OPINION

GOULD, Circuit Judge:

Elaine Martin appeals her convictions for subscribing false federal tax returns and her sentence for those convictions and several fraud-related convictions.1 First, we address Martin's contention that the district court abused its discretion in admitting evidence that, years before the conduct underlying this case, she had submitted Idaho state tax returns on which she improperly characterized several thousand dollars in personal expenses as deductible farm expenses and had been audited by Idaho tax authorities. Second, we address whether the district court misapplied the Sentencing Guidelines in calculating the losses that resulted from Martin's fraud, where her company was awarded government contracts under programs meant to aid disadvantaged businesses, for which Martin's company did not legitimately qualify. We have jurisdiction under 28 U.S.C. § 1291. For the reasons that follow, we vacate Martin's tax convictions, vacate her sentence, and remand for further proceedings.

I

Martin owned a construction company, MarCon, which specialized in installing steel guardrails and concrete barriers on public highways. MarCon also earned revenue by selling used materials the company removed from its construction sites. But Martin never reported the income from the used material sales to the IRS, and instead kept it off the company books and sent it to a bank account hidden from her external accountants.2 By keeping several hundred thousand dollars of this income off of her personal and company tax returns between 2002 and 2008, Martin avoided paying about $100,000 in income taxes.

Martin also fraudulently obtained government contracts by misrepresenting her assets to qualify for programs designed to aid disadvantaged businesses. A federal program run by the Small Business Administration ("SBA") qualifies small businesses owned by socially and economically disadvantaged persons for certain federal contracts without going through the normal competitive bidding process. Martin also obtained contracts through a state-administered Disadvantaged Business Enterprise ("DBE") program, which sets targets for awarding a percentage of federally-funded contracts to participants. Between 1999 and 2006, MarCon received nearly $20 million from 85 contracts awarded through the DBE program, and successfully performed each contract. MarCon was admitted to the SBA program and awarded three contracts worth nearly $3 million, all of which the company successfully performed.

To prove that Martin knew she had a duty to truthfully report her income on her tax returns, the government was allowed to introduce evidence that Idaho tax authorities had audited Martin and that in tax years 1996 and 1997 she had improperly claimed less than $3,000 as deductible farm expenses on her state tax returns. Martin was accused of incorrectly characterizing student loan payments for her children and expenses related to her divorce as farm expenses. Martin settled the issue without conceding liability.

During closing arguments, the government reminded the jury in its rebuttal of the Idaho audits and argued that Martin knew what she was doing when she subscribed false tax returns because she had tried it before:

The jury convicted Martin of the tax counts and of several fraud-based counts.

At sentencing, Martin, relying on the "procurement fraud rule" found in application note 3(A)(v)(II) of § 2B1.1 of the Sentencing Guidelines, argued for a loss amount of zero. Relying on the "government benefits rule" found in application note 3(F)(ii), the government advocated for a loss amount equal to the total value of the contracts — about $22 million — and the resulting 22-level enhancement that loss amount permitted.

The district court held that the government benefits rule applied. It disagreed, however, that the loss under that rule was $22 million and instead set the loss amount at $3 million, the profit from Martin's fraudulently obtained contracts. Acknowledging that its focus on profit was possibly erroneous, it invoked application note 20(C) and found that a higher loss amount would "overstate the actual loss."

The district court's loss calculation led to an 18-level enhancement. With a base offense level of 7 and additional enhancements for defendant's role and sophisticated means, the adjusted offense level was 31, for which the Guidelines range for someone in criminal history category I is 108 to 135 months. The district court imposed a sentence of 84 months.

The district court also entered an order of forfeiture, pursuant to the parties' stipulation, requiring Martin to forfeit over $3 million.

Martin timely appealed her convictions and sentence.

II

We review a district court's evidentiary decisions for an abuse of discretion. United States v. McFall, 558 F.3d 951, 960 (9th Cir.2009). Even if an evidentiary ruling was incorrect, we will vacate a conviction only if that ruling "more likely than not affected the verdict." United States v. Pang, 362 F.3d 1187, 1192 (9th Cir.2004) (internal quotation marks and citation omitted).3

The district court's interpretation of the sentencing Guidelines is reviewed de novo. United States v. Treadwell, 593 F.3d 990, 999 (9th Cir.2010).

III

We first address Martin's argument that the district court abused its discretion by admitting evidence about her audits by Idaho state tax authorities. We agree and conclude that the error was not harmless. As a result of this substantial error, we vacate Martin's convictions for subscribing false tax returns.

Federal Rule of Evidence 404(b) "provides that evidence of `other crimes, wrongs, or acts' is inadmissible to prove character or criminal propensity but is admissible for other purposes, such as proof of intent, plan or knowledge." United States v. Rizk, 660 F.3d 1125, 1131 (9th Cir.2011) (quoting Fed.R.Evid. 404(b)).

This general rule reflects our concern that a person charged with a crime be convicted only if its elements are proved beyond a reasonable doubt. A person should not be convicted merely because he or she has done prior bad acts. Rule 404(b) will not be violated if the prior bad acts are relevant on some issue in the current prosecution, such as "motive, opportunity, intent, preparation, plan, knowledge, identity, absence of mistake, or lack of accident." Fed.R.Evid. 404(b). But when bad acts are not relevant, they can only be viewed as being presented to inflame prejudice in the trier of fact, in which case they are at odds with our fundamental premises on the need for a fair trial. And even when relevant on some issue, evidence of prior bad acts should not, under Federal Rule of Evidence 403, be admitted when its "probative value is substantially outweighed by dangers of unfair prejudice, confusion on issues, waste of time, or needlessly presenting cumulative evidence." Fed.R.Evid. 403.

In United States v. Bailey, 696 F.3d 794 (9th Cir.2012), the government, prosecuting a defendant for the sale of unregistered securities, introduced an SEC civil complaint alleging that the defendant had previously issued securities in violation of the same SEC rules as those at issue in the criminal trial. We held that the admission of the complaint was an abuse of discretion that required a new trial. Id. at 800-05. We outlined the four part test for admitting evidence under Rule 404(b): the government must show that "(1) the evidence tends to prove a material point; (2) the other act is not too remote in time; (3) the evidence is sufficient to support a finding that defendant committed the other act; and (4) (in certain cases) the act is similar to the offense charged." Id. at 799 (quotations omitted). "If the evidence meets this test under Rule 404(b), the court must then decide whether the probative value is substantially outweighed by the prejudicial impact under Rule 403." Id. (quotation omitted).

Under these standards, admitting evidence of the prior state tax audit for a prosecution of federal tax violations was serious error. Here, the state tax auditors described their investigation and the settlement agreement that Martin had signed, providing more information than merely the civil complaint introduced in Bailey. But this is a distinction that makes no substantive difference. The government introduced evidence that Martin was accused of incorrectly deducting farm expenses on a state tax form in 1996 and 1997, apparently to show her knowledge of federal tax laws related to reporting income in the mid-2000s. But we can perceive no relevant connection between Martin's awareness of rules about the characterization of farm expenses under Idaho tax law, and whether she had knowledge of federal tax law governing the reporting of income. Moreover, there is a substantial probability that the jury took this evidence as proof that Martin is a liar who does not want to pay taxes and will cheat to avoid them — a theme the government emphasized at closing, and a line of thinking the evidence rules are meant to discourage. The government has failed to meet its burden under our normal four-part test for admitting evidence under Rule 404(b). Also, even if relevant, introducing this evidence fails the Rule 403 balancing test.

The government argues that unlike the securities violation in Bailey, the government in criminal tax cases must prove that the defendant knew the tax laws, and that extending Bailey to prohibit evidence of prior audits in criminal tax cases would impair the government's ability circumstantially to prove a defendant's knowledge of the tax laws. We disagree. When the government seeks to admit evidence of a defendant's knowledge, we have "emphasized that the government must prove a logical connection between the knowledge gained as a result of the commission of the prior act and the knowledge at issue in the charged act." United States v. Mayans, 17 F.3d 1174, 1181-82 (9th Cir.1994). Mayans instructs that in cases such as this one, the materiality and similarity prongs of the four-part test merge essentially into one: "similarity is necessary to indicate knowledge and intent because it can furnish the link between knowledge gained in the prior act and the claimed ignorance of some fact in the offense charged." Id. at 1182 (internal quotation marks omitted).

Evidence of an audit by, or settlement with, state authorities for unrelated conduct is only minimally — if at all — probative of Martin's knowledge of the federal tax laws at issue in this case, and there is "an insufficient connection, for Rule 404(b) purposes, between [the prior audit] and the knowledge, in the context of the crime charged," of federal tax laws governing the reporting of income. Id.

To show that the admission of the evidence here was not an abuse of discretion, the government cites several criminal tax cases where evidence of prior encounters with tax officials was used. But all of the cases the government cites involve prior run-ins with the IRS, not state authorities. See United States v. Jackson, 565 Fed. Appx. 662, 662 (9th Cir.2014) (evidence that defendant continued filing false returns after IRS instructed him that his conduct was illegal used to show willfulness); United States v. Matthies, 319 Fed. Appx. 554, 557 (9th Cir.2009) (introduction of IRS publication related to tax protestor arguments used to show defendant was on notice of legal duty to pay income taxes); United States v. Voorhies, 658 F.2d 710, 715 (9th Cir.1981) (evidence that defendant was put on notice of tentative tax deficiencies by IRS audit used to prove willfulness when defendant moved assets overseas the next year). None suggests that learning about obligations related to claiming personal expense deductions for state tax purposes shows knowledge of federal tax laws barring the under-reporting of income. We conclude that the state tax audit evidence was not relevant on the federal tax claims and so should have been excluded under Rule 404(b). But even if relevant, it was unduly prejudicial and not admissible under Rule 403. The government in substance told the jury that Martin had lied on her taxes before and should be convicted of doing so again — an argument not supported by the facts and barred under the rules of evidence.

Was this mistake harmful or harmless? The evidence about the audit was introduced through the testimony of two witnesses and several documents and the government emphasized its importance in closing. Rather than merely arguing that the evidence showed Martin's knowledge of federal tax laws, the government also insinuated, impermissibly, that it showed Martin to be a dishonest person: "[D]oes [a farm] really need to pay for student loans? Well, in Elaine Martin's book it does." Cf. United States v. Brooke, 4 F.3d 1480, 1488 (9th Cir.1993) (stating that evidentiary ruling was not harmless in light of the volume of testimony and references to it in the government's closing argument). The government was permitted to argue at closing that Martin knew what she was doing when she under-reported her income because "she had been there before," and "she tried this before." The government incorrectly used the state audit to make a propensity argument that more likely than not affected the verdict on the false tax return charges. Cf. Bailey, 696 F.3d at 805 (noting government's numerous references to the prior SEC civil complaint at closing); United States v. Brown, 880 F.2d 1012, 1016 (9th Cir.1989) (stating that "continued references to [defendant's] prior bad acts at the Government's closing arguments make it impossible... to say" the error was harmless).4 Martin's convictions for subscribing false tax returns must be vacated.

Considering the totality of the circumstances, however, we reach a different conclusion on the fraud and obstruction of justice convictions. There are several reasons for this. First, except for a brief reminder that income and net worth matter with regard to the DBE and SBA programs at the close of the discussion of the Idaho audit, the government's remarks at closing about the audit related only to the charges of subscribing false tax returns. Second, this propensity evidence likely affected the jury's decision differently on the tax charges than on the other charges. If jurors think that a person cheats on state taxes, they are likelier to infer that such a person cheats on federal taxes than to infer that the person is guilty of a more complex fraud scheme. Third, during a trial that lasted twenty-seven days, there was overwhelming evidence presented that Martin had fraudulently qualified her business for the DBE and SBA programs and had obstructed justice by concealing her true net worth. We see no realistic possibility that a jury would have reached a different conclusion on the fraud and obstruction charges if the state audit had not been mentioned. We conclude that unlike Martin's tax convictions, it was not more likely than not that the evidence of the Idaho tax audit affected the jury's decision on Martin's other charges.

Martin is entitled to a new trial on the tax charges, but not on the other convictions.

IV

Because we vacate Martin's tax convictions, we vacate Martin's sentence. See United States v. Bennett, 363 F.3d 947, 955 (9th Cir.2004) (vacating defendant's entire sentence where one count of conviction was vacated). Martin must be resentenced after liability on a potential retrial for tax violations is resolved. But because we affirm Martin's fraud convictions in a memorandum disposition, the issue of calculating the losses from Martin's fraud is certain to come up again at re-sentencing, and any error by the district court in interpreting the Guidelines may likely be repeated unless we provide guidance here. Accordingly, we next address how the district court should have calculated loss where MarCon gave valuable construction services under the contracts that it gained, but Martin defrauded the government into wrongly concluding that MarCon was qualified to participate in the DBE and SBA programs.

A district court must correctly calculate the Sentencing Guidelines range before imposing a reasonable sentence. See Gall v. United States, 552 U.S. 38, 51, 128 S.Ct. 586, 169 L.Ed.2d 445 (2007). The "commentary in the Guidelines Manual that interprets or explains a guideline is authoritative unless it ... is inconsistent with, or a plainly erroneous reading of, that guideline." Stinson v. United States, 508 U.S. 36, 38, 113 S.Ct. 1913, 123 L.Ed.2d 598 (1993); see also United States v. Jackson, 697 F.3d 1141, 1146 (9th Cir. 2012) (per curiam).

As the general rule for fraud cases, the Guidelines define loss as "pecuniary harm." U.S.S.G. § 2B1.1 cmt. nn.3(A)(i, ii). Pecuniary harm is "harm that is monetary or that otherwise is readily measurable in money." Id. cmt. n.3(A)(iii). They further state that "[l]oss shall be reduced" by "the fair market value of ... the services rendered ... by the defendant ... to the victim before the offense was detected." Id. cmt. n.3(E)(i). This is consistent with the idea that fraud is not always the same as theft. Sometimes, the scheme is to obtain a contract or other opportunity; the scheme still amounts to fraud if a person gains by deceit something to which the person was not entitled, "but [the person] means to perform the contract (and is able to do so) and to pocket, as the profit from the fraud, only the difference between the contract price and [the person's] costs." United States v. Schneider, 930 F.2d 555, 558 (7th Cir.1991); see also United States v. Kopp, 951 F.2d 521, 529 (3d Cir.1991); United States v. Smith, 951 F.2d 1164, 1167 (10th Cir.1991).

Although the value of the contracts in this case is a matter of record, the government does not argue that the United States suffered that amount of "pecuniary harm." It is uncontested that MarCon successfully performed the contracts. Rather, the government contends that one of the "special rules" of loss calculation applies. It invokes the "government benefits" rule of application note 3(F)(ii), which the district court applied, and also invokes the "regulatory approval rule" of application note 3(F)(v). See U.S.S.G. § 2B1.1 cmt. nn.3(F)(ii, v). These special rules apply "[n]otwithstanding" the general rules of application note 3(A). Id. cmt. n.3(F). But in our view, neither special rule applies.

The government benefits rule says that "[i]n a case involving government benefits (e.g., grants, loans, entitlement program payments), loss shall be considered to be not less than the value of the benefits obtained by unintended recipients or diverted to unintended uses, as the case may be." Id. cmt. n.3(F)(ii). Several circuits have held that this rule applies to DBE programs. See United States v. Maxwell, 579 F.3d 1282, 1306 (11th Cir.2009) (citing United States v. Leahy, 464 F.3d 773, 790 (7th Cir.2006), and United States v. Brothers Constr. Co. of Ohio, 219 F.3d 300, 317-18 (4th Cir.2000)).

Leahy reasons that the DBE program "was an affirmative action program aimed at giving exclusive opportunities to certain women and minority businesses. The contracts which these businesses received pursuant to this type of program constitute government benefits." 464 F.3d at 790. "Unlike standard construction contracts, these contracts focus mainly on who is doing the work." Maxwell, 579 F.3d at 1306.

We agree that an "exclusive opportunity" might be a benefit in some sense, but the Guidelines' focus on pecuniary harm suggests a more concrete meaning. The examples given — loans, grants, and entitlement program payments — confirm that this comment deals with unilateral government assistance, such as food stamps, not a fee-for-service business deal. Had Martin been issued food stamps — an entitlement program payment — due to her fraud, the government's loss would be the full value of the stamps. But here Martin obtained contracts, albeit contracts reserved for a special class of contractors of which Martin and her company were not legitimately a part.

It is a "basic canon of statutory construction that when general and specific words are associated ..., then the general words are construed to embrace things similar to those enumerated by the specific words." Hamilton v. Madigan, 961 F.2d 838, 840 (9th Cir.1992); see also Cal. State Legislative Bd., United Transp. v. Dep't of Transp., 400 F.3d 760, 763 (9th Cir.2005). Moreover, if there is any lingering ambiguity as to whether a DBE program is a "government benefit," then the application note cannot apply. See United States v. Leal-Felix, 665 F.3d 1037, 1040 (9th Cir.2011) ("If, after applying the normal rules of statutory interpretation, the Sentencing Guideline is still ambiguous, the rule of lenity requires us to interpret the Guideline in favor of [the defendant].").

Here, the government received significant value from the contracts with Martin because MarCon fully performed. The government has offered no persuasive reason to impose a rule whereby the entire value of the contract would be deemed losses for the government, with no credit given for the value of the services returned. We conclude that the government benefits rule does not apply.

We reach the same conclusion regarding the "regulatory approval" rule, which provides:

U.S.S.G. § 2B1.1 cmt. n.3(F)(v).

The Seventh Circuit has held that the use of fraud to secure minority-business certification fits "squarely within the scheme considered by Application Note 3(F)(v)." United States v. Giovenco, 773 F.3d 866, 871 (7th Cir.2014). While the analogy is fairly arguable, we disagree with the Seventh Circuit's decision to apply that rule. Martin did not falsely pose as a licensed professional or supply goods without obtaining required regulatory approval. Here, too, the rule of lenity counsels against an expansive interpretation of the application note, particularly where, as discussed below, another application note is a closer fit to these circumstances.

We agree with Martin that fraudulently obtaining contracts for disadvantaged businesses falls under the procurement fraud rule, which says:

U.S.S.G. § 2B1.1 cmt. n.3(A)(v)(II). The application note's example of "fraud affecting a defense contract award" is a close fit for the circumstances here. Moreover, the procurement fraud's rule placement within application note 3(A), rather than in note 3(F) with the special rules, indicates that procurement fraud cases fall under the general rule for calculating actual and intended loss. We have said that district courts should "take a realistic, economic approach to determine what losses the defendant truly caused or intended to cause, rather than the use of some approach which does not reflect the monetary loss." United States v. Crandall, 525 F.3d 907, 912 (9th Cir.2008) (quotations omitted). We have also said that "district courts should give credit for any legitimate services rendered to the victims." United States v. Blitz, 151 F.3d 1002, 1012 (9th Cir.1998). Applying the general rule in this and similar cases lets district courts do just that. Applying the special rules, which apply notwithstanding application note 3(A), would not. By fully performing all of the contracts, Martin gave the government considerable value. It would be unjust to set the loss resulting from her fraud as the entire value of the contracts, as the district court itself recognized.

Having decided that the procurement fraud rule, which falls within the general rule for loss calculation, applies, we also reject Martin's contention that the loss amount is nothing because MarCon performed the contracts. The government concedes that the procurement fraud rule's reference to administrative costs is inapplicable because there were no such costs in this case. But neither that nor MarCon's performing the contracts necessarily means that there was no pecuniary harm to the government. The DBE and SBA programs are designed to benefit disadvantaged businesses. It is conceivable that the government paid a premium contract price above what it would pay for other contracts under normal competitive bidding procedures. Any such difference would be an actual loss resulting from Martin's fraud. There was some evidence at trial suggesting that prices paid on DBE and SBA contracts may be higher than those paid for similar services outside those programs. But the government did not show whether there was any such price difference for the contracts awarded to MarCon, or what that difference was. In these circumstances, it is in the interest of justice to remand on an open record to let both the government and Martin submit further evidence and argument on loss amount. On remand, the government may attempt to prove any actual or intended losses resulting from Martin's fraud, including whether there was any pecuniary harm to the government from paying a premium on top of the normal contract price for services comparable to those MarCon provided.

If it is not feasible to determine the actual or intended loss, district courts may use the defendant's gain as another way to measure the loss. See U.S.S.G. § 2B1.1 cmt. n.3(B) ("The court shall use the gain that resulted from the offense as an alternative measure of loss only if there is a loss but it reasonably cannot be determined."). In this case, the government stated below that "the loss from Defendant Martin's fraud can be determined...." This may be a binding admission that precludes reliance on Martin's gain as an alternative measure for loss on remand. Or, in context, it may have been premised on the applicability of the government benefits rule, under which the total value of the contracts awarded to MarCon — a known quantity — would be the loss amount.

Because we conclude that the government benefits rule does not apply, the district court should decide in the first instance whether the government may use the gain rule as an alternative measure for loss. The premium, if any, paid by the government on the contracts in this case is presumably a determinable amount. If that proves to be the case — and if there is no other theory of loss for the district court to consider — the gain rule would not apply.

Finally, there may be other, non-pecuniary losses to the government insofar as Martin's fraud harmed the integrity of the programs, which were designed to help legitimately disadvantaged businesses. There may also be harm, pecuniary or otherwise, to legitimate program participants whose businesses might have received the contracts that were awarded to MarCon. The Guidelines themselves recognize that "there may be cases in which the offense level determined under [§ 2B1.1] substantially understates the seriousness of the offense," U.S.S.G. § 2B1.1 cmt. n.20(A), and give as an example warranting an upward departure a scheme that "caused or risked substantial non-monetary harm," id. cmt. n.20(A)(ii). Even without the authority to depart, district courts have the ability to base an upward variance on a broader concept of harm than the Guidelines contemplate. Nothing in our ruling today is meant to limit district courts' discretion to depart or vary from the Guidelines in appropriate cases, but a sentence must begin with a proper calculation of the Guidelines sentencing range.

The district court misinterpreted the Guidelines and applied the wrong rule. On remand, the losses resulting from Martin's fraud should be calculated under the general rules of application note 3(A) of § 2B1.1 rather than under any of the special rules of application note 3(F), and resentencing should be on an open record to permit both the government and Martin to submit evidence supporting their theories of loss.5

V

We vacate Martin's tax convictions and her entire sentence, and remand for further action consistent with this opinion.

VACATED and REMANDED.