concurring in part and concurring in the judgment.
Meadows has challenged three rulings by the district court: denial of her motion to dismiss the indictment as the product of vindictive prosecution; allowing the government, during Meadows’ cross examination, to refer to that denial; and allowing the government to'make-an improper closing argument based on mathematical probability. I concur in the court’s disposition of her second challenge. Op. 1315-18. As to her first and third challenges,, I concur in part. For the first challenge, the court need only decide, assuming a presumption of vindictiveness, whether the government has satisfied its “minimal” burden to offer “any objective evidence, justifying the prosecutor’s actions,” United States v. Safavian, 649 F.3d 688, 694 (D.C. Cir. 2011); it has, and Meadows does not contend she provided any evidence of “actual vindictiveness,” see United States v. Meyer, 810 F.2d 1242, 1245 (D.C. Cir. 1987). See Op. 1314-15. My analysis of her third challenge differs from that of the court. Given well settled law, the prosecutor’s probability theory first presented in rebuttal closing argument to the jury cannot be excused as merely a “rhetorical flourish.” Op. 1318. In the absence of evidence to support the theory as presented, the error was “plain.” United States v. Olano, 507 U.S. 725, 734, 113 S.Ct. 1770, 123 L.Ed.2d 508 (1993). I nevertheless concur in the judgment because Meadows fails to show the requisite prejudice to obtain reversal of her convictions.
*1320I.
The court affirms the district court’s ruling that there was not a realistic likelihood of vindictiveness because nothing more is involved than a “run-of-the-mill pretrial situation,” see Op. 1312 (quotation omitted), and because Meadows’ “disparate treatment” argument lacked merit, id. at 1314. This court recently acknowledged again that Supreme Court precedent leaves little room for a defendant to demonstrate prosecutorial vindictiveness at the pretrial stage. See United States v. Slatten, Slip. No. 15-3078, at 47, 865 F.3d 767, 2017 WL 3318837, at 20 (D.C. Cir. Aug. 4, 2017) (citing United States v. Goodwin, 457 U.S. 368, 102 S.Ct. 2485, 73 L.Ed.2d 74 (1982)). Still, it is unclear Meadows’ circumstances involved no more than a run-of-the-mill situation.
In explaining the government’s decision to charge Meadows with six counts of unemployment fraud, the Assistant United States Attorney told the district court judge that all other suspects involved in the D.C. Jail contraband scheme “have either been found guilty or plead [sic] guilty. We have just one person—Ms. Meadows—who’s still hanging out there,... We can’t just walk away from that.” Hearing Tr. 22 (Aug. 12, 2014). Additionally, Meadows presented evidence that the United States Attorney’s Office rarely prosecutes unemployment fraud; her case was the single unemployment fraud prosecution in four years even though the D.C. Department of Employment Services had referred more than 30 cases to the Office for prosecution involving amounts greater than the $14,173 Meadows owed. See Def.’s Ex. 1 (A. 116-23). True, Meadows did not show she was treated differently from others who “participated in the same [offense] and conducted themselves in the same manner.” Cf. Op. at 267 (quoting Meyer, 810 F.2d at 1246). But the issue is whether “all of the circumstances, when taken together, support a realistic likelihood of vindictiveness.” Meyer, 810 F.2d at 1246.
Here, the circumstances could suggest that the prosecutor had a “personal stake in [obtaining a] conviction and [a] motivation to engage in self-vindication.” Safavian, 649 F.3d at 692 (quoting United States v. Stanfield, 360 F.3d 1346, 1362 (D.C. Cir. 2004)). In addition to the government’s indictment of Meadows for obstruction of justice when she repeatedly refused to cooperate in the investigation of her boyfriend for murder, prosecutorial frustration was evident from Meadows’ refusal to plead to the contraband charges. Arguably it was evident as well from the Office’s pursuit of rarely prosecuted unemployment fraud charges. Indeed, the district court judge inquired, if the government was so interested in punishing Meadows for her role in the contraband scheme, then “why do you not charge her with thatV’ Hearing Tr. at 22 (Aug. 14, 2014) (emphasis added).
But even assuming that Meadows established a presumption of vindictiveness as to the unemployment fraud charges, see Op. 1314-15, the government offered objective evidence, see Safarían, 649 F.3d at 694, that its charging decision was based on a different “act of dishonesty” by Meadows, Op. 14 (quoting Mem. Op. 15), rather than Meadows’ repeated invocation of her legal rights in the contraband scheme or murder investigation. For this reason, I concur in holding her prosecutorial vindictiveness challenge fails.
II.
The prosecutor’s error in rebuttal closing argument is not so easily dismissed. First, it is well established that counsel may not argue facts that are not supported by evidence before the jury. United States *1321v. Maddox, 156 F.3d 1280, 1282 (D.C. Cir. 1998) (collecting cases). Recently the court repeated “that the prosecutor may not refer in the opening or closing statement to evidence not admitted at trial.” United States v. Davis, 863 F.3d 894, 901 (D.C. Cir. 2017) (quoting United States v. Valdez, 723 F.3d 206, 209 (D.C. Cir. 2013)); see also United States v. McGill, 815 F.3d 846, 916 (D.C. Cir. 2016) (quoting United States v. Jones, 482 F.2d 747, 753 (D.C. Cir. 1973)). The reason is self-evident if a defendant is to receive a fair trial and justice be done. See Berger v. United States, 295 U.S. 78, 88, 55 S.Ct. 629, 79 L.Ed. 1314 (1935). Consequently, repeated prosecuto-rial overreaching during rebuttal closing arguments to the jury is disheartening. See, e.g., United States v. Williams, 836 F.3d 1, 11-12 (D.C. Cir. 2016); McGill, 815 F.3d at 918-19; United States v. Becton, 601 F.3d 588, 593-94, 598 (D.C. Cir. 2010); United States v. Williams, 212 F.3d 1305, 1310-11 (D.C. Cir. 2000). Second, as the last word the jury hears from the parties, the government’s rebuttal 'closing argument can be appropriately forceful in responding to the defendant’s closing argument. At the same time and for the same reason, “the potential for prejudice from an impermissible [rebuttal] closing statement is heightened.” Davis, 863 F.3d at 907 (citing United States v. Holmes, 413 F,3d 770, 776 (8th Cir. 2005) (collecting cases)). When counsel for the parties con-fíne their remarks to the evidence admitted at trial, and reasonable inferences therefrom, the justice system should work as intended. But when, as here, the rebuttal closing argument launches into new terrain and the sandbagging of the defendant is evident, a court must evaluate the likely prejudice to a defendant with considerable care. See McGill, 815 F.3d at 918; cf. United States v. Moore, 651 F.3d 30, 50-51 (D.C. Cir. 2011).
Today the court takes the position that the prosecutor’s use of numbers to calculate probability during rebuttal closing argument was merely a “rhetorical flourish.” Op. 1318. But cf. Davis, 863 F.3d at 907 (quoting United States v. Monaghan, 741 F.2d 1434, 1440 (D.C. Cir. 1984)). Perhaps had the Assistant U.S. Attorney told the jury in rebuttal closing argument that “there is a one in a ga-zillion chance,” or even “a one in a billion chance,” that the computer glitch occurred as Meadows testified, then a court might dismiss her objection as a grandiose or a stylistic form of rhetoric. But that is not what occurred at Meadows’ trial. During rebuttal closing argument, the prosecutor introduced his probability theory in a deliberate and detailed manner, presenting a step-by-step calculation for the jury in attacking Meadows’ defense. Trial Tr. 79-81 (Dec. 22, 2014). To determine “the odds” that a computer glitch affected Meadows and only Meadows, the prosecutor told the jury that first, he multiplied the 49 weeks that Meadows applied for unemployment by the 20,000 claims submitted each week. Second, that he multiplied that number by the “61 key strokes that Ms. Meadows had to make as she sits at her computer entering her name, the Social Security number, the week ending days, her e-mail address, plus the six questions [concerning her eligibility].” Third, that he found a “one in nearly 60 million” chance by dividing 1 by the product of the multiplication at the first two steps. As presented, the prosecutor’s probability theory was more like an indisputable mathematical formula than a permissible rhetorical flourish supported by the evidence. Yet because the theory was unsupported by evidence, the prosecutor’s argument was impermissible under well-settled controlling precedent, see supra at 1320-21, making the error “plain,” see United States v. Williams, 773 F.3d 98, 105 (D.C. Cir. 2014) (citation omitted). *1322Further, the prosecutor’s statements on probability were both flawed and presented for-the first time during rebuttal closing argument.
Implicitly, the prosecutor’s calculation assumed that no other glitches of any kind for any other applicant occurred during the relevant 49-week period. Yet no such evidence was before the jury. At most, a software technician at the Department of Employment Services testified that he was unaware of any widespread computer problems and thought a systemic problem would have been uncovered during the annual audit. Trial Tr. 32-33," 40-41 (Dec. 18, 2014). But a lack of widespread problems would not invariably mean no other errors occurred in the computer system. Indeed, problems with the computer system and website were apparently common enough for the Department to establish a streamlined process for applicants to submit “ticket[s]” to report computer problems to the technicians who were to “£&[] [such] issues” 'as they arose. Id. at 40-41. Because the Department deletes records of specific computer problems after three years, no records were before the jury on the "number of computer errors in 2009 when Meadows was submitting claims. Trial Tr. 175 (Dec. 17, 2014).
Additionally, the prosecutor’s calculation based on no evidence presented a risk of juror confusion. The prosecutor purported to determine “the odds” that a glitch had occurred as Meadows described. Trial Tr. 79 (Dec. 22, 2014). What he really did, however, was assume that a glitch occurred as Meadows described and that no other glitches occurred during that time, and then calculate the error rate based on those assumptions. Oral Arg. 34;07-36;28. Given the subtle distinctions between error rates and the probability an error occurred, and the consequent potential for juror confusion, it may well be that any “analysis of mathematical probability ... must be offered through expert testimony.” Commonwealth v. Ferreira, 460 Mass. 781, 955 N.E.2d 898, 903 (2011); see Op. 1319. None was offered at Meadows’ trial, and no other evidence supported the prosecutor’s newfound theory and its assumptions.
The government responds that the standard jury instructions at the end of the trial before the jury retired to deliberate “alleviated any potential prejudice arising from the prosecutor’s mathematical argument.” Appellee’s Br. 70. The standard instruction does inform the jury that arguments of counsel are not, evidence. Such generic instructions are not inevitably “a cure-all for an erroneous statement.” United States v. Hall, 610 F.3d 727, 742 (D.C. Cir. 2010). Here, the jury “instructions d[id] not address the prosecutor’s error in [rebuttal] closing argument, and the error affect[ed] a central issue,” namely Meadows’ defense to the charges. United States v. Williams, 836 F.3d 1, 16 (D.C. Cir. 2016) (quoting United States v. Watson, 171 F.3d 695, 702 (D.C. Cir. 1999)).
Nevertheless, Meadows has not met her •burden to show prejudice under a plain error standard of review. The government’s case against her was hardly close; it was overwhelming. Put otherwise, the error did not affect the outcome of her trial. See United States v. Gartmon, 146 F.3d 1015, 1026 (D.C. Cir. 1998). Meadows’ testimony about a computer glitch causing her false responses on 49 occasions was facially dubious absent any supporting evidence. The software technician testified that the “glitch” she described would have been “an error in programming and ..". more, systemic than just a one-off occurrence,” Trial Tr. 180 (Dec. 17, 2014); no evidence showed an audit had uncovered such a computer problem. At sentencing, the district court judge even found that *1323Meadows had perjured herself. Sent. Hg. Tr. 12-13 (July 16, 2016). Regardless, based on the evidence before the jury, there is no basis on which the court could conclude that the prosecutor’s probability theory affected a reasonable jury’s rejection of Meadows’ bare assertion that she was the repeated victim of .a computer glitch.
Accordingly, I concur in part and I concur in the judgment affirming Meadows’ convictions.