OPINION

We are asked to decide whether certain allegations of fraud, some of which were known before the parties settled and some of which came to light after settlement, rise to the level of fraud on the court such that relief from the settlement agreement is warranted under Federal Rule of Civil Procedure 60(d)(3). Because the instances of alleged fraud known before settlement cannot justify relief, and the instances discovered after settlement do not rise to the level of fraud on the court under Rule 60(d)(3), we affirm.

I

This case arises from a forest fire that broke out on private property near the Plumas National Forest in northern California on September 3, 2007. The Moonlight Fire, as it came to be known, eventually burned 46,000 acres of the Plumas and Lassen National Forests and resulted in the United States bringing a civil action against private forestry operators, Sierra Pacific Industries, Inc. (“Sierra Pacific”) and Howell’s Forest Harvesting Company (“Howell”), and other individuals to recover damages for that fire.

A

Sierra Pacific contracted with Howell to conduct logging operations on the land where the Moonlight Fire is believed to have started. On the morning of the fire, two Howell employees had been operating bulldozers in the area, but they left without inspecting the site for sparks or signs of fire.

After the fire was spotted from a U.S. Forest Service (“Forest Service”) lookout tower in the early afternoon, Forest Service investigator Dave Reynolds visited the site where the fire was believed to have started. Reynolds interviewed JW Bush, one of the Howell employees who had been working at the site that morning, but the site was too hot to investigate further at the time.

Reynolds returned to the site the following day with Josh White, an investigator from the California Department of Forestry and Fire Protection (“Cal Fire”). According to the Origin and Cause Investigation Report jointly released by the Forest Service and Cal Fire, that day the investigators identified a “general origin area” and a “specific origin area” based on fire indicators in the area. On September 4th and 5th, White and Reynolds took numerous photos and measurements of relevant points within the origin site, and they placed numbered markers and colored flags to mark certain fire indicators and other evidence.

Sierra Pacific and the other defendants allege that White and Reynolds identified a specific point of origin that they marked with a single white flag, and took measurements and photographs of that point. The government denies that the investigators identified this point as the specific point of origin. Instead, the government notes that the investigators took photos of two other rocks, which appeared to have marks from bulldozer blades or treads, and which were ultimately identified in the final report as the points of origin for the fire. Investigators White and Reynolds also used a magnet to search the area and identified metal shavings near these two rocks, which they collected as evidence. Diane Welton, another Forest Service investigator, joined the investigation and visited the origin site on September 8th. Welton agreed with the other investigators’ assessment of the fire’s origin.

Cal Fire and the Forest Service released their joint Origin and Cause Investigation Report in June 2009. The report concluded that one of the Howell bulldozers had caused the fire by striking a rock, which created a spark that ignited forest litter on the ground and eventually broke out into a fire that spread into the surrounding forest.

B

The United States filed this action against Sierra Pacific, Howell, and a num ber of individual defendants (collectively, “the Defendants”) in August 2009. The government sought nearly $800 million in damages caused by the Moonlight Fire and compensation for the resources spent fighting it. The California Attorney General’s office, representing Cal Fire, filed a state court action against the Defendants earlier the same month. The U.S. Attorney and the California Attorney General entered into a joint prosecution agreement, but the two cases proceeded separately.

The parties in this federal case engaged in extensive discovery and motion practice over the next three years. Most relevantly, the government produced a number of documents during discovery that led the Defendants to believe that the government had engaged in fraud during and after its investigation of the Moonlight Fire, in an attempt to blame the fire on them. Specifically, the Defendants discovered photographs and an early sketch that appeared to place the point of origin in a slightly different spot than the final report; an aerial video of the smoke plume that allegedly undermined the government’s point-of-origin determination; an expert report that had used the wrong slope angle in modeling fire dynamics and had not been corrected; and evidence regarding alleged employee misconduct at the Forest Service’s Red Rock Lookout Tower before the fire was spotted. The Defendants also alleged at various points in the pre-trial proceedings that the government had advanced a fraudulent Origin and Cause report based on these cover-ups; had misrepresented the investigator’s interview with Howell employee JW Bush shortly after the fire started; had misrepresented evidence regarding other forest fires started by Howell; had proffered false testimony by the investigators regarding the origin of the fire; and had failed to adequately investigate arson as a possible cause of the fire, particularly in light of evidence that wood cutter Ryan Bauer had been using a chainsaw in the vicinity of the fire on the day it began.

The government moved in limine to exclude much of the evidence supporting the Defendants’ theories of fraud and concealment, and the district court granted this motion in part. The court’s final pretrial order precluded the Defendants from introducing evidence to show conspiracy but permitted them “to introduce evidence that there was an attempt to conceal information from the public or the defense.” The Defendants also wanted to present evidence that the government had failed to investigate possible arson by Ryan Bauer, though the Defendants disavowed any intention of actually proving that Bauer started the fire. The court permitted the Defendants to introduce “evidence indicating arson was not considered to show weaknesses in the investigation following the fire” but precluded evidence demonstrating that a particular person, such as Bauer, had started the fire. Nonetheless, the court’s oral ruling explained that the Defendants would be permitted to present evidence that Bauer was “near the scene, seen by witnesses, and there was no follow-up” during the fire investigation. The court’s written order specifically stated that each of these in limine rulings was “made without prejudice and [was] subject to proper renewal, in whole or in part, during trial.”

Three days before trial was set to begin, the parties reached a settlement agreement under which the Defendants agreed to pay $55 million and transfer 22,500 acres of land to the government. The terms also specified:

Following entry of the settlement agreement, the district court entered judgment dismissing the case with prejudice at the parties’ request.

The state case proceeded after settlement of the federal case. While the state proceedings were pending, several other instances of alleged misrepresentation and fraud came to light. The state case was ultimately dismissed with prejudice before going to trial, and the California Superior Court imposed terminating sanctions on Cal Fire’s attorneys, concluding that they .had engaged in “pervasive misconduct” and “a systematic campaign of misdirection with the purpose of recovering money from the Defendants.”

In the federal case, the Defendants then filed a motion for relief from judgment under Federal Rule of Civil Procedure 60(d)(3), arguing that the government’s alleged misrepresentations throughout the investigation and litigation constituted fraud on the court. That motion is the subject of this appeal.

In addition to the misrepresentations that the Defendants raised prior to settlement, which it re-alleged in the Rule 60(d)(8) motion, the Defendants also alleged newly-discovered fraud. First, Defendants had learned that Ryan Bauer’s father, Edwin Bauer, had accused Sierra Pacific’s legal counsel (apparently falsely) of offering him a bribe to say that his son started the fire. The Defendants alleged that the government knew of this false bribe accusation but fraudulently failed to disclose it, despite representing to the court that there was not a “shred” of evidence pointing to Bauer. The Defendants also alleged that they had learned that the government had instructed the fire investigators to lie about the significance of the white flag by telling them it was a “non-issue” during a meeting prior to the investigators’ depositions.

Finally, the Defendants cited a new report issued by the California State Auditor that some of the funds recovered in state wildfire cases were being put into an extra-legal account called the Wildland Fire Investigation Training and Equipment Fund (“WiFITER”), rather than into the state treasury. In their Rule 60(d)(3) motion, the Defendants alleged that the government had misrepresented the nature of the fund in this federal case. The Defendants also alleged that Cal Fire’s Investigator White stood to benefit from the fund and that his improper financial incentives had tainted the -entire wildfire investigation on which the government had relied.

After an initial status conference on the Rule 60 motion, the district court ordered the parties to submit briefing on the “threshold question” of “whether, assuming the truth of the Defendants’ allegations, each alleged act of misconduct separately or collectively constituted ‘fraud on the court’ within the meaning of Rule 60(d)(3).” The district court also asked the parties to identify whether the Defendants had learned of each alleged act before or after the settlement and dismissal of the case.

After holding an oral hearing on the Rule 60 motion, the district court denied the motion in a detailed written order. With respect to the alleged fraud that the Defendants had known about before settlement — namely the conflicting evidence regarding the point of origin and the alleged misconduct at the lookout tower — the district court concluded that this conduct could not constitute fraud on the court because the doctrine only allows relief from judgment for “after-discovered fraud.” See Hazel-Atlas Glass Co. v. Hartford-Empire Co., 322 U.S. 238, 244 , 64 S.Ct. 997 , 88 L.Ed. 1250 (1944).

With respect to the allegations of fraud that the Defendants claimed to have discovered only after settlement, the district court concluded that relief was barred by the specific terms of the settlement; alternatively, it concluded that relief was unwarranted because the new allegations surrounding the white flag testimony were unsupported by the record, the government did not have a duty to disclose the false bribe accusation made by Edwin Bauer, and the government had not committed fraud on the court through its representations about Cal Fire’s WiFITER fund. The district court concluded that, because none of these allegations constituted fraud on the court, the totality of the government’s conduct similarly failed to rise to that level.

The same day that the district court denied the Defendants’ motion, the U.S. Attorney’s Office for the Eastern District of California posted eight tweets about the outcome of the case via its Twitter account. That evening, a Twitter account allegedly owned by the federal district judge presiding over the Rule 60 motion, which followed the U.S. Attorney’s account, posted a tweet with a link to a news article about the Moonlight Fire. The tweet contained the title of the news article, “Sierra Pacific still liable for Moonlight Fire damages,” as well as a link to the article itself.

The Defendants timely appealed the denial of their Rule 60 motion, arguing that the district court erred in failing to grant the motion and that the judge should be retroactively recused based on the activity of the Twitter account allegedly belonging to him. The district court had jurisdiction over this case under 28 U.S.C. § 1345 , and we have jurisdiction to hear the appeal under 28 U.S.C. § 1291 because the denial of a Rule 60 motion for relief from judgment is a final, appealable order. See United States v. Estate of Stonehill, 660 F.3d 415, 443 (9th Cir. 2011).

In the context of Rule 60(d)(3), we “review denials of motions to vacate for abuse of discretion.” Id. at 443 . Under this standard, we review questions of law de novo, United States v. Hinkson, 585 F.3d 1247, 1261-62 (9th Cir. 2009), and “[a] district court by definition abuses its discretion when it makes an error of law,” Koon v. United States, 518 U.S. 81, 100 , 116 S.Ct. 2035 , 135 L.Ed.2d 392 (1996) (citing Cooter & Gell v. Hartmarx Corp., 496 U.S. 384, 405 , 110 S.Ct. 2447 , 110 L.Ed.2d 359 (1990)). We review the district court’s findings of fact for clear error. Hinkson, 585 F.3d at 1261 . A district judge’s failure to sua sponte recuse himself or herself is reviewed for plain error where, as here, the issue was not raised in the district court. United States v. Spangle, 626 F.3d 488, 495 (9th Cir. 2010).

II

A

Federal Rule of Civil Procedure 60 enumerates several possible grounds for setting aside a judgment. While Rule 60(c) sets a one-year time limit for a Rule 60(b)(3) motion based on “fraud ..., misrepresentation, or' misconduct,” Rule 60(d)(3) provides that “[t]his rule does not limit a court’s power to ... set aside a judgment for fraud on the court” (emphasis added). Therefore, relief based on fraud on the court is not subject to the one-year time limit. Appling, 340 F.3d at 784. Because its motion was made more than a year after the entry of judgment in this case, Sierra Pacific moved for relief under Rule 60(d)(3) and therefore must show fraud on the court, rather than the lower showing required for relief under Rule 60(b)(3).

A court’s power to grant relief from judgment for fraud on the court stems from “a rule of equity to the effect that under certain circumstances, one of which is after-discovered fraud, relief will be granted against judgments regardless of the term of their entry.” Hazel-Atlas, 322 U.S. at 244 , 64 S.Ct. 997 (citing Marine Ins. Co. v. Hodgson, 11 U.S. (7 Cranch) 332 , 3 L.Ed. 362 (1813); Marshall v. Holmes, 141 U.S. 589 , 12 S.Ct. 62 , 35 L.Ed. 870 (1891)). However, the Supreme Court has noted that “[o]ut of deference to the deep-rooted policy in favor of the repose of judgments ..., courts of equity have been cautious in exercising [this] power.” Id. (citing United States v. Throckmorton, 98 U.S. 61 , 25 L.Ed. 93 (1878)). Thus, relief from judgment for fraud on the court is “available only to prevent a grave miscarriage of justice.” United States v. Beggerly, 524 U.S. 38, 47 , 118 S.Ct. 1862 , 141 L.Ed.2d 32 (1998).

Our own cases, similarly, have emphasized that “not all fraud is fraud on the court.” In re Levander, 180 F.3d 1114, 1119 (9th Cir. 1999). “In determining whether fraud constitutes fraud on the court, the relevant inquiry is not whether fraudulent conduct ‘prejudiced the opposing party,’ but whether it ‘harmed the integrity of the judicial process.’ ” Estate of Stonehill, 660 F.3d at 444 (internal alterations omitted) (quoting Alexander v. Robertson, 882 F.2d 421, 424 (9th Cir. 1989)). Fraud on the court must be an “intentional, material misrepresentation.” In re Napster, Inc. Copyright Litig., 479 F.3d 1078, 1097 (9th Cir. 2007), abrogated on other grounds by Mohawk Indus., Inc. v. Carpenter, 558 U.S. 100 , 130 S.Ct. 599 , 175 L.Ed.2d 458 (2009). Thus, fraud on the court “must involve an unconscionable plan or scheme which is designed to improperly influence the court in its decision.” Pumphrey v. K.W. Thompson Tool Co., 62 F.3d 1128, 1131 (9th Cir. 1995) (quoting Abatti v. Commissioner, 859 F.2d 115, 118 (9th Cir. 1988)).

In addition, the relevant misrepresentations must go “to the central issue in the case,” Estate of Stonehill, 660 F.3d at 452 , and must “affect the outcome of the case,” id. at 448 . In other words, the newly discovered misrepresentations must “significantly change the picture already drawn by previously available evidence.” Id. at 435 . In that vein, “[m]ere nondisclosure of evidence is typically not enough to constitute fraud on the court, and ‘perjury by a party or witness, by itself, is not normally fraud on the court’ ” unless it is “so fundamental that it undermined the workings of the adversary process itself.” Id. at 444-45 (quoting In re Levander, 180 F.3d at 1119 ). However, perjury may constitute fraud on the court if it “involves, or is suborned by, an officer of the court.” 12 J.W. Moore, Moore’s Federal Practice § 60.21[4][c]; see In re Intermagnetics Am., Inc., 926 F.2d 912 , 917 (9th Cir. 1991). Despite Sierra Pacific’s arguments to the contrary, our Court and the Supreme Court have consistently applied this standard for fraud on the court even in cases involving' government attorneys, rather than creating some different standard for these cases. Beggerly, 524 U.S. at 47 , 118 S.Ct. 1862 ; Pizzuto v. Ramirez, 783 F.3d 1171, 1181 (9th Cir.2015); Estate of Stonehill, 660 F.3d at 449 .

Finally, relief for fraud on the court is available only where the fraud was not known at the time of settlement or entry of judgment. See, e.g., Hazel-Atlas, 322 U.S. at 244 , 64 S.Ct. 997 (allowing relief for “after-discovered fraud”); Haeger v. Goodyear Tire & Rubber Co., 813 F.3d 1233, 1243-45 (9th Cir. 2016) (analogizing to fraud on the court, where crucial information was concealed until after settlement and entry of judgment), overruled on other grounds, — U.S. -, 137 S.Ct. 1178 , 197 L.Ed.2d 585 (2017); Pumphrey, 62 F.3d at 1133 (finding fraud on the court where crucial information was concealed and came to light after entry of judgment); In re Levander, 180 F.3d at 1120 (same). This limitation arises because issues that are before the court or could potentially be brought before the court during the original proceedings “could and should be exposed at trial.” In re Levander, 180 F.3d at 1120 (citing Gleason v. Jandrucko, 860 F.2d 556, 560 (2d Cir. 1988)); see also id. at 1119-20 (explaining that there is no fraud on the court where “the plaintiff had the opportunity to challenge the alleged perjured testimony or nón-disclosure because the issue was already before the court”). As the district court correctly explained, allowing parties to raise issues that should have been resolved at trial amounts to collateral attack and undermines “the deep rooted policy in favor of the repose of judgments.” Hazel-Atlas, 322 U.S. at 244 , 64 S.Ct. 997 .

The decision in Hazel-Atlas does not undermine this principle, despite the Defendants’ argument that the moving party in that case had some knowledge of the fraud prior to trial and settlement. First, as we have already noted, the Court’s opinion in Hazel-Atlas specifically stated that relief is available for “after-discovered fraud.” Id. And second, the majority opinion in Hazel-Atlas explained that Hazel-Atlas Glass Company had indeed attempted to uncover the suspected fraud before trial, but it had been thwarted by a witness who blatantly lied about the relevant issue. Id. at 242-43 , 64 S.Ct. 997 . After settlement and entry of judgment, it came to light that the witness had been contacted by Hartford-Empire’s attorneys shortly before he lied to Hazel-Atlas’s attorneys, and that Hartford-Empire had compensated the witness shortly thereafter with an $8,000 payment for his lie. Id. Thus, the key information in Hazel-Atlas was revealed only after entry of judgment, ultimately supporting the proposition that relief is available only for fraud discovered after judgment is entered.

Similarly, despite some earlier language suggesting otherwise, see Pumphrey, 62 F.3d at 1133 , our decision in Appling v. State Farm, clarified that where the moving party “through due diligence could have discovered” the alleged perjury or non-disclosure, such fraud does “not disrupt the judicial process” and thus does not constitute fraud on the court. 340 F.3d at 780. Thus, a finding of fraud on the court is reserved for material, intentional misrepresentations that could not have been discovered earlier, even through due diligence.

B

Under the standard described above, the district court properly concluded that Sierra Pacific cannot demonstrate fraud on the court regarding any of the alleged fraud it discovered before settlement. In addition to the fact that these allegations do not constitute “after-discovered fraud,” Hazel-Atlas, 322 U.S. at 244 , 64 S.Ct. 997 , Sierra Pacific had explicitly stated its intention to raise the alleged fraud at trial, and the court’s in limine ruling permitted it “to introduce evidence that there was an attempt to conceal information from the public or the defense.” Thus, “the plaintiff had the opportunity to challenge the alleged perjured testimony or non-disclosure because the issue was already before the court,” In re Levander, 180 F.3d at 1119-20 , and these allegations cannot be grounds for subsequent relief after Sierra Pacific voluntarily settled instead of going to trial.

The district court therefore did riot abuse its discretion in finding that there was no fraud on the court related to the photographs, sketches, and investigator testimony about the white flag; the aerial video and erroneous expert report; the misconduct at the lookout tower; the government’s interview with the Howell employee; the other fires allegedly started by Howell; or the lack of an arson investigation.

C

Nor do the instances of alleged fraud discovered after settlement constitute actionable fraud on the court warranting Rule 60 relief. To begin with, the district court correctly noted that the express settlement terms appear to preclude any relief, even for newly discovered facts or evidence. In agreeing that the “Settlement Agreement ... shall remain in full force as a complete release of Unknown Claims notwithstanding the discovery or existence of additional or different claims or facts' before or after the date of this Settlement Agreement,” it appears that the Defendants bound themselves not to seek future relief, even for fraud on the court. Thus, the district court did not abuse its discretion by finding that relief is precluded on this ground.

Even if the terms of the settlement agreement did not bar relief, the district court properly concluded that relief is unwarranted because the allegations of after-discovered fraud fail to rise to the level of fraud on the court. The Defendants allege three instances of alleged fraud or misrepresentation that they did not discover until after settlement. They argue that each of these allegations demonstrates fraud on the court, and that the district court erred by failing to assume the truth of the allegations, given its specific instructions that the parties brief only the legal sufficiency of the Defendants’ claims. Yet, even assuming the truth of these allegations, we conclude that they do not constitute fraud on the court.

1

First, the Defendants have consistently alleged that Investigators White and Reynolds testified falsely about their investigation of the fire’s origin, specifically regarding the white flag that allegedly marked the initial “concealed” point of origin. The only allegation of after-discovered fraud regarding this dispute is the Defendants’ new allegation that the government attorneys actually suborned this perjury by instructing White and Reynolds to lie in their testimony. During his deposition in the state case, Reynolds mentioned a January 2011 meeting in which the government attorneys spoke with the fire investigators and told them that the issue of the white flag was likely to come up and that the attorneys “saw it as a nonissue.” According to the Defendants, this language is tantamount to the attorneys telling the investigators to conceal any relevant information about the white flag.

Assuming the truth of the Defendants’ allegation on this point, Reynolds’ testimony still does not establish that the investigators were instructed to lie. The attorneys’ comment that they saw the white flag as a “nonissue” is merely an opinion about the relative importance of an element of the case; it is not an instruction to commit perjury. As the government accurately notes, it is not fraud on the court for a party’s attorneys to have their own theory of the case and discuss it with their witnesses. Moreover, the Defendants knew about this meeting before settlement, as Reynolds had explained in his deposition in the federal case that the investigators had met with the attorneys and had discussed the insignificance of the white flag. The slightly different language used by Reynolds in his state deposition did not “significantly change the picture already drawn by previously available evidence,” Estate of Stonehill, 660 F.3d at 435 , nor does it demonstrate that any “grave miscarriage of justice” occurred, Beggerly, 524 U.S. at 47 , 118 S.Ct. 1862 . Accordingly, the district court did not abuse its discretion by denying relief on this ground.

2

The Defendants’ second allegation of after-discovered fraud is that the government failed to disclose Edwin Bauer's accusation that Sierra Pacific’s legal counsel had offered him a bribe to say that his son started the fire. According to the Defendants, this information constituted exculpatory evidence as to the Defendants because it suggests that Edwin Bauer was trying to point investigators away from his son, who may have actually started the fire, by claiming that his son was asked to falsely confess in exchange for a bribe. The Defendants argue that, by withholding this information, the government secured a “critical” in limine ruling limiting the evidence that the Defendants could present regarding its arson theory. The Defendants also contend that the district court failed to accept as true its allegation that this in limine ruling prejudiced the Defendants.

We uphold the district court’s conclusion that relief was unwarranted on these grounds. To begin with, it was not error for the district court to look at the content of the earlier in limine rulings and conclude that the Defendants were not prejudiced by these rulings. In the analogous context of a motion to dismiss, a court can consider matters of public record even when assuming the truth of the allegations, United States v. 14.02 Acres of Land More or Less in Fresno Cty., 547 F.3d 943, 955 (9th Cir. 2008), and the district court here was likewise permitted to consider the record of earlier proceedings even when assuming the truth of the Defendants’ allegations. The court’s oral discussion of the in limine ruling specifically explained that the Defendants would be permitted to present evidence that Bauer was “near the scene, seen by witnesses, and there was no follow-up.” The Defendants thus overstate the impact of the in limine ruling that was allegedly secured through the government’s nondisclosure of the bribe allegation, as the Defendants were still allowed to present evidence relating to its arson theory. Moreover, the district court expressly stated that this ruling was subject to reconsideration during trial. In this context, it was not clearly erroneous for the district court to find that the Defendants were not prejudiced by the ruling.

Next, because the district court correctly concluded that Brady does not generally apply in civil proceedings, see Dist. Attorney’s Office for Third Judicial Dist. v. Osborne, 557 U.S. 52, 69 , 129 S.Ct. 2308 , 174 L.Ed.2d 38 (2009); Fox ex rel. Fox v. Elk Run Coal Co., 739 F.3d 131, 138-39 (4th Cir. 2014), the government did not have a specific duty to disclose the false bribe information, beyond its standard discovery obligations. Under the high standard for a Rule 60(d)(3) motion, a mere discovery violation or non-disclosure does not rise to the level of fraud on the court. Appling, 340 F.3d at 780. In addition, the Defendants could have obtained this information by interviewing Edwin Bauer on their own. See Fed. R. Civ. P. 26(b)(1) (allowing consideration of “the parties’ relative access to relevant information” in determining discovery obligations).

Furthermore, despite the Defendants’ confidence in the probative value of the false bribe accusation, the government’s failure to disclose this information “do[es] not significantly change the story as presented to the district court” prior to settlement, given that the Defendanfs already possessed other circumstantial evidence of arson. Estate of Stonehill, 660 F.3d at 462 . For all of these reasons, the district court did not abuse its discretion by denying relief on this ground.

3

The Defendants’ third allegation is that the government committed fraud on the court by misrepresenting the true nature of Cal Fire’s WiFITER fund, which was later determined by the California State Auditor to be structured such that it was “open to possible misuse.” The Defendants allege that, because the WiFITER fund was not subject to adequate oversight, the funds were used improperly to send Cal Fire investigators to luxury retreats and purchase expensive equipment. The Defendants concede that Cal Fire’s Investigator White had no contingent financial interest in the outcome of the federal case currently before us, because none of the federal recovery was destined for the WiFITER fund, but they argue that White’s contingent interest in the outcome of the state case tainted the entire fire investigation on which both cases relied.

Because our case law requires that a party show willful deception rather than simply reckless disregard for the truth, e.g., Napster, 479 F.3d at 1097 , White’s contingent financial interest only rises to the level of fraud on the court if the government knew about White’s interest and wilfully concealed it. Here, the United States’ only affirmative representations about the nature of the WiFITER fund were that it was “a separate public trust fund to support investigator training and to purchase equipment for investigators” and that it was “a public program established to train and equip .fire investigators.” The Defendants admitted in the disv triet court that they had no evidence that the United States knew of the improper nature of the WiFITER fund; the Defendants alleged only that the government had a duty to fully investigate any agency it was working with and root out any improper motives. The Defendants now argue that, Cal Fire’s knowledge of the fund’s impropriety should be imputed to the United States due to the two entities’ joint prosecution agreement, but the Defendants waived this argument by failing to raise it below. Padgett, 587 F.3d at 985 n.2.

Similarly, the United States could not have had a duty to disclose documents that it did not possess relating to the WiFI-TER fund. The United States represented to the district court that it did not know about or have access to any documents demonstrating the true nature of the fund, and the district court ruled that the Defendants would have to subpoena any such documents from Cal Fire. The Defendants have not challenged the United States’ representation that it did not possess these documents. The Defendants have therefore failed to show that the United States knew about the fund’s improprieties and made “intentional, material misrepresenta tion[s]” on this point. Napster, 479 F.3d at 1097 . Accordingly, the district court did not abuse its discretion by denying relief on this ground.

4

Finally, the Defendants argue that the district court failed to consider the totality of the United States’ conduct, which the Defendants label a “trail of fraud.” See Hazel-Atlas, 322 U.S. at 250 , 64 S.Ct. 997 . Contrary to the district court’s assertion that “the whole can be no greater than the sum of its parts,” a long trail of small misrepresentations — none of which constitutes fraud on the court in isolation — could theoretically paint a picture of intentional, material deception when viewed together. Nonetheless, the instances of possible misinformation in this case do not constitute fraud on the court within the meaning of Rule 60, because almost all of the evidence of alleged fraud was received by the Defendants through discovery and thus was known to them when they made the decision to settle. The three instances of alleged fraud that came to light after settlement, even when viewed together, do not “significantly change the picture already drawn by previously available evidence.” Stonehill, 660 F.3d at 435 . Therefore, the district court did not abuse its discretion by denying relief based on the totality of the circumstances.

5

In sum, none of the allegations of after-discovered fraud, either individually or as a whole, establish that the government committed fraud on the court within the meaning of Rule 60. Accordingly, the district court did not err in denying the Defendants’ motion for relief for judgment under Rule 60(d)(3).

Ill

The Defendants argue that the district court judge assigned to the Rule 60 motion should be recused because of an appearance of bias created by activity on a Twitter account that does not bear his name, but is allegedly controlled by him. As explained above, the Defendants could have raised this issue in the district court following either of the disputed Twitter account’s pre-judgment tweets. Because they failed to do so, plain error review applies. Spangle, 626 F.3d at 495 . The Defendants also filed a motion for judicial notice and a motion for leave to supplement their reply brief with further information regarding the contents of this Twitter account and other related documents. We deny both motions as moot because, under the plain error standard, the allegations do not warrant retroactive recusal even if the judge is the owner of the account.

The Code of Conduct for United States Judges “prescribes ethical norms for federal judges as a means to preserve the actual and apparent integrity of the federal judiciary.” United States v. Microsoft Corp., 253 F.3d 34, 111 (D.C. Cir. 2001). To this end, Canon 2 of the Code instructs judges to “avoid impropriety and the appearance of impropriety in all activities.” Canon 3A(4) prohibits ex parte communications or any “communications concerning a pending or impending matter that are made outside the presence of the parties or their lawyers,” and Canon 3A(6) provides that “[a] judge should not make public comment on the merits of a matter pending or impending in any court.” Canon 3C instructs that a judge must disquali fy himself or herself in a proceeding where his or her impartiality could reasonably be questioned, mirroring the provision of 28 U.S.C. § 455 (a) which mandates that a United States judge “shall disqualify himself in any proceeding in which his impartiality might reasonably be questioned.” The test for recusal under these provisions is “an objective test based on public perception.” United States v. Holland, 519 F.3d 909, 913 (9th Cir. 2008).

The Defendants argue that the judge’s alleged “following” of the U.S. Attorney’s office on Twitter created an appearance of bias, in violation of Canon 2, and constituted an ex parte communication, in violation of Canon 3A(4). They also argue that the judge’s alleged tweet on the evening of his ruling created a further appearance of bias and constituted an impermissible public comment on the substance of a pending case (given the impending appeal), violating Canon 3A(6). Because of these violations, the Defendants argue that the judge was required to recuse himself under Canon 3C and 28 U.S.C. § 455 (a). Even assuming that the judge owned or controlled the disputed Twitter account, these arguments fail.

The claim that an unknown account, not identified with a judge or the judiciary, followed a public Twitter account maintained by the U.S. Attorney does not provide a basis for recusal here. As we know, Twitter is a news and social networking service where users post comments, restricted to 140 characters, in “tweets.” A Twitter account holder may “follow” other Twitter account holders, meaning that the “following” user will receive all of the tweets generated by the other user. Some Twitter users restrict their posts to a private audience. But news organizations, celebrities, and even high-up government officials use Twitter as an official means of communication, with the message intended for wide audiences. Thus, without more, the fact that an account holder “follows” another Twitter user does not evidence a personal relationship and certainly not one that, without more, would require recu-sal. Thus, assuming the account belonged to the district judge, the judge did not plainly err in not recusing himself because he “followed” the U.S. Attorney’s office on Twitter.

For similar reasons, the fact that the Twitter account “followed” the U.S. Attorney does not mean that the public tweets published by the U.S. Attorney constituted improper ex parte communications. The relevant opinion from the Committee on Codes of Conduct explains that concerns of improper communication arise in the context of “the exchange of frequent messages, ‘wall posts,’ or ‘tweets’ between a judge or judicial employee and a ‘friend’ on a social network who is also counsel in a case pending before the court.” Comm, on Codes of Conduct Advisory Opinion 112. The situation in the current case, however, does not present the type of circumstance that the Committee warned against in its opinion. Here, none of the challenged tweets were specifically directed from the U.S. Attorney to the judge, nor have the Defendants alleged that there were any personally directed tweets. Thus, the public tweets did not constitute communication from the U.S. Attorney to the judge. Rather, the relevant tweets from the U.S. Attorney’s account constituted news items released to the general public, intended for wide distribution to an anonymous public audience. Under the circumstances, the social media activity alleged to have occurred in this case did not constitute prohibited ex parte communication.

Finally, the Defendants also allege that the judge’s action in tweeting the link to an allegedly erroneous news article requires recusal. Assuming the challenged tweet was from the judge’s account, it still does not warrant retroactive recusal in this case. The tweet consisted only of the title and link to a publicly available news article about the case in a local newspaper, without any further commentary. Under the standard of review applicable at this stage, the district judge did not plainly err in not recusing himself because he tweeted the link to this news article.

The Defendants rely heavily on United States v. Microsoft Corp., 253 F.3d at 107 , but in fact the conduct in Microsoft was far more problematic: the judge in that case had given numerous secret interviews to the press, in which he spoke extensively about his views on the merits of the case. Id. at 107-11 . Even in In re Boston’s Children First, which the Defendants cite for the proposition that a violation of Canon 3A(c) requires recusal for even the appearance of partiality, the judge had expressed her own views about the case in a published letter to the editor and an interview with a reporter. 244 F.3d 164, 166 (1st Cir. 2001). Here, in contrast, the tweets allegedly posted by the judge expressed no opinion on the case or on the linked news articles. Although “the analysis of a particular section 455(a) claim must be guided, not by comparison to similar situations addressed by prior jurisprudence, but rather by an independent examination of the unique facts and circumstances of the particular claim at issue,” Holland, 519 F.3d at 913 , these cases nonetheless help put the circumstances of the current case in context.

Under the facts and circumstances present here, the single challenged tweet does not amount to “public comment on the merits of a [pending] matter” in violation of Canon 3A(6). Even if the judge’s choice of the particular article he posted and its allegedly inaccurate title could be construed as public commentary, as the Defendants argue, not every violation of the Code of Conduct creates an appearance of bias requiring recusal under § 455(a). Microsoft, 253 F.3d at 114-15 . Here, the three relevant tweets — containing only links to news articles, and coming from an account not publicly identifying a member of the judiciary — do not create an appearance of bias such that recusal is warranted under § 455(a).

For these reasons, under the plain error standard we conclude that there was no appearance of bias created by the instances of alleged conduct in this case, so retroactive recusal is not warranted. Vacatur of the district court’s order is therefore also unwarranted. Nonetheless, this case is a cautionary tale about the possible pitfalls of judges engaging in social media activity relating to pending cases, and we reiterate the importance of maintaining the appear- anee of propriety both on and off the bench.

IV

In deference to the longstanding policy in favor of the repose of judgments, courts have consistently required a very high showing for relief for judgment on the basis of fraud on the court. After voluntarily settling this case and asking the district court to enter judgment based on that settlement, the Defendants’ allegations of newly discovered fraud fail to meet this high standard. We therefore affirm the district court’s denial of the Defendants’ motion for relief from judgment under Rule 60(d)(3), and we decline to order vacatur or direct retroactive recusal.

AFFIRMED.

. Cal Fire and the California Attorney General’s office took no part in the federal case, and the U.S. Attorney’s office took no part in the state case. . Sierra Pacific agreed to pay $47 million and transfer 22,500 acres; Howell’s agreed to pay $1 million; Beaty and the other landowners agreed to pay $7 million. . After the original district court judge re-cused herself from hearing the Rule 60 motion, the case was eventually assigned to a different judge within the Eastern District of California. . The district court also discussed and rejected the Defendants’ allegations regarding an investigator's handwritten notes, and the removal of one of the Assistant United States Attorneys who originally worked on the case. Because the Defendants did not discuss those allegations on appeal, they are waived. Padgett v. Wright, 587 F.3d 983 , 985 n.2 (9th Cir. 2009). .The Defendants argue that de novo review is appropriate because, by asking the parties to brief only the legal sufficiency of the Defendants' allegations, the district court created a procedural posture akin to a Rule 12(b)(6) motion to dismiss. Yet a motion under Rule 60(d)(3) is grounded in the court’s inherent power to set aside a judgment. Such an action “is based on equity,” and we "review a district court's decision to deny equitable relief for an abuse of discretion.” Appling v. State Farm Mut. Auto. Ins. Co., 340 F.3d 769, 780 (9th Cir. 2003). Abuse of discretion review is therefore appropriate here. . The Defendants argue that they had no opportunity to raise this issue in the district court because the challenged tweet was not posted until after judgment was entered. But evidence submitted by the Defendants shows that the same Twitter account had posted several other news articles about the case while proceedings were still ongoing. The Defendants therefore had an opportunity to raise the issue below, and only plain error review is available on appeal. . At the time of Appling , Rule 60(b) contained the language now in Rule 60(d)(3), preserving the court’s right to set aside a judgment for fraud on the court. . Hazel-Atlas does not undermine this conclusion because, unlike in that case where the plaintiffs tried and failed to gain information about the fraud before trial, the Defendants here received numerous documents through discovery that allegedly demonstrated fraud, and they were prepared to present this evidence at trial. . Because we conclude that these allegations do not demonstrate fraud on the court even if taken as true, we need not decide whether the district court erred in failing to assume their truth. . The Defendants also argue that the district court erred by requiring that they act with diligence in attempting to discover the alleged fraud before trial, and that the court made clearly erroneous findings of fact as to whether the Defendants had been diligent. Because none of the alleged instances of fraud rise to the level of fraud on the court regardless of the Defendants’ diligence, we need not and do not reach this issue. . As the Defendants conceded in the district court, they had received the photographs of the white flag and the earlier point of origin sketch during discovery, and the Defendants questioned White and Reynolds extensively about the white flag during their lengthy depositions. . It was proper for the district court to consider the transcript of Reynolds’s deposition that included the “nonissue” comment, which the Defendants filed with the district court in support of their Rule 60 motion. . Indeed, a 2009 internal audit report had failed to reveal any problems with the WiFI-TER fund. . For purposes of this rule, pending matters include those that have been resolved by the court or judge in question but remain pending on appeal. Code of Conduct for United States Judges cmt. 3A(6). . Of course, there are circumstances in which use of social media may create concern. For example, the Judicial Conference of the United States’ Committee on Codes of Conduct has issued an opinion noting that "identifying oneself as a ‘fan’ of an organization” on social media may create the appearance of impropriety. Comm, on Codes of Conduct, Advisory Opinion 112. The ABA’s formal opinion on social media similarly notes that a "judge must be mindful that [an electronic social media] connection may give rise to the level of social relationship or the perception of a relationship that requires disclosure or recusal.” ABA Formal Opinion 462 (Feb. 21, 2013) (emphasis added). Nothing suggests that following a Twitter account under the circumstances here rises to the level of creating an appearance of impropriety. . As explained above, the Defendants could ' have raised this issue in the district court following either of the pre-judgment tweets. Doing so would have allowed a full development of the record. However, because the Defendants’ failed do so, plain error review applies. Spangle, 626 F.3d at 495 . . In making this decision, we do not express any opinion as to the veracity of either party’s factual assertions, attempt to decide any of the underlying issues, or express any opinion as to the troubling issues discussed in the state court opinion. Nor do we make any findings as to the alleged use of the judge’s Twitter account, which was an issue undeveloped in the district court. Those questions must be resolved, if at all, in another forum.