This appeal from the denial of a habeas corpus petition presents a thicket of procedural and substantive issues arising from a murder on the streets of Chicago. On the Fourth of July in 2002, Tony Cox was standing outside a restaurant when he was gunned down by two men. The gunmen fled, but two women driving cars near the scene saw the murder and the shooters’ faces. Not quite two months later, both women independently chose petitioner Eric Blackmon’s photograph out of arrays, identifying him as the second shooter. They repeated those identifications at a live line-up and then again at trial. Primarily on the strength of their testimony, Blackmon was convicted of first-degree murder and sentenced to sixty years in prison.
Blackmon petitioned the state courts for post-conviction relief, arguing that his attorney was constitutionally ineffective because he failed to present certain eyewitness and alibi testimony, and in the alternative, that he was actually innocent. The state courts summarily denied relief. Blackmon then sought relief in federal court under 28 U.S.C. § 2254. The district court denied Blackmon’s petition.
The record before us supports the conclusion that Blackmon’s trial counsel was constitutionally ineffective by failing to investigate the alibi witnesses and shows that the state court’s summary dismissal of the claim was unreasonable. But a “state court’s mistake in summarily rejecting a petition, ie., without fully evaluating conflicting evidence on disputed factual issues, does not necessarily mean the petitioner is ultimately entitled to relief.” Mosley v. Atchison, 689 F.3d 838, 842 (7th Cir. 2012). Because of that summary dismissal, the alibi witnesses have not yet been tested in any sort of adversary proceeding, and the record contains no evidence from Black-mon’s trial counsel as to what he did or did not do. Accordingly, we vacate the denial of the habeas petition and remand to the district court to assess whether Blackmon “is actually ‘in custody in violation of the Constitution or laws or treaties of the United States.’ ” Id., quoting 28 U.S.C. § 2254(a).
I. Background
A. The Murder of Tony Cox
At about 4:30 p.m. on July 4, 2002, Tony Cox was shot and killed in front of a restaurant called Fat Albert’s located at 1143 South Pulaski Road in Chicago. Though the various accounts of the shooting differ in some ways, the basic outline is relatively consistent, with one exception discussed below.
At the time of the shooting, Cox and Richard Arrigo, the owner of Fat Albert’s, were outside the restaurant with two other men. The first assailant shot Cox at least twice, and Cox fell to the ground. The second assailant then shot Cox twice more, and both men fled. The State’s theory is that Blackmon was the second shooter. Cox died from four bullet wounds to the head, each of which would have been fatal *1093alone. The medical examiner recovered two bullets from Cox’s body, and a forensic scientist with the Illinois State Police later testified that those bullets had come from two different guns.
Frencshun Reece and Lisa McDowell, the witnesses who would later form the core of the State’s case, saw Cox’s murder from their respective vehicles. Reece was stopped at a red light when she saw the first assailant shoot Cox twice. According to Reece, both assailants began to flee, but the second assailant turned back and shot Cox in the head again before running away. Reece pulled over and called the police, although Arrigo urged her to hang up and claimed he had already called. Reece remained at the scene to speak with police. Later that day, she viewed a photographic array and selected three pictures that “resembled” one or both of the assailants, though she made no definitive identification that day. Blackmon was not part of this first photo array.
McDowell was driving south when she stopped at a red light, about two car-lengths back from the intersection. She did not see the first two gunshots, but she looked over after hearing them and saw Cox lying on the ground, with Arrigo standing nearby. She then saw the two assailants approach Cox. One of them shot him twice more, and both men fled. McDowell also called 911 but did not remain at the scene, and she did not view photographs of possible suspects that day.
Arrigo, too, witnessed the crime. According to his account, he was locking the doors of his restaurant when he heard two gunshots. He turned and saw the second assailant shoot Cox twice and flee with the first. Like Reece, Arrigo remained at the scene and spoke to police, but he told them he had not recognized the shooters and could not identify them.
B. The Investigation
A few days into the investigation, a possible explanation for the shooting began to emerge. George Davis, also known as “Booney Black,” was the founder of the New Breed gang, of which Cox was a member. According to an “Investigation Time-Line” produced during discovery, police received a tip indicating that Davis had lent Cox money to start a drug corner but had ordered him killed when things went badly. There was also evidence suggesting that Arrigo, who was Davis’s friend, helped arrange the murder. The investigators found a message from Arrigo on Cox’s voicemail instructing Cox to meet him at the restaurant. Arrigo confirmed it was his voice on the recording but claimed not to remember leaving the message, and he denied arranging the murder. Cell phone records also showed that Arrigo had called Davis immediately after the shooting.
Police also received information that two individuals nicknamed “Pride” and “Keno” may have killed Cox. “Keno” was the nickname of Michael Davis, the nephew of George Davis; “Pride” was the nickname of a man named Eric Bridges, who, like Cox, was a member of the New Breed gang.
At some point, the police began to focus on Eric Blackmon instead, though the record does not show how he became a suspect. The “Investigation Time-Line” mentions “Pride” and “Keno” but does not even mention Blackmon (whose nickname is or was “Forty”). The rest of the record provides little additional insight. A detective testified during grand jury proceedings that Blackmon and Cox were “members of the same street gang,” but the State’s counsel conceded at oral argument that no trial evidence supports this assertion. There was also testimony at trial that “family members” had verified Blackmon *1094was “involved” in the murder, but the detective did not explain further and the “Investigation Time-Line” does not mention them.
Regardless, Blackmon did come under suspicion, and his picture was placed in a photo array that Reece, McDowell, and Arrigo all viewed in late August or early September, about two months after the shooting. Arrigo did not identify anyone from the photo array as one of the shooters. Reece and McDowell, however, both selected Blackmon’s photograph. A live line-up yielded the same results: Arrigo did not identify anyone, while both Reece and McDowell chose Blackmon, who was arrested and charged with Cox’s murder.
C. Blackmon’s Trial
At a bench trial in September 2004, the State presented the testimony of McDowell and Reece to prove that Blackmon was the second shooter on July 4, 2002. McDowell testified that she was stopped in traffic on South Pulaski Road when she heard a gunshot. When she looked to her left, she saw “an Italian guy” — that is, Arrigo — and then watched as two men came around a building. One of the men, she testified, stood over a body lying on the sidewalk and then fired two more shots at the victim. She described that shooter as a black male, a “tall guy, maybe about six feet, slender build.” She testified that she saw the gunman’s face and identified Blackmon in court as the man she saw. McDowell said she had seen the shooter from about fifteen to sixteen feet away, and that after firing, both the shooter and the other man had run across Pulaski Road.
On cross-examination, Blackmon’s counsel tried to undermine the reliability of McDowell’s identification. For example, he elicited testimony that she had seen the shooter’s face straight-on for only about five seconds. McDowell also testified that as she watched the shooter running away, she had only a view of his profile in her car’s rearview mirror from about forty-five feet away. Counsel also emphasized that McDowell had not viewed the photo array until weeks after the incident, and he sought to draw out discrepancies in her testimony as compared to her previous statements. Nevertheless, McDowell did not waver on her identification.
Reece testified that she was stopped at a red light on Pulaski Road when she noticed a group of four men, three black and one white, gathered outside Fat Albert’s. She said that she had heard a noise she initially believed to be firecrackers. Then, she said, she saw one of the black men pull out a gun and shoot toward Cox’s head. Cox went down on his hands and knees, and when he tried to rise, the first man shot him again “in his neck.” Reece testified that both assailants then “took off running,” but the second assailant returned, “came very close” to the victim, “pointed the gun directly into his eye,” and shot him. The second assailant then looked up at Arrigo, “shook his head,” and fled. Like McDowell, Reece identified Blackmon in court as the second shooter.
Reece testified that she pulled over, called the police, and waited at the scene until they arrived. Several hours later, she went to the police station and viewed a photo array (the one without Blackmon’s picture). She testified that she did not “really” identify anyone then but admitted saying some of the photos “looked familiar.” According to the officer who administered the photo array, Reece also selected a third picture depicting a subject who had a hairstyle similar to one of the shooters. Near the end of August, Reece viewed another photo array, this time with Black-mon’s picture included. She testified that she had identified Blackmon as the second *1095shooter from July 4 after viewing the second array based on the “bone protrusion” she could see through his shirt. On cross-examination, she elaborated that she had also identified Blackmon because he had “braids in his hair” and because he “looked like Michael Jackson.”
As he had with McDowell, Blackmon’s attorney attempted to discredit Reece’s identification. Reece said that she had seen the shooting from about 130 feet away (though she testified that at some point she began driving, getting “so close that I can hit them with my car”) and had seen the second shooter’s face close up for only three seconds. The attorney also pointed out that the men in the photos Reece had selected on July 4 did not look like Black-mon, although Reece maintained that she had chosen two photos that both resembled the first shooter, not the second. Blackmon’s attorney also attacked Reece’s identification by cross-examining Detective Gregory Jones, who administered the July 4 photo array. Jones testified that Reece had never said the second shooter looked like Michael Jackson, nor had she pointed out his bone structure. He also testified that Reece had not chosen two photographs that resembled the first shooter, as she claimed, but had in fact chosen two photographs “that resembled two of the individual offenders” — helpful to Blackmon because Reece conceded that neither subject resembled Blackmon.
Blackmon’s defense consisted of both alibi testimony and competing eyewitness testimony. To support his alibi, he offered testimony from two witnesses, Tomeka Wash and Selena Leavy. Wash testified that on July 4, 2002, she had hosted a barbecue in the lot across the street from her home. At about 1:00 p.m., she had seen Blackmon at the barbecue firing up the grill, and by 2:00 p.m., between twenty and forty people had arrived at the picnic, with guests coming and going throughout the afternoon. About half an hour before the murder took place, Wash said, Blackmon was still at the picnic playing chess and barbecuing. She said that she stayed until 10:00 p.m. that night and never saw Black-mon leave. According to Wash, she was close to Blackmon throughout the picnic and could see him “all the time.”
Leavy testified that she attended the same barbecue, arriving about 2:45 or 3:00 p.m. When she arrived, she saw Blackmon at the barbecue cooking and she said that at approximately 4:00 p.m., Blackmon had fixed her a plate of food. She also said that between 3:00 and 4:00 or 4:30 p.m., Black-mon was at the picnic playing dominos or chess with his friends and that he never left the barbecue at any time between her arrival and her departure around 8:00 p.m. Finally, she testified that Blackmon’s car did not leave the barbecue, either: the attendees had been using its radio for music and so presumably would have noticed if someone had driven it away.
On cross-examination, the prosecution brought out that Wash had two felony convictions and that Leavy was Black-mon’s cousin. The prosecution also emphasized that two years had passed since the barbecue; that Wash and Leavy claimed, somewhat improbably, to have kept Black-mon continuously in their sight for hours; and that guests had been coming and going throughout the picnic, presumably making it less likely that either woman could have kept track of a single guest without interruption.
In addition to the two alibi witnesses from the barbecue, the defense presented testimony from a competing eyewitness to the murder, Terrance Boyd. He testified that on July 4, 2002, he and Cox planned to go drinking together. When he arrived on Pulaski, he saw Cox standing in the street and spoke with him. Boyd said that *1096another man named “Booney” was also present, as was Eric Bridges. According to Boyd, Cox said he needed to talk to Eric, so Boyd walked away and turned the corner into an alley. He then heard gunshots and peered around the corner, where he saw Bridges shooting Cox. Boyd testified that he “trotted away,” fearing for his own safety, but returned once Bridges left the scene. He repeated that “Booney” had been present but initially did not remember seeing anyone else there. Later, Boyd said there may have been another “black person” present “at the curb, maybe coming from across the street,” but that he did not know who it was. He never mentioned Arrigo.
On cross-examination, Boyd acknowledged his multiple prior convictions, including one for perjury. He admitted that he first came forward with his account of the shooting two years after the fact when he was seeking to “work a deal” with the U.S. Attorney in a federal case. The prosecution also emphasized the flaws in Boyd’s story. For example, Boyd testified he and Cox had been friends for ten years, but Boyd never checked on Cox after the shooting. Boyd also said that “Booney” ran away during the shooting, but the prosecution introduced testimony that George Davis had an “apparent disability with his left leg” and that he had “a hard time walking.” Further, Boyd testified that he had seen only Eric Bridges shoot Cox, but the other eyewitnesses described two shooters, and ballistics evidence showed that two different weapons were used.
The judge found Blackmon guilty of murder. He summarized the testimony and evidence before making explicit credibility determinations about the witnesses. He found McDowell and Reece credible and persuasive because they expressed confidence in their identifications, withstood extensive cross-examination, and told consistent stories despite having had no contact with one another. The judge rejected Boyd’s alternative version of the shooting, finding that his testimony had “less than zero credibility.”
The judge also rejected Blackmon’s alibi, finding Wash’s and Leav/s stories to be highly unrealistic. He also noted that Wash’s credibility was undermined by her felony convictions, her testimony that she hosted the barbecue but knew only “the defendant and two other people,” and the fact that Leavy claimed to have spoken to Wash at the barbecue but Wash did not remember her being present. The judge also pointed out that Leavy’s family ties to Blackmon were revealed only on cross-examination, calling that fact “interesting.” Based on his assessment of the evidence, the judge found Blackmon guilty of first-degree murder. He sentenced Blackmon to sixty years in prison. The conviction and sentence were affirmed on direct appeal.
D. Collateral Proceedings
1. State Posir-Conviction Proceedings and the Discovery of New Witnesses
Blackmon then petitioned for state post-conviction relief. He claimed his trial counsel provided ineffective assistance by: (1) failing to investigate and call additional alibi witnesses from the barbecue; and (2) failing to call Arrigo as a trial witness. The state court summarily dismissed the petition, holding that the alibi witnesses’ testimony would have been cumulative, and that whether to call a witness like Arrigo was a matter of trial strategy.
The Illinois Appellate Court affirmed. First, it found that Blackmon’s evidence did not demonstrate actual innocence, which in Illinois means total vindication or exoneration of the defendant. People v. Williams, 394 Ill.App.3d 236, 333 Ill.Dec. 222, 914 N.E.2d 641, 651 (2009). The court *1097also agreed with the trial court that the alibi witnesses’ testimony would have been cumulative and held that neither that testimony nor Arrigo’s testimony was newly discovered, as required to state an actual innocence claim. See People v. Ortiz, 235 Ill.2d 319, 336 Ill.Dec. 16, 919 N.E.2d 941, 950 (Ill. 2009). Finally, the appellate court agreed with the trial court on the merits of the ineffective-assistance claims. It held that decisions regarding which witnesses to call are usually strategic and saw “no reason to set aside the usual deference to counsel’s trial strategy under such circumstances .... ” The court did not separately address the failure to investigate those witnesses. The Supreme Court of Illinois denied Blackmon’s petition for leave to appeal.
Blackmon later discovered additional witnesses helpful to his case: Latonya Thomas and Lajuan Webb, who were both employees of a barber shop adjacent to the crime scene. Both were working the day of the shooting and signed affidavits with their observations of that day.
Thomas swore that on July 4, 2002, she was working when her attention was drawn to the street outside by what she believed to be fireworks. As she looked through the window, she saw the victim fall to the ground, and a man stepped into her line of sight and shot the victim as he tried to rise. At that point, Thomas crouched beside a chair but could still see what was going on outside. She said that she saw a second man whom she knew as “Pee” approach the victim and shoot him “several more times.” Both men then fled north. Thomas acknowledged that she had not previously spoken to the police but said she had feared for her safety. She was adamant that she could identify the shooters because she had “seen them both hanging out on the street around the salon countless times.” She testified that neither man was Blackmon, whom she has never met.
Webb swore that on July 4, 2002, he was also working when he heard five or six gunshots outside. He then saw two black men with guns run past the shop. Webb, too, said he had seen both gunmen in the neighborhood before the shooting and that neither one was Blackmon. Webb said he gave his name to a police officer who came to investigate the shooting, but no one ever contacted him for questioning.
Based on these exculpatory affidavits, Blackmon sought leave to file a successive post-conviction petition in state court, arguing that Webb’s and Thomas’s testimony supported his actual innocence claim. He also re-asserted his claim that his counsel was ineffective for failing to call Arrigo at trial. The state trial court denied his request, and Blackmon appealed.
The Illinois Appellate Court affirmed again. People v. Blackmon, No. 1-11-1908, 2013 WL 2145922 (Ill. App. May 14, 2013). It noted that Illinois law ordinarily contemplates only one post-conviction proceeding, but a showing of either cause and prejudice or actual innocence could overcome that bar. Id. at *5. “Where defendant seeks to relax the bar against successive post-conviction petitions on the basis of actual innocence, leave of court should be denied only where it is clear from a review of the successive petition and supporting documentation that, as a matter of law, defendant cannot set forth a colorable claim of actual innocence.” Id., citing People v. Edwards, 360 Ill.Dec. 784, 969 N.E.2d 829, 836 (Ill.2012). To establish a claim of actual innocence under Illinois law, a defendant must present evidence that is newly discovered, material, and not merely cumulative, and of such conclusive character that it would probably change the result on retrial. Id. The court concluded that the affidavits of Webb and Thomas *1098could have been obtained long ago, stating that “any reasonable investigation would have included inquiries as to who was working at the salon at the time of the shooting” and that there “was no apparent obstacle to obtaining this information.” Id. at *6.
The court also held that Blackmon had not raised “the probability that it is more likely than not that no reasonable juror would have convicted defendant in light of the new evidence,” concluding he had not set forth a colorable actual innocence claim. Id. at *7. The court focused on the following facts: (1) both Webb and Thomas waited nearly eight years to disclose their observations of the shooting; (2) Thomas viewed the incident while crouched behind a chair; (3) Thomas merely contradicted the credible testimony of McDowell and Reece; and (4) Webb did not even see the shooting take place. Id. at *6-7. Finally, the court rejected on res judicata grounds the claim that counsel should have called Arrigo to testify. Id. at *7. The Supreme Court of Illinois denied Blackmon’s petition for leave to appeal.
2. Federal Habeas Corpus Proceedings
While the state courts were still considering Blackmon’s request to file a successive post-conviction petition, Blackmon filed a petition for a writ of habeas corpus in federal court. He asserted three distinct grounds for relief. Two were theories of ineffective assistance of counsel based on the failure to present Arrigo’s testimony and the failure to investigate and call additional alibi witnesses. The third was a claim of actual innocence based on newly discovered evidence — that is, the Webb and Thomas affidavits.
The federal court held his petition in abeyance pending resolution of the state post-conviction proceedings. The petition was briefed after the state court proceedings ended, with Blackmon filing a supplemental memorandum that expanded on his earlier-asserted claims. The memorandum also suggested for the first time that his claim of actual innocence was meant not as a stand-alone ground for relief but as a “gateway claim” to an otherwise-defaulted theory of ineffective assistance of counsel and a violation of Brady v. Maryland, 373 U.S. 83, 83 S.Ct. 1194, 10 L.Ed.2d 215 (1963).1 Turning to the merits of those defaulted claims, Blackmon argued that his trial counsel was constitutionally ineffective by failing to conduct an adequate investigation of businesses near the crime scene, which would have led him to the exculpatory testimony of Webb and Thomas.
When the ‘State filed its response to Blackmon’s § 2254 petition, it ignored the memorandum and Blackmon’s new characterization of the actual innocence claim as a “gateway” to an ineffective-assistance claim. Instead, it responded to the original § 2254 petition, limiting its analysis of Webb’s and Thomas’s testimony to the observation that federal habeas corpus law has not recognized freestanding claims of actual innocence.
The district court denied relief, concluding that the state court had not unreasonably applied the principles of Strickland v. Washington, 466 U.S. 668, 104 S.Ct. 2052, 80 L.Ed.2d 674 (1984). United States ex rel. Blackmon v. Hardy, No. 11 C 2358, 2014 WL 3511497, at *2-3 (N.D. Ill. July 16, 2014). The district court also upheld the state court’s rejection of the actual innocence claim, finding the analysis of the Thomas and Webb affidavits was reasonable. Id. at *4. We granted a certificate of *1099appealability as to whether Blackmon’s trial counsel “was ineffective for failing to call an occurrence witness and numerous alibi witnesses.” We also instructed the parties to address whether Blackmon had defaulted any of his ineffective-assistance theories; whether he had demonstrated actual innocence excusing default; and whether 28 U.S.C. § 2254(d) applies to state court findings regarding actual innocence.
II. Analysis
We review de novo the district court’s decision denying habeas relief. Stitts v. Wilson, 713 F.3d 887, 891 (7th Cir. 2013). Blackmon is in state custody, so we review his claims under 28 U.S.C. § 2254, as amended by the Antiterrorism and Effective Death Penalty Act of 1996 (“AED-PA”). Under § 2254(d), federal courts may not grant relief on any claim adjudicated on the merits by a state court unless the adjudication “resulted in a decision that was contrary to, or involved an unreasonable application of, clearly established Federal law, as determined by the Supreme Court of the United States,” or “resulted in a decision that was based on an unreasonable determination of the facts in light of the evidence presented in the State court proceeding.”
Blackmon relies on the “unreasonable application” prong of § 2254(d)(1). Under this provision, a federal court may grant habeas relief if the state court correctly identified the governing legal principle from the Supreme Court’s case law but unreasonably applied it to the facts of the petitioner’s case. Wiggins v. Smith, 539 U.S. 510, 520, 123 S.Ct. 2527, 156 L.Ed.2d 471 (2003) (citations omitted). This is a high standard: to grant relief, the state court’s decision must be objectively unreasonable, not merely incorrect. Id. at 520-21, 123 S.Ct. 2527 (citations omitted); Ba-delle v. Correll, 452 F.3d 648, 654-55 (7th Cir. 2006). The Supreme Court has warned that “even a strong case for relief does not mean the state court’s contrary conclusion was unreasonable.” Harrington v. Richter, 562 U.S. 86, 102, 131 S.Ct. 770, 178 L.Ed.2d 624 (2011); see, e.g., White v. Wheeler, — U.S.-, 136 S.Ct. 456, 193 L.Ed.2d 384 (2015) (per curiam) (summarily reversing grant of habeas corpus relief for failure to extend sufficient AEDPA deference to state court’s determination).
This appeal also presents questions of procedural default, which we review de novo. Richardson v. Lemke, 745 F.3d 258, 269 (7th Cir. 2014). A state prisoner must give the state an opportunity to correct alleged violations of federal rights by fairly presenting his claim through a full round of state court review. O’Sullivan v. Boerckel, 526 U.S. 838, 844-45, 119 S.Ct. 1728, 144 L.Ed.2d 1 (1999). The failure to do so results in procedural default. Id. at 848, 119 S.Ct. 1728. Federal courts will not review a procedurally defaulted claim unless the petitioner can demonstrate cause for and prejudice stemming from that default, or, alternatively, that the denial of relief will result in a miscarriage of justice. Lewis v. Sternes, 390 F.3d 1019, 1026 (7th Cir. 2004) (citations omitted). The miscarriage of justice exception “applies only in the rare case where the petitioner can prove that he is actually innocent of the crime of which he has been convicted.” McDowell v. Lemke, 737 F.3d 476, 483 (7th Cir. 2013). To meet that standard, he must demonstrate, based on new, reliable evidence, that “ ‘in light of new evidence, it is more likely than not that no reasonable juror would find him guilty beyond a reasonable doubt.’” Coleman v. Hardy, 628 F.3d 314, 319 (7th Cir. 2010), quoting House v. Bell, 547 U,S. 518, 537, 126 S.Ct. 2064, 165 L.Ed.2d 1 (2006).
*1100Procedural default is not jurisdictional. Trest v. Cain, 522 U.S. 87, 89, 118 S.Ct. 478, 139 L.Ed.2d 444 (1997); see Canaan v. McBride, 395 F.3d 376, 382 (7th Cir. 2005). Rather, it is “an affirmative defense that the State is obligated to raise and preserve, and consequently one that it can waive.” Perruquet v. Briley, 390 F.3d 505, 515 (7th Cir. 2004) “As with any other right or defense, the State will waive procedural default by intentionally relinquishing its right to assert that defense,” either explicitly or implicitly. Id.
A. Procedural Default of Claims Based on Webb’s and Thomas’s Testimony
Blackmon argues that his trial counsel’s failure to locate and investigate Webb and Thomas, who worked in the barber shop, was one of several errors amounting to ineffective assistance. The State responds that Blackmon proeedurally defaulted this claim by failing to raise it through one complete round of state court review. Blackmon contends in turn that the State implicitly waived its right to assert procedural default by failing to raise it in the response to his § 2254 petition filed in the district court.
Working our way backwards through these contentions, we first reject the argument that the State implicitly waived or forfeited the procedural default defense. Blackmon did not raise the failure to discover Webb and Thomas as a basis for his ineffective-assistance claim in the district court until he filed his supplemental memorandum, and the district court never ordered the State to respond to it. Under these unusual procedural circumstances, the State did not show “the intent to relinquish the defense that is the essence of true waiver.” Perruquet, 390 F.3d at 517 (finding no waiver where state had at most remained silent on procedural default issue). Nor would it be fair to hold the State to forfeiture where Blackmon introduced the claim so late.
Next, Blackmon proeedurally defaulted this basis for his ineffective-assistance claim. To preserve a claim for federal habeas review, a state prisoner must fairly present the operative facts and legal principles controlling the claim through a full round of state court review. Mulero v. Thompson, 668 F.3d 529, 536 (7th Cir. 2012), quoting Smith v. McKee, 598 F.3d 374, 382 (7th Cir. 2010). This rule requires that the factual and legal substance of the claim remain essentially the same when the petitioner has exhausted his state court remedies and moved on to federal court. Anderson v. Benik, 471 F.3d 811, 814-15 (7th Cir. 2006). With ineffective-assistance claims, we have held that a petitioner proeedurally defaults individual claimed deficiencies when he does not fairly present them to the state courts, even if he presented an ineffective-assistance claim based on other alleged errors. See Mulero, 668 F.3d at 536.
Blackmon presented an ineffective-assistance claim through one full round of state-court review, but it was based on the other matters we discuss below. His second post-conviction petition included a claim based on the Webb and Thomas affidavits, but he framed that as a freestanding actual-innocence claim, not an ineffective-assistance claim. Thus, neither the legal nor the factual substance of the claim he now seeks to assert was before the state courts, meaning he has procedurally defaulted that claim.
We might nevertheless address the merits of Blackmon’s claim if he could show cause and prejudice or a miscarriage of justice — that is, “the conviction of an innocent person.” Hayes v. Battaglia, 403 F.3d 935, 938 (7th Cir. 2005). He relies on the *1101latter, arguing that the affidavits of Webb and Thomas constitute “new reliable evidence,” Schlup v. Delo, 513 U.S. 298, 324, 115 S.Ct. 851, 130 L.Ed.2d 808 (1995), and that in light of that new evidence, it is “more likely than not that no reasonable juror would have convicted him,” id. at 327, 115 S.Ct. 851.2
As noted above, Schlup requires Blackmon to show that “more likely than not, in light of the new evidence, no reasonable juror would find him guilty beyond a reasonable doubt — or, to remove the double negative, that more likely than not any reasonable juror would have reasonable doubt.” House v. Bell, 547 -U.S. 518, 538, 126 S.Ct. 2064, 165 L.Ed.2d 1 (2006). The Schlup standard “is demanding and permits review only in the ‘extraordinary’ case.” Id., quoting Schlup, 513 U.S. at 327, 115 S.Ct. 851; see also McQuiggin v. Perkins, — U.S.-, 133 S.Ct. 1924, 1936, 185 L.Ed.2d 1019 (2013) (“We stress once again that the Schlup standard is demanding.”); Gladney v. Pollard, 799 F.3d 889, 896 (7th Cir. 2015) (“The actual innocence gateway is narrow.”). Our function “is not to make an independent factual determination about what likely occurred, but rather to assess the likely impact of the evidence on reasonable jurors.” House, 547 U.S. at 538, 126 S.Ct. 2064, citing Schlup, 513 U.S. at 329, 115 S.Ct. 851. In applying this standard, we must consider all the evidence, both old and new, incriminating and exculpatory, without regard to whether it would necessarily be admitted at trial. Id., citing Schlup, 513 U.S. at 327-28, 115 S.Ct. 851. We then make a “probabilistic determination about what reasonable, properly instructed jurors would do.” Id., quoting Schlup, 513 U.S. at 329, 115 S.Ct. 851.
The new evidence from Webb and Thomas does not meet this demanding standard for actual innocence. We recognize that the State’s case against Black-mon had important weaknesses: no physical evidence tied him to the murder, and the State introduced no evidence of a motive. But to counter, the State had matching independent identifications from Reece and McDowell. We have repeatedly recognized high rates of error in eyewitness identifications of strangers. United States v. Bartlett, 567 F.3d 901, 906 (7th Cir. 2009); Raygoza v. Hulick, 474 F.3d 958, 965 (7th Cir. 2007); Newsome v. McCabe, 319 F.3d 301, 305 (7th Cir. 2003). But we have also recognized that those findings “have only limited application when multiple witnesses identify the same person.” Bartlett, 567 F.3d at 907; see also Morales v. Johnson, 659 F.3d 588, 601 (7th Cir. 2011) (“Each witness’s identification of Morales as the shooter corroborated the other’s testimony.”); United States v. Williams, 522 F.3d 809, 812 (7th Cir. 2008) (“[T]he number of identifications supplies valuable information. Even if the risk that any one identification would be mistaken is *1102substantial, the risk that multiple witnesses would make the same error is smaller.”)-3 Furthermore, some of the State’s criticisms of Blackmon’s defense evidence are well founded. Arrigo had shown himself unreliable (and possibly even involved in the murder). And the alibi witnesses’ testimony is fairly subject to the criticism that it was acquired years after the barbecue in question and purports to recount fairly specific details of an unremarkable gathering. See Woods v. Schwartz, 589 F.3d 368, 377 (7th Cir. 2009) (rejecting claim of actual innocence based on alibi affidavits of family members that were prepared years after the murder and noted “with incredible particularity the most pedestrian details of that night,” even though “the night was like any other”).
Webb and Thomas saw the shooters for no longer than Reece and McDowell (in fact, Webb did not even see them until they were already fleeing). Furthermore, they did not come forward until eight years after the murder, a substantial delay that could affect their memories and/or their credibility. See Schlup, 513 U.S. at 332, 115 S.Ct. 851 (holding that “the court may consider how the timing of the submission and the likely credibility of the affiants bear on the probable reliability of that evidence”).
This sort of balance between inculpatory and exculpatory witnesses is not enough to meet the demanding Schlup standard for actual innocence. See, e.g., Smith v. McKee, 598 F.3d 374, 388 (7th Cir. 2010) (two exculpatory eyewitnesses insufficient to counter state’s two inculpatory eyewitnesses); Hayes v. Battaglia, 403 F.3d 935, 938 (7th Cir. 2005) (“Suppose that the six alibi witnesses had been called. That would at best have produced a draw: six eyewitnesses identify Hayes as the culprit, six others exculpate him. That cannot establish that ‘no reasonable factfinder would have found the applicant guilty of the underlying offense[.]’ ”); see also id. (“[pjroof of innocence must be considerably more than the proof required to establish prejudice” needed for ineffective assistance of counsel). Blackmon has not shown the miscarriage of justice needed to excuse his procedural default, so we do not consider the merits of the claim based on the testimony of Thomas and Webb.
B. Merits of the Remaining Ineffective-Assistance Claims
Blackmon asserts two non-defaulted bases for his claim of ineffective assistance of counsel under Strickland v. Washington, 466 U.S. 668, 104 S.Ct. 2052, 80 L.Ed.2d 674 (1984): the failure to call Ar-rigo as a witness at trial, and the failure to investigate and present additional alibi testimony. The operative question is whether the state court reasonably applied the familiar two-pronged test of Strickland: (1) whether counsel’s performance “fell below *1103an objective standard of reasonableness,” Strickland, 466 U.S. at 688, 104 S.Ct. 2052, and (2) whether counsel’s errors prejudiced the defendant, requiring him to show a reasonable probability that, but for those errors, the result of the proceeding would have been different, id. at 694, 104 S.Ct. 2052; see also Allen v. Chandler, 555 F.3d 596, 600 (7th Cir. 2009). The state court found that counsel’s performance was not constitutionally deficient, so we must give AEDPA deference to that conclusion, upholding it so long as it is not objectively unreasonable. Badelle v. Cornell, 452 F.3d 648, 654-55 (7th Cir. 2006).
1. Richard Arrigo
We cannot disagree with the state court’s finding that the decision not to call Arrigo was a reasonable strategic decision. “The Constitution does not oblige counsel to present each and every witness that is suggested to him.” United States v. Berg, 714 F.3d 490, 499 (7th Cir. 2013), quoting United States v. Best, 426 F.3d 937, 945 (7th Cir. 2005). Rather, counsel need only investigate possible lines of defense and make an informed decision. Id. “If counsel has investigated witnesses and consciously decided not to call them, the decision is probably strategic.” Best, 426 F.3d at 945. Strategic decisions like these, so long as they are made after a thorough investigation of law and facts, are “virtually unchallengeable.” Strickland, 466 U.S. at 690, 104 S.Ct. 2052.
Arrigo looks helpful at first glance. We must assume he would have supported Blackmon’s version of events by testifying that he saw the shooting and that Black-mon was not there. But Blackmon glosses over serious potential problems with that testimony. Arrigo had been caught lying to the police during the course of the investigation. He claimed that he did not have Cox’s phone number but later admitted that a voicemail on Cox’s cell phone featured his voice. His cell phone records showed he called Cox before the murder, and George Davis after it. Blackmon’s attorney could reasonably have concluded that putting Arrigo on the stand would do more harm than good. For example, a trier of fact might have believed that Arrigo helped arrange the murder and sought to protect his co-conspirator. Arrigo’s testimony would also have undermined the defense testimony from Boyd, who claimed that Arrigo’s friend George Davis was present at the shooting. Arrigo consistently maintained that he did not recognize either shooter. Also, Boyd did not even remember seeing Arrigo at the scene. Blackmon’s attorney could reasonably have taken this into account as well.
Blackmon argues that he can “overcome the presumption that, under the circumstances, the challenged action ‘might be considered sound trial strategy.’” Strickland, 466 U.S. at 689, 104 S.Ct. 2052, quoting Michel v. Louisiana, 350 U.S. 91, 101, 76 S.Ct. 158, 100 L.Ed. 83 (1955). He contends that the failure to call Arrigo was not strategic but was caused by counsel’s accidental failure to secure his attendance at trial. See Mosley v. Atchison, 689 F.3d 838, 848 (7th Cir. 2012) (presumption that strategic judgments are reasonable “applies only if the lawyer actually exercised judgment”). Blackmon relies on this exchange between counsel and the trial judge:
THE COURT: You’ve got two more witnesses?
[COUNSEL]: I have one for certain and I had spoken with another gentleman that was suppose[d] to be here but he’s a little — I don’t want to—
THE COURT: Okay.
[COUNSEL]: — characterize him in any fashion.
*1104Blackmon argues that the most reasonable inference is that counsel was referring to Arrigo, who failed to show up. That is one possible inference, but not the only one. Nothing in this exchange shows that the “gentleman” in question was Arrigo or that counsel failed to secure his attendance through inattention. The state court did not act unreasonably by finding that the decision not to call Arrigo was reasonable as a matter of strategy. In addition, given all the baggage Arrigo would have carried as a witness, we could not say that the failure to call him prejudiced Blackmon. That failure would not undermine confidence in the verdict.
2. Alibi Witnesses from the Barbecue
The alibi witnesses present a very different problem. As noted above, the decision not to call a particular witness is frequently strategic, insulated from attack on ineffective-assistance grounds. “An outright failure to investigate witnesses, however, is more likely to be a sign of deficient performance.” United States v. Best, 426 F.3d 937, 945 (7th Cir. 2005), citing Rompilla v. Beard, 545 U.S. 374, 387, 125 S.Ct. 2456, 162 L.Ed.2d 360 (2005); see also Strickland, 466 U.S. at 690-91, 104 S.Ct. 2052 (“strategic choices made after less than complete investigation are reasonable precisely to the extent that reasonable professional judgments support the limitations on investigation”).
On this record, the State must concede, at least for purposes of this appeal, that counsel did not interview any of the additional alibi witnesses whom Black-mon identified, but the State argues it was enough that counsel “learned the substance” of their testimony from interviews with Blackmon and his family members. Blackmon’s affidavit, which the State cites for support, says only that Blackmon informed his counsel he was at the barbecue, that plenty of people could vouch for his presence, and that he eventually provided his counsel with names, addresses, telephone numbers, and as much information as he could on potential alibi witnesses. This does not mean that counsel knew the substance of those witnesses’ testimony. Counsel had no way of knowing, for example, if any of the witnesses could definitively place Blackmon at the barbecue at 4:30 p.m. Such testimony could have provided Blackmon with a much stronger alibi. Nor does it appear that counsel or the state court considered the benefits of alibi testimony from disinterested witnesses who, as far as we know (and unlike Wash and Leavy, who did testify for Blackmon), had no family ties to Blackmon and no felony convictions. See Raygoza, 474 F.3d at 963 (noting that “witnesses both related and unrelated to Raygoza could have been called”); Washington v. Smith, 219 F.3d 620, 634 (7th Cir. 2000) (testimony of additional alibi witnesses without criminal records “would have added a great deal of substance and credibility” to alibi).
Counsel’s failure to investigate undermines the state court’s analysis, which appears to assume that counsel knew, somehow, that the additional alibi witnesses would offer purely cumulative testimony. If counsel never learned what the witnesses would have said, he “could not possibly have made a reasonable professional judgment that their testimony would have been cumulative.” Mosley, 689 F.3d at 848; see also Campbell v. Reardon, 780 F.3d 752, 764 (7th Cir. 2015) (rejecting state court’s assumption that lawyer’s decision not to interview eyewitnesses was reasonable; proper inquiry was “whether the investigation supporting counsel’s decision” not to call the witnesses “was itself reasonable”). The unreasonableness of counsel’s failure to investigate is further bolstered by the significant potential benefits of obtaining alibi testimony from wit*1105nesses unimpaired by family ties to Black-mon or prior convictions, another point the state court apparently did not consider.
The State also argues that the individual alibi witnesses would themselves have had vulnerabilities. That’s possible, of course, but counsel could not have known those vulnerabilities without doing at least some investigation of the witnesses and the testimony they could provide. There is also no indication he considered the effect all the witnesses might have had in combination, any individual weaknesses notwithstanding. See Raygoza, 474 F.3d at 964 (rejecting attorney’s decision not to call additional alibi witnesses where he “picked off’ each witness based on potential vulnerabilities without considering the cumulative impact of their testimony).
Strickland “permits counsel to ‘make a reasonable decision that makes particular investigations unnecessary.’ ” Harrington, 562 U.S. at 106, 131 S.Ct. 770, quoting Strickland, 466 U.S. at 691, 104 S.Ct. 2052. But the record provides no support to treat as reasonable a decision not to investigate further the available alibi witnesses from the barbecue. Blackmon’s location at 4:30 p.m. was the pivotal issue for the defense. Additional disinterested and credible alibi witnesses could have made a significant difference in the viability of Blackmon’s defense, especially given the problems with the alibi witnesses who did testify. See Washington, 219 F.3d at 631 (attorney’s failure to try to contact any witnesses besides one was ineffective, and state court’s decision to the contrary was an unreasonable application “of Strickland’s requirement that an attorney conduct a reasonable investigation in connection with his client’s case”). Nothing in the record shows that investigating those witnesses would have been “fruitless or harmful,” Campbell, 780 F.3d at 765, citing Strickland, 466 U.S. at 691, 104 S.Ct. 2052, and the benefits could have been enormous. Just one witness might have been able to give Blackmon a true alibi. At a minimum, all of them could have bolstered his claim of being at the barbecue all afternoon. It is not reasonable strategy to leave such possible testimony unexplored under these circumstances. So even giving both counsel and the state court the substantial deference they are due under Strickland and AEDPA, respectively, the state court’s finding with respect to trial counsel’s performance was, on this record, unreasonable.
Finally, we must determine whether “there is a reasonable probability that, but for counsel’s unprofessional errors, the result of the proceeding would have been different.” Strickland, 466 U.S. at 694, 104 S.Ct. 2052. The state court never reached the prejudice question, so we review de novo that prong of Strickland. See Sussman v. Jenkins, 636 F.3d 329, 350 (7th Cir. 2011), quoting Toliver v. McCaughtry, 539 F.3d 766, 775 (7th Cir. 2008). “A reasonable probability is ‘a probability sufficient to undermine confidence in the outcome.’ ” Mosley, 689 F.3d at 851, quoting Strickland, 466 U.S. at 694, 104 S.Ct. 2052. Unlike the more demanding Schlup inquiry, the issue here is not whether Blackmon is actually innocent, but instead whether he would have had a “reasonable chance” of acquittal absent counsel’s errors. Stanley v. Bartley, 465 F.3d, 810, 814 (7th Cir. 2006) (noting also that “it needn’t be a 50 percent or greater chance”). In undertaking this inquiry, we must “consider the totality of the evidence before the judge or jury. A verdict or conclusion that is overwhelmingly supported by the record is less likely to have been affected by errors than one that is only weakly supported by the record.” Hough v. Anderson, 272 F.3d 878, 891 (7th Cir. 2001), citing Strickland, 466 U.S. at *1106696, 104 S.Ct. 2052; see, e.g., Thomas v. Clements, 789 F.3d 760, 771-72 (7th Cir. 2015).
We begin there, with the weakness of the State’s case. No physical evidence tied Blackmon to the crime, nor did the State present any evidence of motive or connection between Blackmon and the victim. The only evidence connecting Blackmon to the murder was the eyewitness testimony of McDowell and Reece. That was enough to support the guilty verdict but is by no means ironclad. Neither witness saw the second shooter’s face for long, and both were strangers to Blackmon, increasing the danger of mistake. See United States v. Brown, 471 F.3d 802, 804 (7th Cir. 2006) (“Even under the best circumstances, the probability of erroneous identification of a stranger seen briefly is uncomfortably high.”), citing Elizabeth F. Loftus, Eyewitness Testimony (1979), and Daniel L. Schacter, The Seven Sins of Memory 88-137 (2001). As noted above, the two corroborating identifications help make mistakes less likely, but mistakes are certainly still possible where two eyewitnesses identify the same alleged perpetrator. Garrett, supra, at 50-51. And their confidence in their identifications (a point that impressed the trial judge in their favor) is not a reliable indicator of accuracy. See Newsome v. McCabe, 319 F.3d 301, 305 (7th Cir. 2003) (“Psychological research has established that the witnéss’s faith is equally strong whether or not the identification is correct.”); Garrett, supra, at 63 (noting that almost all the eyewitnesses who testified at exonerees’ trials “expressed complete confidence at trial that they had identified the attacker”); see also Williams, 522 F.3d at 812 (discussing how witnesses’ memories realign over time to match earlier statements,- so “trial testimony may reflect more confidence than is warranted”).
Against this weak case for Blackmon’s guilt we balance the defense bolstered by several additional alibi witnesses, most of whom are, as far as we know now, disinterested and unencumbered by prior convictions. See Washington, 219 F.3d at 634 (“Rather than one direct alibi witness with a criminal record, Washington could have had three potentially more credible witnesses[J”). Six of the seven new witnesses recall being at the barbecue at 4:30 p.m., when the murder occurred. Some of them specifically recall interacting with Black-mon at the picnic; all say they never saw Blackmon leave or noticed he was gone. At a casual gathering with people coming and going, it is perhaps unlikely that a single person could vouch for Blackmon’s continued presence there. But adding another witness who can say the same thing makes it less likely that Blackmon would have been able to slip away unseen long enough to commit the murder. Adding six such witnesses makes it less likely still, particularly when some of those witnesses can testify to actual interactions with Black-mon during the course of the afternoon, like Lashun Melton, who says she played cards with him sometime between about 4:00 and 6:00 p.m., or Tiara Topps, who says she flirted with him sometime between about 3:00 and 6:00 p.m. In a large-group setting like this one, the collective weight of the other guests’ testimony is greater than the sum of their individual accounts.
The alibi witnesses’ testimony is not definitive, of course. Maybe they are honestly mistaken or even lying. It’s also possible that everyone who would testify that Blackmon was at the barbecue all afternoon missed seeing him leave during the critical window of time. No one can testify specifically to seeing him close to the critical time of 4:30 p.m., though Sheryce Crowder, the mother of Blackmon’s daugh*1107ter, claimed she was near him all afternoon. All of the witnesses are vulnerable to attacks on their memory at this point. Many of them did not know until years after the barbecue that Blackmon had been arrested for murder. At the time, they would have had no reason to consider the barbecue or Blackmon’s presence at it of any particular importance.
These are fair criticisms. But to establish prejudice, Blackmon does not have to prove actual innocence; he does not even have to show that counsel’s errors more likely than not altered the outcome in his case. Harrington v. Richter, 562 U.S. 86, 111-12, 131 S.Ct. 770, 178 L.Ed.2d 624 (2011), quoting Strickland, 466 U.S. at 693, 697, 104 S.Ct. 2052; Campbell, 780 F.3d at 769; Raygoza, 474 F.3d at 963. He must show only a reasonable likelihood that the outcome would have been different — that is, a likelihood that is “substantial, not just conceivable.” Richter, 562 U.S. at 112, 131 S.Ct. 770, citing Strickland, 466 U.S. at 693, 104 S.Ct. 2052. Given the weakness of the State’s case— the complete lack of any motive, the dearth of physical evidence, and the heavy reliance on the eyewitness identifications of two strangers who saw the killers for only seconds — we conclude that on this record, it is “substantially likely that [Blackmon] could have raised at least a reasonable doubt and had a different outcome at trial” if counsel had provided adequate representation. Thomas, 789 F.3d at 772.
With our § 2254(d)(1) analysis limited to the record that was before the state court, we conclude that the state court’s decision was unreasonable. See Cullen v. Pinholster, 563 U.S. 170, 181, 131 S.Ct. 1388, 179 L.Ed.2d 557 (2011); Mosley, 689 F.3d at 841. We have recognized, however, that although the conclusion that a state court’s summary decision made on the basis of affidavits was an unreasonable application of federal law will often show the petitioner is entitled to relief, “it will not do so always and automatically.” Mosley, 689 F.3d at 853. As in Campbell and Mosley, the state court here summarily dismissed Blackmon’s petition for post-conviction relief. It never tested the truth of the alibi witnesses’ affidavits through any form of adversarial process. “Having concluded that the affidavits and statements, if true, are sufficient to warrant habeas relief, we still have no factual findings on these questions to review, and the record” — which contains no affidavit from defense counsel, for example — “is otherwise ambiguous.” Campbell, 780 F.3d at 772. Under these circumstances, “an evi-dentiary hearing is needed to develop the record on (1) the extent of counsel’s actual pretrial investigation and (2) what these witnesses would have said if called to testify at trial.” Id. Whether Blackmon is entitled to habeas corpus relief will depend on the outcome of that hearing and the district court’s factual findings. See id. at 772-73.4
We VACATE the denial of Blackmon’s petition for a writ of habeas corpus, and REMAND this matter to the district court for consideration of whether Blackmon is actually in custody in violation of the United States Constitution.
. The purported Brady violation was based on the failure of the police to provide Blackmon with Webb’s contact information; Blackmon has not pursued that claim in this appeal.
. In assessing whether Blackmon satisfied the Schlup standard, the district court appears to have extended § 2254(d) deference to the state court’s resolution of Blackmon's actual-innocence claim. We asked the parties to address whether 28 U.S.C. § 2254(d) applies to state court findings regarding a claim of actual innocence. They agree that under these circumstances it does not. In fact, the State relies on Sharpe v. Bell, 593 F.3d 372, 378 (4th Cir. 2010), which states that § 2254(d) “has no application in the context of a Schlup claim because it pertains only to a 'claim that was adjudicated’ in state court.” State courts simply have no occasion to adjudicate Schlup claims as such because those claims address only federal procedural obstacles to relief. In light of the parties' agreement, we assume for the present appeal that no § 2254(d) deference is due to the state court’s resolution of Blackmon’s actual innocence claim and review de novo the district court's decision as to whether no reasonable juror would convict given all the evidence. Coleman v. Lemke, 739 F.3d 342, 349 (7th Cir. 2014).
. Certainly, there are situations in which multiple identifications are of limited value — if the eyewitnesses have contact with one another, for example, or if they make those identifications due to similarly suggestive police procedures. See Brandon Garrett, Convicting the Innocent: Where Criminal Prosecutions Go Wrong 50-51 (2011) (noting that 36% of exon-erees whose cases involved eyewitness identification were wrongly identified by multiple eyewitnesses); Deborah Davis & Elizabeth F. Loftus, The Dangers of Eyewitnesses for the Innocent: Learning from the Past and Projecting into the Age of Social Media, 46 New Eng. L. Rev. 769, 807 (2012) (noting that "converging identifications seem particularly persuasive if one does not recognize that: (1) witnesses can influence one another, and (2) separate witnesses can each be affected by the same suggestive influences (such as suggestive police procedures)”). Blackmon has not attacked the police procedures, however, and nothing in. the record suggests McDowell and Reece had any contact with one another before or while making their identifications.
. We see no reason why Blackmon would not be able to call the salon witnesses, Latonya Thomas and Lajuan Webb, as witnesses in the evidentiary hearing on remand. We have explained why an independent claim of ineffective assistance based on their testimony has been procedurally defaulted (i.e., forfeited). Nevertheless, their testimony might well corroborate the testimony of the alibi witnesses from the barbecue and thus be relevant to the factual findings the district court will need to make at least on the prejudice prong of the Strickland, analysis.