, 87 S.Ct. at 1933.10
In light of these concerns, we must assess the likelihood that Mendez’s in-court identification was free from the taint of the illegal line-up by considering such factors as
the prior opportunity to observe the alleged criminal act, the existence of any discrepancy between any pre-lineup description and the defendant’s actual description, any identification prior to lineup of another person, the identification by picture of the defendant prior to the lineup, failure to identify the defendant on a prior occasion, and the lapse of time between the alleged act and the lineup identification.
Id. at 241, 87 S.Ct. at 1940. In addition, we look to such particulars as the lighting conditions under which Mendez viewed the assailant, her level of distraction, and any hesitation on her part in making other pretrial identifications. United States v. Crews, 445 *1241U.S. 463, 473 n. 18, 100 S.Ct. 1244, 1251 n. 18, 63 L.Ed.2d 537 (1980).
Applying these factors in Crews, the Court held that the trial court’s finding of an independent source for the victim’s courtroom identification was “ampl[y] support[ed]” where the victim had viewed her assailant at close range for five to ten minutes under “excellent lighting conditions and with no distractions,” had provided a description which “closely matched” the defendant, had failed to identify anyone else and had twice identified the defendant, and only a week had passed between the crime and her first identification of him. Id. at 473, 100 S.Ct. at 1250-51. The case, however, did not involve a possibly suggestive pretrial line-up, but merely an illegal arrest where a “court-ordered,” id. at 467, 100 S.Ct. at 1247-48, and presumably legal line-up was conducted.
After analyzing the Wade factors, the district court here held that Mendez’s in-court identification was derived from an independent source.11 In doing so, it relied on several “significant events” that took place before the illegal line-up: Mendez’s observations of the assailant at the time of the crime; her subsequent description of Tomlin to the police; and her selection of his picture in a high school yearbook and in a photographic lineup. One problem with its analysis is that one of these “significant events” didn’t occur: It was Tillman’s picture, not Tomlin’s, that Mendez selected from the yearbook. We are also left to wonder whether the district court took into account the government’s burden of proof in seeking to show that the in-court identification was constitutional—which would have been by proof of “clear and convincing” evidence. Wade, 388 U.S. at 240, 87 S.Ct. at 1939. After considering the Wade factors, the court merely stated it was “reasonable to conclude that Mendez would have been allowed to identify petitioner in court.” ER 20.
Reviewing its findings de novo,12 we part company with the district court’s determination that there was an “independent source” for Mendez’s in-court identification. As the district court recognized, Mendez had an opportunity to observe the assailant for four to five minutes. And, as the state points out, Mendez testified at trial that she had a good chance to look at the assailant’s face. RT 208. The circumstances under which she viewed him, however, limit to some extent the reliability of any observations Mendez made. Not only did the assailant suddenly barge into the truck while she and Stewart were stopped in a dark alley, but he also sat next to her. To observe him, she would have had to turn her head, and even then it would have been difficult to see him full-face. When one of the first things the assailant did was to fire his gun at close range within the enclosed space of the truck cab, it’s doubtful Mendez would then have spent much time turned towards him, calmly noting his facial features and other aspects of his appearance. The situation here is thus very different from that in Crews, where the victim had viewed her assailant frontally for five to ten minutes under the bright lights of a public restroom. Crews, 445 U.S. at 473 n. 18, 100 S.Ct. at 1251 n. 18.
We also don’t have here a suspect who “closely matched the description given by the victim immediately after the robbery.” Id. Contrary to what the district court found, Mendez did not “immediately identiffy] the petitioner to a police officer.” ER 20. She identified a black man, twenty-five years in age, five feet six to eight inches tall, heavy build, with a possible mustache and an inch and a half to two-inch afro, wearing blue jeans and a Pendleton-type shirt. Tomlin, by contrast, although a black man who was then *1242twenty-six years old, is six-feet tall, and at the time was thin and wore a shoulder-length, straightened permanent hair style. RT 345-46, 348-49; see Transcript of Evi-dentiary Hearing at 14. Tomlin and other witnesses testified that he did not own a Pendleton-type shirt. RT 348. Considering the heavy clothing worn by the assailant and the fact that Mendez based her description of his height on her estimation while she sat next to him, RT 208, we might discount the difference between Mendez’s description and Tomlin’s actual height and build.13 However, there’s no discounting the discrepancy between Tomlin’s hair style and Mendez’s description.14 While a physical description need not be perfect to be rehable, see United States v. Monks, 774 F.2d 945, 956-57 (9th Cir.1985), the discrepancies here — never mentioned by the district court — east additional doubt on the reliability of Mendez’s description. Their shadow thus extends to her later in-court identification.
We must also consider any pre-lineup picture identifications. Mendez chose Tomlin’s picture from a photographic line-up. At the time of trial, she testified that there had been no undue influence by the detective involved when she made this identification. When testifying during the state evidentiary hearing on Tomlin's habeas petition, however, Mendez indicated that, at the time of this line-up, she was most likely under the influence of tranquilizers, and that when she was shown Tomlin’s photograph, the detectives exchanged glances that led her to choose it. She also said that she’d expressed uncertainty at the time, and told the officers she needed to see the suspect in person as it was “hard to tell from a photo.” Transcript of Evidentiary Hearing at 116-120.
The state court concluded her testimony at the evidentiary hearing was not credible, and we accord this determination a presumption of correctness. 28 U.S.C. § 2254(d); Marshall v. Lonberger, 459 U.S. 422, 434, 103 S.Ct. 843, 850-51, 74 L.Ed.2d 646 (1983). To reject this finding, we must conclude that it “laek[s] even ‘fair support’ in the record.” Id. at 432, 103 S.Ct. at 850. The record from the evidentiary hearing, however, indicates Mendez _ waffled considerably between recanting her trial testimony and confirming it. The number of contradictions and uncertainties we find in reading through it leads us to find “fair support” for the state court’s determination that Mendez’s evidentiary hearing testimony that the photographic line-up was suggestive was not credible.15 The result of that line-up, therefore, remains reliable.16 And it cuts in the government’s favor that there was only a short time that passed between the crime and this first identification of Tomlin. Crews, 445 U.S. at 473 n. 18, 100 S.Ct. at 1251 n. 18.
*1243Of course, it is also significant that Mendez never identified another individual as the assailant and, as had the victim in Crews, did identify Tomlin before trial. Wade, 388 U.S. at 241, 87 S.Ct. at 1939-40. But the police gave Mendez only one other chance to identify her assailant, and that was the illegal lineup. And we cannot consider the live line-up to bolster her in-court identification:17 “[T]he accurate pretrial identifications assume significance only to the extent that they indicate that the witness’ ability to identify respondent antedated any police misconduct. ...” Crews, 445 U.S. at 473 n. 18, 100 S.Ct. at 1251 n. 18. As noted, the prosecution did nothing to mitigate the effects of its misconduct, see n. 8 supra, which, in turn, would have helped carry its heavy burden of showing an independent source.
This is a close case. Before an illegal lineup was ever conducted, and soon after the shooting, Mendez identified Tomlin in a photographic line-up. Yet, although she did observe the assailant for a few minutes at the time of the murder, various constraints necessarily limited her observations. And immediately afterward, she gave the police a description of the assailant that differed significantly from Tomlin’s actual appearance. Considering all of the evidence, we cannot find, as Strickland requires, “a reasonable probability” that, had Tomlin’s counsel objected to the in-court identification, the government would have been able to show— clearly and convincingly — that Mendez’s ability to identify Tomlin was not influenced by the illegal line-up. Tomlin thus was prejudiced by his lawyer’s failure to challenge Mendez’s in-court identification. And, there is a serious risk that Tomlin was, in fact, wrongly identified as the assailant. His petition must be granted.
REVERSED.
. We note that at various points she testified that she'd described his height as being between five feet six to seven inches tall.
. Although Mendez first told police the assailant was a hitchhiker she and Stewart had picked up, she testified at trial that her description was always of the actual assailant.
. The prosecution stipulated, during the course of the state habeas proceedings, that Mendez’s eyewitness testimony "was the sole evidence tying [Tomlin] as the perpetrator." Transcript of State Habeas Evidentiary Hearing ("Evidentiary Hearing”) at 65.
.The district court also rejected another claim raised by Tomlin, that counsel was ineffective for failing to investigate a rumor that a certain other individual committed the murder. As Tomlin has not pursued this claim on appeal, we do not address it.
. The police officer who interviewed Tillman after his arrest claimed Tillman had written the name "Treetop” on a piece of paper when asked to identify the assailant. He failed, however, to retain the paper. Transcript of Evidentiary Hearing at 285-87. At the evidentiary hearing, Tillman denied having done so and having ever pointed the finger at Tomlin. In fact, he claimed to have turned down a deal for a three year sentence if he would testify that Tomlin was the assailant. Instead, he went to trial and was convicted of murder and robbery. Id. at 270-78.
. Tomlin's lawyer was apparently unaware of Tillman's refusal to testify against Tomlin. Instead, he thought Tillman had falsely identified Tomlin out of fear of the actual assailant, and might do so again at trial. Id. at 257.
. As the D.C. Circuit explained in Clemons v. United States, 408 F.2d 1230, 1237 (D.C.Cir.1968), “If the judge regards only the in-court identification as admissible ... the defense may, as a matter of trial tactics, decide to bring out the pre-trial confrontation itself, hoping that it can thus detract from the weight the jury might otherwise accord the in-court identification.”
. After the prosecutor learned of the illegal lineup, he might have arranged for a second line-up, involving several panels not all of which included Tomlin, with defense counsel present. Mendez's reactions might have indicated her first pick was in error or somehow influenced by the police. Also, knowing full well that the government was culpable in conducting the illegal line-up, it would certainly have been possible to limit the damage by seeking a pretrial ruling on whether Mendez's in-court identification could properly be introduced under Wade. See Clemons, 408 F.2d at 1237.
. Tomlin's counsel asked four questions pertaining to the line-up; the prosecutor asked twenty.
. That there were pictures of the line-up mitigates the prejudice, but does not remove it. Although the pictures would show the other individuals who participated in the line-up, they obviously cannot disclose anything that happened in the witness room. See LaPierre, 998 F.2d at 1464 (videotape of line-up where counsel not present did not satisfy Wade's concerns because it did not record anything that occurred in witness room).
. Although we are referring to the magistrate’s findings and recommendations, they were adopted in full by the district court.
. Whether Mendez’s identification was derived from an independent source is a mixed question of law and fact, as " 'the historical facts are admitted or established, the rule of law is undisputed, and the issue is whether the facts satisfy the [relevant] ... [constitutional] standard, or to put it another way, whether the rule of law as applied to the established facts is or is not violated.’’’ Norris v. Risley, 878 F.2d 1178, 1181 (9th Cir.1989) (quoting United States v. McConney, 728 F.2d 1195, 1200 (9th Cir.1984) (en banc)) (alteration in original). As the mixed question here involves constitutional rights, de novo review is particularly appropriate. Id.
. Her description of his weight was fairly accurate, but this seems to cut the other way considering the height discrepancy. A five foot seven individual who weighs 150 pounds looks significantly different from a six foot individual at that weight.
. The state argues that we shouldn't be influenced by this discrepancy, because the arresting officer testified at the evidentiary hearing that, at the time of Tomlin's arrest several days after the crime, he had his hair combed in such a way that it could look like a short natural in the dark. Perhaps it's possible to comb a shoulder-length straightened style so that it looks like a short afro but, considering these are vastly different hairstyles, we cannot entirely discount the significance of this discrepancy.
. For example, Mendez at one point testified that she'd been truthful and had no doubt at the time she'd made trial and pretrial identifications of Tomlin, but later stated she was unsure at the time she identified him. Transcript of Evidentia-ry Hearing at 142, 228. Although she first stated that during the photographic line-up the detectives "exchanged glances” when she reached Tomlin's photo, on cross-examination she testified that when she was going through the lineup, she stopped at Tomlin's photograph because, at the time, it looked like the assailant. Id. at 117, 131.
.In determining whether Tomlin was prejudiced by his counsel's failure to challenge Mendez's in-court identification, we note that had he demanded an evidentiary hearing on the independent source issue, Mendez's testimony would have been key. As Tomlin now argues, we don't know what might have come out at that time concerning the conduct of the photo line-up; perhaps details would have been elicited along the lines of some of her statements at the eviden-tiary hearing that would have led the trial judge to exclude the photo line-up evidence. The burden, after all, was on the state to establish an independent source by clear and convincing proof. Wade, 388 U.S. at 240, 87 S.Ct. at 1939.
. Again, although Mendez testified that the lineup was not suggestive, her testimony is necessarily somewhat unreliable on this point. See Wade, 388 U.S. at 230, 87 S.Ct. at 1933-34.