PRECEDENTIAL
UNITED STATES COURT OF APPEALS FOR THE THIRD CIRCUIT _______________
No. 23-1757 _______________
KELVIN ROSA
v.
ATTORNEY GENERAL OF NEW JERSEY; ADMINSTRATOR OF EAST JERSEY STATE PRISON, Appellants _______________
On Appeal from the United States District Court for the District of New Jersey (D.C. No. 1:20-cv-04686) Chief District Judge: Honorable Renée M. Bumb _______________
Argued: March 26, 2025
Before: BIBAS, PHIPPS, and AMBRO, Circuit Judges
(Filed: June 30, 2025)
Adam Elewa [ARGUED] WHIPPLE AZZARELLO 177 Madison Avenue 2nd Floor Morristown, NJ 07960 Counsel for Appellee
Jennifer B. Paszkiewicz [ARGUED] BURLINGTON COUNTY PROSECUTOR’S OFFICE 49 Rancocas Road P.O. Box 6000 Mount Holly, NJ 08060 Counsel for Appellants
_______________
OPINION OF THE COURT _______________
BIBAS, Circuit Judge. Our justice system counts on effective adversaries to ensure just results. That is why the Sixth Amendment guarantees criminal defendants effective assistance of counsel. Kelvin Rosa, a state prisoner seeking federal habeas relief, says he was denied that. He claims that his lawyer failed to do much when the state presented hours of prior-bad-acts evidence suggesting that he had a propensity to commit the very crimes for which he was on trial. The state habeas court denied his claim. To win, he must show that it applied federal law unreasonably. Though this bar is high, Rosa clears it. Given his defense lawyer’s woe- ful inaction, no fair-minded judge could have confidence in Rosa’s convictions. So we will affirm the District Court’s grant of habeas.
2 I. MOUNDS OF PRIOR-BAD-ACTS EVIDENCE AT TRIAL, WITH FEW OBJECTIONS
About two decades ago, New Jersey prosecuted Kelvin Rosa for burglarizing a check-cashing store and shooting a police officer who responded to the burglary. Nearly four months after the burglary, police tried to pull over a driver who had aroused their suspicions. The driver took off, and police followed in a high-speed chase. As they followed the fleeing car, officers watched as its passengers tossed crowbars, screw- drivers, and two-way radios out the window. The fleeing sus- pects eventually crashed, abandoned the car, and fled on foot into a nearby swamp, where they were captured. One of the passengers whom police captured was Rosa. A few days later, in bushes near the car-chase scene, police recovered a stolen Sig Sauer 9 mm pistol. The car’s driver, Mariano Nunez, told police that he and Rosa had burglarized multiple stores together, and that during one of them, Rosa had shot and injured an officer. New Jersey charged Rosa with several types of burglary, theft, armed robbery, aggravated assault, and attempted mur- der—all from the burglary of a check-cashing store. At trial, to place Rosa at the check-cashing store with the attempted- murder weapon, the state put on evidence that Rosa had (1) sto- len the gun the night before, while burglarizing a beef distrib- utor, and (2) carried that gun during the police chase and tossed it out the window. New Jersey law typically excludes such prior-bad-acts evidence as unduly prejudicial. See N.J. R. Evid. 404(b). But it admits that evidence for non-propensity reasons like proving identity. Id.; State v. Cofield, 605 A.2d 230, 234 (N.J. 1992).
3 At a pretrial hearing, the judge thus ruled that this prior- bad-acts evidence could come in—but only to link Rosa to the gun and, with it, to the burglary and shooting. Because “[t]he crimes [we]re not similar in nature,” the jury could not use the evidence to infer that the same person must have committed both; they were not “signature crimes.” JA 96 (first quotation); State v. Fortin, 745 A.2d 509, 516 (N.J. 2000) (second quota- tion) (internal quotation marks omitted). The judge also excluded another piece of evidence. The state had planned to elicit testi- mony from Nunez that Rosa had intended to shoot a police officer during the chase. The judge ruled that this testimony would be so prejudicial that not only would he keep it out, but, if it were introduced, he would declare a mistrial. The trial ended in a hung jury. The state then retried Rosa. At the retrial, a new judge pre- sided but adopted the first judge’s rulings limiting prior-bad- acts evidence. Right before the jury was set to enter the court- room, Rosa’s lawyer tried to stipulate to the identity of the gun, conceding that the “gun retrieved from the side of the road … was the weapon that was fired on the day of the [check-cashing shooting].” JA 101. The state argued that it still needed to intro- duce testimony about the beef-distribution burglary and police chase to link the gun to Rosa—in other words, to show “who possessed it and how he came into possession” of it. JA 102. The trial judge agreed with the state, rejecting the stipulation and letting the evidence in for that limited purpose. Though much of this evidence was properly admitted to tie Rosa to the gun, the state’s presentation of it blasted past the lawful limits. From the jump, it highlighted similarities between the beef-distributor burglary and the cash-checking one. In his
4 opening statement, the prosecutor told the jury that the case against Rosa began not with the check-cashing-store burglary but “actually began the day before,” when Rosa had burglar- ized the beef distributor. JA 109. He also noted how the doors at both facilities were pried and peeled back in a similar way. That opening set the tone for the whole trial. First, an early witness, the cop who responded to the beef-distributor bur- glary, testified in detail about the crime scene, including the distinctive way that the burglars broke in. She told the jury that “the alarm boxes [were] removed,” electrical and phone lines were cut, and a window was broken. Supp. App. 22. Plus, she went well beyond discussing the stolen gun, veering into other weapons and cash that had been stolen from the distributor. The prosecution even showed a photo of the beef distributor’s door peeled back like a tuna can—just as the prosecutor’s opening had described the check-cashing store’s door after it was burglarized. Rosa’s counsel did nothing. After hearing from officers who responded to the check- cashing-store burglary, the jury headed to lunch, with time to mull the similarities. When they returned, the state put on its main witness: Nunez, who had agreed to cooperate with the prosecution. He accused Rosa of belonging to the burglary ring that Nunez led. The prosecutor elicited detailed testimony from Nunez about how he had carried out his burglaries, including the tools and approach he typically used. The prosecutor also asked Nunez whether he had a “routine or a protocol” for each, and Nunez confirmed that they were “all done the same way.” JA 126. Nunez then testified in detail about the beef-distributor burglary, including how he set it up just like the check-cashing- store burglary. He described how they “cut the alarm,” broke a
5 window, and peeled back the door with a crowbar. JA 130. Defense counsel again did nothing. The prosecutor then segued into asking Nunez about the burglary for which Rosa was on trial. Nunez explained that he had picked the check-cashing store because it fit the profile of places he liked to burgle. He testified that Rosa was there, car- rying the gun that Rosa had stolen the night before. Underscor- ing that the crew had done this job just like the others, he explained how they “cut the wires” to the alarm, used a crow- bar, and “pr[ied] the door open.” JA 135–36. The burglary ended when Nunez heard shots and ran outside to find Rosa reloading. Nunez claimed that Rosa had admitted shooting an approaching police officer. The testimony then veered again to other crimes. Nunez be- gan discussing the police chase that had occurred four months after the charged crimes. He said that he and Rosa were starting yet another burglary, this time of a cellphone store, when they heard that police were approaching. Nunez started testifying that the crew had used “the same procedures” in this robbery. JA 142. As he told the jury about casing the store, getting the crew together, and going into the store, Rosa’s counsel finally objected that this testimony did not go to the identity of the gun. After the Court overruled the objection and let the testi- mony continue, Rosa’s counsel objected again. The judge called a sidebar, at which the state argued that this testimony went to whether Rosa had the gun at this robbery and thus dur- ing the ensuing car chase. The judge agreed that setting the background by showing why Rosa had the gun that night “does an awful lot to identify Mr. Rosa as the possessor,” a legitimate end. JA 147. But it ultimately sustained the objection because
6 there was no reason the jury needed to understand all the details of the burglary; it just needed to know that Rosa took part in the attempted burglary that night, possessed the gun, and was in the ensuing car chase. Still, defense counsel did not ask for a limiting instruction, which could have told the jury not to treat the third (attempted) burglary as evidence that Rosa had done the second one. Nunez kept discussing the police car chase. He testified that as Rosa was tossing burglary materials out of the car’s window, Rosa took out the gun. And as the officer approached, Rosa “cock[ed] it.” JA 153. The prosecutor then cut him off, direct- ing him to discuss how Rosa threw the gun out the window. Even though the judge presiding over the first trial had ruled that testimony about Rosa’s intent to shoot the officer would be so prejudicial that it would trigger a mistrial, defense coun- sel never objected or asked for a limiting instruction. Nunez then started telling the jury about fleeing with Rosa through frozen swamps after they had abandoned the car. Rosa’s lawyer objected, noting that Nunez had said enough to place the gun in Rosa’s hand. At sidebar, the judge allowed the evidence in to show Rosa’s consciousness of guilt. But defense counsel failed to ask for a limiting instruction along those lines. Nor did he object or seek a limiting instruction when police testified in lurid detail about this chase, including their daring rescue of Rosa from the frozen swamps. Despite all these witnesses’ testimony dripping with preju- dicial prior-bad-acts evidence, the court issued limiting instruc- tions before only one witness’s testimony and at the end of the trial.
7 The state also put on the officer who arrested Rosa in the Dominican Republic, where he had gone after he made bail on the New Jersey charges. She testified that Rosa had admitted upon arrest “that he had to fire a weapon, [but] that he had not injured anyone.” Appellate Dkt. D.I. 65 at 13. After the state rested its case, Rosa’s counsel moved at side- bar to dismiss, arguing that evidence of the attempted cell- phone-store burglary and eluding incident was unduly prejudi- cial. The judge denied the motion. Rosa then put on his case. He denied taking part in any of the burglaries or shooting the officer. He admitted being a part of the chase but offered an innocent explanation. He said he was heading to a construction job with Nunez when it hap- pened. Nunez began to swerve when he tried to get a phone he had dropped on the floor, so the cops tried to pull them over. Rosa claimed that Nunez had refused to stop because his license was suspended and his insurance was fake. And Rosa testified that he had run out of the car only because police can shoot at evading cars in the Dominican Republic, where he was from. Plus, he said he had left for the Dominican Republic soon after posting bail only because his father became sick and was hos- pitalized there. Finally, he told the jury that he thought Nunez was trying to frame him because he had dated Nunez’s girl- friend while she had been dating Nunez. The jury convicted Rosa on both counts, the judge sen- tenced him to thirty years in prison, the court of appeals af- firmed, and the New Jersey Supreme Court denied review. On state habeas, Rosa claimed that his trial lawyer had been constitutionally ineffective for not objecting to and seeking
8 limiting instructions for the evidence of the beef-distribution and (attempted) cellphone-store burglaries and the police chase. The state habeas court denied his petition. In about two pages of analysis, it rejected that claim, holding that counsel was effective and (in any event) did not prejudice him. On federal habeas, Rosa filed a pro se petition, raising the same ineffective-assistance claim. The District Court granted habeas, holding in a thorough opinion that the state habeas court had unreasonably applied Strickland v. Washington, 466 U.S. 668 (1984). The state now appeals. II. THE STATE HABEAS COURT APPLIED STRICKLAND UNREASONABLY To prove ineffective assistance of counsel, Rosa must show both that (1) his trial counsel performed deficiently and (2) as a result, he suffered prejudice. Id. at 687. On federal habeas, we must defer greatly to the state habeas court’s conclusion that Rosa could not prove these two elements. We presume that its factual findings are right. 28 U.S.C. § 2254(e)(1). On a chal- lenge to its legal ruling, we can grant relief only if its legal ruling was “contrary to, or involved an unreasonable applica- tion of, clearly established Federal law.” § 2254(d)(1). Strickland is clearly established federal law, so we must decide whether it was applied unreasonably. Williams v. Super- intendent Greene SCI, 112 F.4th 155, 167 (3d Cir. 2024). Because the District Court granted habeas without holding an evidentiary hearing, we review its decision de novo. Saranchak v. Beard, 616 F.3d 292, 301 (3d Cir. 2010). We agree with the District Court: In denying relief, the state habeas court applied Strickland unreasonably.
9 A. The state court unreasonably found counsel adequate A lawyer’s performance is deficient if it falls below “an objective standard of reasonableness” measured against “pre- vailing professional norms.” Strickland, 466 U.S. at 688. Under the Antiterrorism and Effective Death Penalty Act (AEDPA), our review of defense counsel’s performance must be “doubly def- erential.” Knowles v. Mirzayance, 556 U.S. 111, 123 (2009). We already defer to trial counsel’s choices, presuming that his “conduct [fell] within the wide range of reasonable profes- sional assistance.” Strickland, 466 U.S. at 689. And under AEDPA, we defer once again to the state habeas court’s analysis of counsel’s performance. Even so, this state habeas ruling “cannot reasonably be jus- tified” under Strickland. Rosen v. Superintendent Mahanoy SCI, 972 F.3d 245, 252 (3d Cir. 2020). The only reason why the state court held that counsel was effective was because he had objected to all the prior-bad-acts evidence at the pretrial hearing. But its analysis ended there. It failed to consider whether counsel should have asked for a limiting instruction after the court overruled his objection and admitted the evi- dence. It also failed to consider whether counsel should have objected later when the testimony at trial exceeded the allowed purpose. There is no room for “fair-minded disagreement” that this analysis, which disregards prevailing professional norms, unreasonably applied Strickland’s deficiency standard. Har- rington v. Richter, 562 U.S. 86, 103 (2011); see Medina v. Diguglielmo, 461 F.3d 417, 428–29 (3d Cir. 2006). 1. When prior-bad-acts evidence floods the state’s case, counsel must act. Even after the court admits prior-bad-acts for
10 a limited purpose, counsel must keep striving to mitigate its impact. “[T]he fact that evidence is admissible does not decide the question whether a limiting instruction should still have been requested by counsel.” Albrecht v. Horn, 485 F.3d 103, 127 (3d Cir. 2007). This is because, as both the U.S. and New Jersey Supreme Courts have long recognized, prior-bad-acts evidence is highly prejudicial. State v. Reddish, 859 A.2d 1173, 1205 (N.J. 2004) (stressing need for “particular caution”); Old Chief v. United States, 519 U.S. 172, 180–82 (1997); N.J. R. Evid. 404(b); Fed. R. Evid. 404(b). It primes juries to draw a forbidden inference: “[B]ecause he did it before, he must have done it again.” Breakiron v. Horn, 642 F.3d 126, 144 (3d Cir. 2011) (internal quotation marks omitted). And it is doubly dan- gerous when “it reveals that the defendant previously commit- ted the very kind of crime” for which he is on trial. Id. Still, prior-bad-acts evidence may be admitted for limited purposes, like proving identity. See N.J. R. Evid. 404(b)(2); Fed. R. Evid. 404(b)(2). But New Jersey courts strictly limit how it may be presented. First, even when evidence is admitted to prove identity, it can go too far. State v. Gillispie, 26 A.3d 397, 416 (N.J. 2011). For instance, when the jury is “exposed to detailed testimony” about the nature of the prior bad act while hearing proof of identity, the admission of such details is “unduly prejudicial and … not outweighed by any probative value.” Id. So even when a party may bring in evidence for a permissible purpose, it may not layer on extra details that do not serve that purpose. Second, in New Jersey, the jury must get timely and tailored limiting instructions, “both when the evidence is admitted and in the final charge,” to focus it on that limited purpose. Id. at
11 417 (emphases added); see also Gov’t of Virgin Islands v. Toto, 529 F.2d 278, 283 (3d Cir. 1976) (requiring 404(b) evidence in federal court to be subject to “stringent instruction at the time of its reception”). Plus, each “instruction should be formulated carefully to explain precisely the permitted and prohibited pur- poses of the evidence, with sufficient reference to the factual context of the case to enable the jury to comprehend and appre- ciate the fine distinction to which it is required to adhere.” Gil- lispie, 26 A.3d at 417 (internal quotation marks omitted). Though our dissenting colleague thinks otherwise, New Jersey law treats timeliness and tailoring as critical to making limiting instructions effective. See id. at 417; State v. Blakney, 912 A.2d 140, 144 (N.J. 2006) (holding limiting instruction insufficient in part because it was not “contemporaneous”). What is defense counsel’s role here? Counsel is not always “constitutionally required” to (1) object each time the prosecu- tion presents 404(b) evidence beyond its limited purpose or (2) “request a limiting instruction any time one could be given.” Albrecht, 485 F.3d at 127. Rather, counsel’s duties depend on the context. We presume that counsel’s silence was sound trial strategy; sometimes, silence is the right call to avoid “high- lighting” fleeting prejudicial material. Buehl v. Vaughn, 166 F.3d 163, 170 (3d Cir. 1999); Albrecht, 485 F.3d at 127; see Strickland, 466 U.S. at 689. But when the 404(b) evidence is not fleeting but floods the state’s case, counsel’s silence cannot reasonably be described as strategic. See Albrecht, 485 F.3d at 127. 2. By failing to dam the flood, counsel fell below constitu- tional norms. Because prejudicial evidence flooded the state’s
12 case against Rosa, his counsel fell below constitutional norms by not taking enough remedial action. First, counsel failed to object when the state put on detailed evidence about the beef-distribution burglary that went beyond proving the identity of the gun. The 404(b) order provided that the evidence could be used only to link Rosa to the weapon and thus the charged crimes. Gillispie, 26 A.3d at 413–14. It spe- cifically noted that it was not admitting the evidence to prove that the other alleged burglaries were “similar” to the burglary for which Rosa stood trial. JA 96. But the prosecution blew past those clear bounds. To prove the identity of the gun, it needed to present testimony only that Rosa had (1) stolen the gun at the first burglary and (2) possessed it at the attempted third one, “without the details involving the actual” heist plans. Gillispie, 26 A.3d at 416. But from the start, the prosecution slipped extraneous, prejudicial details into the trial. In his opening statement, the prosecutor detailed how the beef-distributor burglary was executed and likened it to the check-cashing-store burglary. A cop testified vividly about what the scene of the beef-distribution burglary looked like and the unique procedures that the burglars used. And Nunez explained to the jury how each burglary was “planned and executed” like the others. Id. Such detailed testimony was not needed to prove identity, and it was severely prejudicial. It blurred the crimes charged with the alleged other crimes and primed the jury to conclude that Rosa had a propensity to commit burglaries—a forbidden inference. Any effective defense lawyer should have known to object as soon as this evidence went beyond the 404(b) order. Counsel
13 did so later on during the state’s presentation, when Nunez began to testify about the details of the attempted cellphone-store bur- glary. But his failure to do so throughout this extensive earlier testimony is inexplicable. Second, and more importantly, defense counsel never once asked for a limiting instruction for any of the 404(b) evidence. Even if the state had presented all the 404(b) evidence within the bounds of the 404(b) order, it still would have risked severely prejudicing Rosa absent timely and specific limiting instructions. Any reasonable defense counsel should have known this. And there was no downside to asking for them. As in Albrecht, objecting or instructing the jury would not have inadvertently underscored “fleeting[ ]” evidence that the jury might other- wise have missed; the state had spent substantial time dragging the jury through it. 485 F.3d at 127. Given these facts, we cannot imagine any valid reason not to object or to seek prompt, focused limiting instructions— especially when counsel knew the evidence was objectionable. There was zero reason not to. Still, our dissenting colleague paints counsel’s silence as a sound tactic worthy of deference. In his view, counsel did not need to protect Rosa from this evidence because his whole theory of the case was that Rosa did not commit these prior bad acts. True, we strongly presume that defense counsel’s decisions were tactical. Strickland, 466 U.S. at 689. Counsel must ration his objections and avoid calling attention to matters that the jury might overlook. But calling something a “tactic” does not make it so. Sometimes counsel’s performance is so woeful that it “cannot be characterized as the product of strategic
14 judgment.” United States v. Gray, 878 F.2d 702, 711 (3d Cir. 1989). As the Supreme Court instructs, we may defer to tactics only if they are reasonable and well-founded. Roe v. Flores- Ortega, 528 U.S. 470, 481 (2000); Strickland, 466 U.S. at 689–91. Counsel’s inaction here was neither. He simply messed up. If counsel’s theory was that Rosa never committed these other burglaries, then it made no sense to let the jury hear for hours about how he carried them out. Far from bolstering counsel’s theory, his silence undermined it. In fact, under any “objective standard of reasonableness,” a lawyer implementing that the- ory should have tried to stop this highly suggestive evidence from coming in or at least mitigate its harm once it did, “even [if it came] at the expense of purportedly clever theories.” Workman v. Superintendent Albion SCI, 915 F.3d 928, 944 (3d Cir. 2019). Doing so would not have been “tangential to core defense strategy,” but key to Rosa’s defense. Dissent at 2. So no reasonable trial strategy supported counsel’s inaction; he was simply asleep at the wheel. His silence was so unjustified that the state habeas court’s contrary ruling was unreasonable. B. The state court unreasonably found no prejudice Defense counsel’s performance is prejudicial when it cre- ates “a reasonable probability” of a different result. Strickland, 466 U.S. at 694. In making this judgment, we look at all the evidence at trial. Id. at 695. An error is far more likely to be prejudicial if the state’s case is weak. Id. at 696. The state habeas court concluded that counsel did not prejudice Rosa because a state court on direct appeal had found that admitting
15 the evidence did not prejudice him. We reject that conclusion as unreasonable too. The state habeas court discounted prejudice because it thought the evidence of guilt was strong. But the record tells a different story. No physical evidence linked Rosa to the crime. The shooter could have been Rosa, Nunez, or any of the look- outs at the burglary. For example, ballistics evidence showed that the gun used to shoot the officer at the check-cashing-store burglary matched the gun found on the road after the police chase. But that does not show who used the gun at the check- cashing store. In fact, when police arrested Nunez and sug- gested that he might have shot the officer himself, Nunez did not deny it, but said only, “you got to prove it.” JA 162. Only later did Nunez say Rosa did it and testified against him. And Nunez had every incentive to blame Rosa: Nunez not only evaded a murder charge, but also got a big sentence discount in exchange. All the other evidence against Rosa at most showed that he was guilty of something, not that he was guilty of the specific crimes for which he stood trial. For instance, he was in the car that eluded police and from which the gun and burglary tools were tossed. Plus, Rosa fled through the swamps (showing consciousness of guilt). And he admitted on arrest that he had fired a gun but injured no one. This evidence certainly supports the inference that Rosa had something to hide. But to sustain a conviction, the state needed more than that. It needed to prove that Rosa was guilty of burglarizing the check-cashing store or attempted murder—crimes that required proving specific intent. N.J. Stat. Ann. § 2C:18-2 (burglary); State v. Rhett, 601 A.2d 689, 692 (N.J. 1992) (attempted murder). And at least one juror
16 agreed that the state’s case was weak. When the state first tried Rosa, nearly the same evidence produced a hung jury, a telltale sign that the case was close. Bey v. Superintendent Greene SCI, 856 F.3d 230, 242 (3d Cir. 2017). So the incriminating evi- dence was “far from overwhelming.” Breakiron, 642 F.3d at 140, 147. Given the thin proof of guilt, prejudice is not a close call. There is a reasonable probability that defense counsel’s failure to do more to check hours of prior-bad-acts evidence influ- enced the jury’s verdict. The state habeas court unreasonably found otherwise. Cf. Albrecht, 485 F.3d at 129 (holding in a nearly identical context that prejudice was “very close” even though the evidence of guilt was “ample” and corroborated). Yet the state and our dissenting colleague claim that the two belated, boilerplate limiting instructions were enough to guard against prejudice. But they were not. True, we ordinarily pre- sume that the jury followed those instructions. United States v. Allinson, 27 F.4th 913, 926 (3d Cir. 2022). But as both New Jersey courts and this court have repeatedly recognized, even a well-heeded instruction may not suffice to purge the taint. Toto, 529 F.2d at 282–83; State v. Stevens, 558 A.2d 833, 844 (N.J. 1989) (“[T]he inherently prejudicial nature of such evi- dence casts doubt on a jury’s ability to follow even the most precise limiting instruction.”). After all, once prior-bad-acts evidence reaches the jury, it becomes “difficult, if not impossible, to assume continued integ- rity of the presumption of innocence. A drop of ink cannot be removed from a glass of milk.” Toto, 529 F.2d at 283. So even with the two catch-all instructions, Rosa may well have been
17 prejudiced by defense counsel’s failure to try to keep as much of the extraneous 404(b) evidence out as possible. Of course, the judge could have overruled his objections. But when defense counsel finally roused himself to object to evidence of the cell- phone-store burglary, the trial judge partially sustained his objec- tion. So if counsel had done his job properly by fighting to keep this evidence out altogether, he could have reduced the prejudice. Plus, the limiting instructions had two defects that reduced what little remedial effect they would have had: First, they came too late. They were not given “at the time the evidence [was] admitted.” United States v. Caldwell, 760 F.3d 267, 277 (3d Cir. 2014); see also Gillispie, 26 A.3d at 417 (“The instructions should be timely given both when the evi- dence is admitted and in the final charge.”). As New Jersey and federal courts have long recognized, prior-bad-acts evidence “may indelibly brand the defendant as a bad person.” Blakney, 912 A.2d at 143. Contemporaneous instructions reduce the risk that this implication will stick in jurors’ minds, leaving them unable to evaluate the rest of the state’s case against the defend- ant carefully and correctly. See id. at 143–44. But here, the jurors had heard hours of prior-bad-acts evidence and gone off together to eat lunch before they were told not to draw the infer- ence that they likely already had. Second, and relatedly, neither instruction was tailored to each piece of prejudicial evidence. Neither one forbade the jury to consider the gun-cocking testimony, even though the 404(b) order had barred that testimony as so prejudicial it could cause a mistrial. That testimony was dangerous. It could have led the jury to think Rosa had a habit of shooting at cops and so must
18 have done it earlier during the check-cashing-store burglary. The jury should have been told that it could not consider that evidence at all. So the two catch-all instructions came far too late to undo the damage done. We do not, as our dissenting colleague insists, adopt “a blanket rule” that limiting instructions are ineffective unless given immediately. Dissent at 7. Prejudice under Strickland al- ways depends on context. Our ruling depends on the amount and severity of the prior-bad-acts evidence here (both of which are astonishing) and the thin evidence of guilt, as shown by the hung jury at the first trial. To be clear, even woeful deficiency will not always be prej- udicial. But here, the evidence was weak, and defense counsel’s silence mattered. On this record, we see no room for debate: Rosa’s trial was unfair. There is a reasonable chance that effec- tive defense counsel would have moved the needle, ending the second trial in at least a mistrial or a hung jury like the first one. The state habeas court’s contrary ruling was unreasonable. ***** The federal standard for habeas is “difficult to meet,” not impossible. Harrington, 562 U.S. at 102. On habeas, federal courts must defer to state ones, doubly so when checking defense counsel’s performance. But the record reveals that this is the rare case where the state-court decision is unjustifiable “beyond any possibility for fair-minded disagreement.” Id. at 103. Defense counsel denied Rosa a just trial by not doing more to shield him from many hours of prior-bad-acts evidence. Because the state habeas court’s contrary ruling was unreasonable, we will affirm the District Court’s grant of habeas corpus.
19 Rosa v. Attorney General of New Jersey, No. 23-1757 PHIPPS, Circuit Judge, dissenting.
A claim of ineffective assistance of counsel under the Sixth Amendment requires two showings: objectively deficient performance by counsel and resulting prejudice. See Strickland v. Washington, 466 U.S. 668, 687 (1984). When an application for habeas corpus under 28 U.S.C. § 2254 relies on such a claim to challenge a state-court judgment, two tiers of deference apply. First, for the performance prong, the well- founded strategic choices of state-court defense counsel receive near-absolute deference. See id. at 690 (emphasizing that tactical choices “made after thorough investigation of law and facts relevant to plausible options are virtually unchallengeable”). 1 Second, the “last reasoned” state-court rulings on the performance and the prejudice prongs receive deference under the Antiterrorism and Effective Death Penalty Act of 1996, Pub. L. No. 104-132, § 104, 110 Stat. 1214, 1219 (codified in relevant part at 28 U.S.C. § 2254(d)). Laird v. Sec’y, Pa. Dep’t of Corr., 129 F.4th 227, 243 (3d Cir. 2025) (quoting Abdul-Salaam v. Sec’y of Pa. Dep’t of Corr., 895 F.3d 254, 265 (3d Cir. 2018)).
In evaluating Kelvin Rosa’s § 2254 application claiming ineffective assistance of counsel, the Majority Opinion does not grant either tier of deference. Instead of respecting defense counsel’s trial strategy – one that previously resulted in a hung 1 See also Cullen v. Pinholster, 563 U.S. 170, 196 (2011) (requiring courts to “‘indulge [the] strong presumption’ that counsel ‘made all significant decisions in the exercise of reasonable professional judgment’” (alteration in original) (quoting Strickland, 466 U.S. at 689–90)); Yarborough v. Gentry, 540 U.S. 1, 8 (2003) (per curiam) (explaining the “strong presumption” that decisions by trial counsel were strategic rather than neglectful – a presumption that “has particular force where a petitioner bases his ineffective- assistance claim solely on the trial record”).
1 jury – the Majority Opinion faults Rosa’s counsel for failing to incessantly object or request limiting instructions with respect to an issue that was tangential to core defense strategy. And rather than affording AEDPA deference to the last reasoned state-court decision on the performance and the prejudice prongs, see State v. Rosa, No. 06-10-01443-I, slip op. at 14–18 (N.J. Super. Ct. Law Div. Feb. 27, 2018), the Majority Opinion declares that “no fair-minded judge could have confidence in Rosa’s convictions.” Maj. Op. at 2. In my view, Rosa cannot overcome either layer of deference, and I respectfully dissent.
A. The Majority Opinion Does Not Afford the Requisite Deference to the Strategic Choices of Defense Counsel. As to the first tier of deference, Rosa’s trial counsel had a strategy. He did not contest that the burglaries took place; he argued instead that Rosa did not commit them and that Rosa was framed by Mariano Nunez – a key witness for the prosecution. Under this theory, Nunez was portrayed as a dangerous and “conniving” “mastermind” of a burglary ring who had decided to pin his crimes on Rosa, an innocent man who happened to be in Nunez’s car when Nunez was caught by the police. Retrial Day 1 Tr. 24:20–21, 25:10 (App. 116–17). Thus, it was not necessary to exclude Nunez’s testimony as to the other crimes at all costs: under this theory, those burglaries, like the burglaries and robberies with which Rosa was charged, were committed by persons other than Rosa. Any abstract doubt about the effectiveness of this strategy gives way to a concrete result – the first trial resulted in a hung jury. And because trial tactics grounded in reason and professional judgment are “virtually unchallengeable,” counsel’s pursuit of a strategy that previously proved successful qualifies for that high level of deference. Strickland, 466 U.S. at 689–91 (explaining that the degree of deference increases for the decisions of counsel that are well founded).
2 Affording such deference is not the same as endorsing counsel’s performance or finding it beyond reproach. Rather, deference to counsel’s performance must be understood in light of the rule that a guilty verdict requires juror unanimity. See N.J. Ct. R. 1:8-9 (requiring all criminal jury verdicts to be unanimous effective September 1, 1998). See generally Ramos v. Louisiana, 590 U.S. 83, 93 (2020) (holding that “the Sixth Amendment’s right to a jury trial requires a unanimous verdict”). And with the need to persuade only one juror that the prosecution has not proved its case, it may be rational for defense counsel to pursue high-variance strategies principally aimed at persuading a single juror of reasonable doubt, even if that approach would be unpersuasive to the average juror – or even most jurors.
Even still, the Majority Opinion views defense counsel’s high-variance strategy as constitutionally deficient. According to the Majority Opinion, trial counsel should have objected and, if he lost, requested a limiting instruction as to each reference to the Amaro and Nextel burglaries.
That is a strange fault to find because Rosa’s trial counsel secured two limiting instructions that governed the jury’s consideration of those burglaries. The first one was given on the second day of trial when the judge announced that the other-crimes evidence was admitted only for the “very limited purpose of permitting the State to demonstrate or attempt to prove possession of the handgun involved and the identity of the person who possessed it.” Retrial Day 2 Tr. 107:11–15 (App. 164). In giving that instruction, the judge underscored that even if the jury found that Rosa “possessed or . . . was involved at Amaro Foods or at the Nextel store,” that would not allow the conclusion that “he’s, therefore, guilty of the present offense.” Id. at 107:16–18. The judge gave a second, very similar limiting instruction on the last day of trial before the jury began deliberating. He explained that it is proper to “exclude evidence that a defendant has committed other
3 crimes, or wrongs or acts when it is offered only to show that he had a disposition or a tendency to do wrong and, therefore, must be guilty of the charges before the [c]ourt.” Retrial Day 5 Tr. 16:2–7 (App. 192). That instruction made clear, however, that such other-crimes evidence may be introduced “for a very specific and narrow purpose,” such as “the issue of identity or possession of the gun.” Id. at 16:12–18.
Under New Jersey law, a limiting instruction is adequate if it distinguishes between the proper and improper purposes of the evidence. See State v. Fortin, 745 A.2d 509, 518 (N.J. 2000); State v. Gillispie, 26 A.3d 397, 417 (N.J. 2011) (citing Fortin with approval); State v. Rose, 19 A.3d 985, 1000 (N.J. 2011). Both limiting instructions did so, very much along the lines of a limiting instruction previously approved by the New Jersey Supreme Court for this purpose. See Fortin, 745 A.2d at 519 (including an example text of a limiting instruction on other-crimes evidence that would satisfy the requirements of New Jersey law).
Moreover, as a principle of law, jurors are presumed to follow limiting instructions. See United States v. Scarfo, 41 F.4th 136, 181 (3d Cir. 2022); State v. Burns, 929 A.2d 1041, 1054 (N.J. 2007) (stating that this presumption is “[o]ne of the foundations of our jury system”). Thus, Rosa’s counsel cannot be credibly criticized for not seeking additional limiting instructions: jurors are presumed to follow a single instruction, and here two were given. See Albrecht v. Horn, 485 F.3d 103, 127 (3d Cir. 2007) (disabusing the notion that “[t]rial counsel is . . . constitutionally required to request a limiting instruction any time one could be given”). Trial counsel’s choice to refrain from additional objections and requests for limiting instructions makes even more sense in light of the reality that it may be “strategically preferable to omit such a request since the instruction might have the undesired effect of highlighting the other crimes evidence.” Buehl v. Vaughn, 166 F.3d 163, 170 (3d Cir. 1999) (Alito, J.).
4 The Majority Opinion nonetheless declares the performance of Rosa’s trial counsel constitutionally deficient. That degree of second guessing – bolstered by fourteen years of hindsight – exceeds the permissible scope of federal court review of an ineffective-assistance-of-counsel claim in a § 2254 petition. Rather, defense counsel’s trial strategy, which previously resulted in a hung jury, merits deference.
B. The Majority Opinion Does Not Afford AEDPA Deference to the New Jersey State-Court Decisions on the Performance and Prejudice Prongs. As to the second tier of deference, the last reasoned state- court decision on Rosa’s ineffective-assistance-of-counsel claim concluded that Rosa could not satisfy either prong of that claim. Rosa, No. 06-10-01443-I, at 14–18; see also State v. Rosa, 2020 WL 1907810, at *4 (N.J. Super. Ct. App. Div. Apr. 20, 2020) (per curiam) (denying the claim and incorporating the reasoning of the Law Division’s decision). Under AEDPA deference, those determinations should be upheld.
In evaluating the performance prong, the Superior Court hewed closely to precedent and concluded that defense counsel’s performance was not unconstitutionally deficient. Rosa, No. 06-10-01443-I, at 16 (“The record shows . . . that trial counsel acted objectively reasonably . . . .”). But even if reasonable minds could differ on the performance of Rosa’s counsel, that would not be enough to overcome AEDPA deference. See Laird, 129 F.4th at 241. More forcefully still, even if the Superior Court had erred in its assessment of counsel’s performance, that alone would not overcome AEDPA deference, which requires “far more than that the state court’s decision was ‘merely wrong’ or ‘even clear error.’” Shinn v. Kayer, 592 U.S. 111, 118 (2020) (per curiam) (quoting Virginia v. LeBlanc, 582 U.S. 91, 94 (2017) (per curiam)). Instead, to overcome AEDPA deference, the Superior Court’s decision must have been “so obviously wrong that its error lies 5 ‘beyond any possibility for fairminded disagreement.’” Id. (quoting Harrington v. Richter, 562 U.S. 86, 103 (2011)). And the decision of the Superior Court is not so obviously wrong; indeed, the strategic choices of Rosa’s counsel had previously resulted in a hung jury. Rather than respect the rulings of the court with direct review authority over New Jersey’s evidence rules, the Majority Opinion draws its own conclusions regarding counsel’s performance relative to those rules, and in so doing it disregards AEDPA deference. See 28 U.S.C. § 2254(d)(1); see also Bradshaw v. Richey, 546 U.S. 74, 76 (2005) (per curiam) (“[A] state court’s interpretation of state law . . . binds a federal court sitting in habeas corpus.”).
The Majority Opinion continues that trend with respect to the prejudice prong. In adjudicating Rosa’s post-conviction relief challenge, the Superior Court concluded that the trial judge’s limiting instructions were sufficient to negate any hypothetical prejudice. Rosa, No. 06-10-01443-I, at 15–18 (ruling both that Rosa suffered no prejudice from the ‘other crimes’ evidence and that the limiting instructions were sufficient); see also Rosa, 2020 WL 1907810, at *5 (affirming the prejudice ruling); State v. Rosa, 2014 WL 10186979, at *12–13 (N.J. Super. Ct. App. Div. Aug. 3, 2015) (per curiam) (finding no prejudice on direct appeal). There is nothing infirm about that decision. In light of the “almost invariable assumption of the law that jurors follow their instructions,” it may be supposed that the jurors followed the clear, specific instructions that they twice received. Scarfo, 41 F.4th at 181 (quoting Richardson v. Marsh, 481 U.S. 200, 206 (1987)); see also Burns, 929 A.2d at 1054. Those instructions stated that the other-crimes evidence could be considered only for identity purposes, not for propensity purposes. See Retrial Day 2 Tr. 107:15–25 (App. 164); Retrial Day 4 Tr. 16:1–23 (App. 192). There is no prejudice stemming from a lack of additional, repetitive limiting instructions: again, jurors are
6 presumed to follow a single instruction, and here there were two. 2
To reach a contrary outcome, the Majority Opinion reasons that to be effective, a limiting instruction must be issued immediately following the at-issue evidence. That is a new rule. At most, this Court has emphasized the value that temporal proximity plays in limiting instructions, but it has not gone so far as to hold that delay extinguishes the efficacy of a limiting instruction. See United States v. Caldwell, 760 F.3d 267, 277 (3d Cir. 2014) (stating that limiting instructions “must” be given if requested but only “should” be issued close to the evidence’s introduction). Nor is such a blanket rule tenable. The Supreme Court has already spoken on this issue and in doing so has made clear that to overcome the presumption that a jury followed an instruction, there must be an “‘overwhelming probability’ that the jury will be unable to follow” the instruction. Greer v. Miller, 483 U.S. 756, 766 n.8 (1987) (quoting Richardson, 481 U.S. at 208). Applied here, providing limiting instructions at two times during the trial – on day two (of a five-day trial) and before the deliberations – hardly established an overwhelming probability that a jury would not follow either instruction. Moreover, applying this new rule here is inconsistent with AEDPA deference: even a modicum of respect for New Jersey state courts would require deference to their assessment of the efficacy of the limiting instructions, especially those, like the ones here, that were
2 Even if there were a shortcoming in those jury instructions, that alone would not satisfy the prejudice prong. Indeed, in State v. Stevens, 558 A.2d 833 (N.J. 1989), a case relied upon by the Majority Opinion, the New Jersey Supreme Court held that deficiencies in the limiting instructions about three pieces of other-crimes evidence involving the sexual exploitation of women in the trial of a police officer charged with official misconduct and criminal coercion related to the sexual exploitation of two women were not prejudicial. Id. at 844. 7 modeled off instructions endorsed by the New Jersey Supreme Court. See Fortin, 745 A.2d at 519; see also Laird, 129 F.4th at 240 (explaining that AEDPA deference can be overcome only by a showing that the state court “arrive[d] at a conclusion opposite to that reached by [the Supreme] Court on a question of law” (second alteration in original) (quoting Williams v. Taylor, 529 U.S. 362, 413 (2000))). 3
*** In sum, this is a double deference case, and affording deference to the strategic choices of defense counsel and AEDPA deference to the decision of a New Jersey state court means that Rosa’s § 2254 application cannot be granted.
3 The Majority Opinion also appears troubled by the fact that Rosa’s trial counsel did not move for a mistrial once Nunez testified that Rosa cocked his gun as officers approached. Those misgivings, however, have no bearing on Rosa’s claim here because he does not argue that his counsel was ineffective for failing to move for a mistrial on that ground. Nor did Rosa raise that argument in his state-court collateral attacks on his conviction. See Rosa, 2020 WL 1907810, at *3. 8