Swiss Bank Corporation employed Judith Perlman in Chicago as a lawyer until September 1994, when she quit. Ever since, Perlman and Swiss Bank have been contesting whether she is entitled to disability benefits. Perlman has serious medical problems. An automobile accident in 1988 caused trauma to her digestive system and led to some additional problems, such as migraine headaches. Perlman did not work for a year after the accident and received disability benefits during that period. She took shorter leaves in 1991 and 1992, again receiving disability benefits. Perlman regularly sees a psychiatrist for *977assistance in coping with the stress caused by the accident and her medical conditions. But between September 1992 and September 1994 she worked full time. Swiss Bank was satisfied with her work, believes that she is physically and mentally able to continue, and contends that she quit for personal rather than medical reasons, seizing an opportunity to leave the labor force and move to the Wisconsin countryside.
unum Life Insurance Company administers Swiss Bank’s disability plan, a “welfare benefit plan” covered by ERISA. For short-term disability UNum acts solely as an administrator; payment comes from Swiss Bank’s accounts. For disability after the first 26 weeks unum is both administrator and insurer; it pays any award. Immediately after quitting, Perlman applied for short-term benefits. After seeking and obtaining additional medical records, unum said no: its letter states that the medical conditions documented in the information Perlman had provided “do not prevent you from performing the material duties of your occupation.” Perlman appealed to a higher level of unum’s staff but was unsuccessful. A letter in July 1995 explained: “We do not see a change in your medical condition which necessitated you to stop work. The records do not show a level of impairment which would restrict or limit you from performing the duties of your regular job given that you have worked with these conditions in the past.”
Section 502(a)(1)(B) of ERISA, 29 U.S.C. § 1132(a)(1)(B), makes decisions of this kind reviewable in federal court, and Perl-man asked the district judge to direct unum to pay both short- and long-term disability benefits. The court concluded that it was authorized to consider both short- and long-term claims, even though Perlman had sought only short-term benefits from unum, because an award of short-term benefits is a condition to receipt of long-term benefits; the parties disagree about whether Perlman is disabled, not the duration of any disability.. 979 F.Supp. 726, 731 n. 6 (N.D.Ill.1997).
After stating that unum’s decision was “arbitrary and capricious” because it failed to obtain the assistance of any outside experts, and did not perform a detailed, study of Perlman’s job duties, the judge directed unum to reconsider Perlman’s application in light of the analysis in its opinion. Both sides have appealed—unum because it believes that its decision should have been sustained, Perlman because she believes that the court should have ordered unum to pay benefits without giving it an opportunity to compile a better record. A second set of cross-appeals concerns attorneys’ fees. The district court held that Perlman is the prevailing- party and ordered unum to pay $44,020 under 29 U.S.C. § 1132(g)(1). 990 F.Supp. 1039 (N.D.I11.1998). unum now contends that an award of fees is impermissible until after the court has directed the erisa plan to pay benefits; Perlman contends that $44,-000 is too low (it is only a third of her request).
Appellate jurisdiction is the first, and potentially the last, issue we need to consider. The district court entered a judgment that reads in full: “IT IS ORDERED AND ADJUDGED that both parties’s motions for summary judgment are denied. The case is remanded to unum for a new determination in accordance with the Memorandum Opinion and Order, and judgment is, therefore, entered accordingly.” “IT IS ORDERED AND ADJUDGED ... [that] judgment is ... entered” is not helpful. A judgment must be self-contained (which this is not) and specify the relief to which the prevailing party is entitled (which this does not). See Massey Ferguson Division of Varity Corp. v. Gurley, 51 F.3d 102 (7th Cir.1995); American Interinsurance Exchange v. Occidental Fire & Casualty Co., 835 F.2d 157 (7th Cir.1987). What is more, it is impossible to see how a “case” (as opposed to a “claim for benefits”) may be returned to unum. Litigants should flag such shortcomings *978for the district judge, so that they can be corrected and jurisdictional issues avoided. Like the parties to other recent cases, these litigants ignored the jurisdictional problems rather than helping to avoid or cure them. See Continental Casualty Co. v. Anderson Excavating & Wrecking Co., 189 F.3d 512, 515-16 (7th Cir.1999); Health Cost Controls, Inc. v. Washington, 187 F.3d 703, 706 (7th Cir.1999).
Neither side contends that the judgment is too uncertain to be enforced, however, and it is easy enough to see from the opinion what the judge had in mind. Because the district court obviously does not plan to enter any further orders, here as in Continental Casualty and Health Cost Controls the formal defects do not prevent appellate review. Compare Otis v. Chicago, 29 F.3d 1159 (7th Cir.1994) (en banc), with Buck v. U.S. Digital Communications, Inc., 141 F.3d 710, 711 (7th Cir.1998). But the nature of the relief, a remand to unum, may do so. Although it is doubtful as an original matter that a district court may “remand” ERisa claims, as if to administrative agencies, we have held that courts may treat welfare benefit plans just like administrative law judges implementing the Social Security disability-benefits program. Quinn v. Blue Cross & Blue Shield Ass’n, 161 F.3d 472, 476-78 (7th Cir.1998); Schleibaum v. Kmart, 153 F.3d 496, 503 (7th Cir.1998). That makes it necessary to determine whether a remand is appealable as a final decision under 28 U.S.C. § 1291.
One court of appeals has answered “yes” without analysis. Snow v. Standard Insurance Co., 87 F.3d 327, 332 (9th Cir.1996). Two courts have answered “no,” analogizing the remand to a district court’s order setting the case for a new trial. Petralia v. AT&T Global Information Solutions Co., 114 F.3d 352 (1st Cir.1997); Shannon v. Jack Eckerd Corp., 55 F.3d 561 (11th Cir.1995). Petralia thought it implicit in a remand that the parties may return to the district court without filing a new complaint following the plan’s fresh decision, and if so then the district court’s decision cannot be called final. A more recent decision of the ninth circuit called Petralia’s assessment persuasive, Williamson v. unum Life Insurance Co., 160 F.3d 1247, 1252 (9th Cir.1998), though because the judgment in Williamson was not final by any standard (the district court had resolved a few disputed issues but had not awarded any relief) Snow was not overruled. A fourth court of appeals has noticed the conflict without taking a stand. Crocco v. Xerox Corp., 137 F.3d 105, 108 (2d Cir.1998). Now it is our turn.
If the justification for a remand to an ERisa plan’s administrator is that the plan makes the same kind of decisions as the Social Security Administration, then it is important to know the jurisdictional consequences of a remand in a Social Security case. Two sentences of 42 U.S.C. § 405(g) authorize remand in different situations. Sentence four authorizes the district court to enter “a judgment affirming, modifying, or reversing the decision of the Commissioner of Social Security, with or without remanding the cause for a rehearing.” Sentence six allows the court to remand for the receipt of new evidence, but without entering a judgment determining the propriety of the decision previously rendered. Sullivan v. Finkelstein, 496 U.S. 617, 110 S.Ct. 2658, 110 L.Ed.2d 563 (1990), holds that a sentence-four remand, which depends on a finding of error in the Commissioner’s decision, is appealable under § 1291 as a final decision, while a sentence-six remand is not final or appeal-able because no adjudication has taken place. See also Forney v. Apfel, 524 U.S. 266, 118 S.Ct. 1984, 141 L.Ed.2d 269 (1998). A sentence-four remand concludes the litigation in the district court; any protest about the Commissioner’s decision on remand requires a new suit. But a sentence-six remand works like a yo-yo; once the record has been enlarged, the district court finally decides whether the administrative decision is tenable.
*979Remands to plan administrators serve the same functions as remands to the Commissioner, which implies the same jurisdictional treatment for purposes of § 1291. If the district court finds that the decision was erroneous and enters a judgment wrapping up the litigation, that decision is appealable even if extra-judicial proceedings he ahead; but if the court postpones adjudication until after additional evidence has been analyzed, then it has not made a final decision. The Supreme Court drew this line in Finkelstein in part because it was concerned that it would otherwise be impossible for the Commissioner to obtain review of an adverse decision by a district court. If the court directs the Commissioner to apply a specific rule, or accept specific evidence, and the Commissioner carries out that directive and awards benefits, the case may never return to court, unum may say the same about the remand ordered here. Petralia assumed that the case was bound to come back to the district court, but as this example shows it need not do so. At all events, the Court added in Forney, an appeal under § 1291 does not depend on who won in the district court or whether the remand is likely to prevent renewed litigation. The rule is mechanical: all sentence-four remands are appealable. For-ney holds that the claimant is as entitled to appeal a remand as is the Commissioner. Forney contended that she was entitled to immediate victory without the need for a remand, and the Supreme Court held that Forney could present this claim to the court of appeals before the remand occurred. Perlman occupies Forney’s position, and unum occupies the Commissioner’s position in Finkelstein, for the district court’s remand is based on a finding of error (parallel to sentence four) rather than new evidence (sentence six). Just as in Finkelstein and Forney, the district court entered a Rule 58 judgment indicating that the decision to deny benefits was in error, and that the court is done with the case.
Before concluding that ERisa remands are just like Social Security remands, we must consider a second possibility: that an ERisa remand is most similar to a remand to an arbitrator. Like an arbitrator, the administrator of an erisa plan is a private dispute resolver. It therefore may be instructive to explore whether an order declining to enforce an award but directing additional arbitral proceedings is appeal-able.
Our court first considered that question in Shearson Loeb Rhoades, Inc. v. Much, 754 F.2d 773 (7th Cir.1985). Recognizing that the circuits then were divided on the question, we came down firmly on the side of no appellate jurisdiction. Contemporaneously with Much, another circuit adopted a contrary view, United Steelworkers v. Adbill Management Corp., 754 F.2d 138, 140 & n. 1 (3d Cir.1985) (equating a remand to an order requiring arbitration in the first place), but we reaffirmed Much in United Steelworkers v. Aurora Equipment Co., 830 F.2d 753 (7th Cir.1987). Much and Aurora Equipment give a common reason for their conclusion: that remands to arbitrators should be treated just like remands to administrative agencies, which at the time of Much and Aurora Equipment were widely thought to be non-final. Three years after Aurora Equipment, that premise was undermined by Finkelstein, and Forney has completed the process of making administrative remands generally appealable. So if we adhere to the rationale of Much and Aurora Equipment, the comparison of erisa remands to arbitration remands does not lead in a new direction; it leads right back to the administrative-law analogy, and thus (today) to appellate jurisdiction.
Neither Much nor Aurora Equipment has been cited since Finkelstein, but for a reason that does not reflect a change of heart about the classification of arbitral remands under § 1291. In 1988 Congress changed the rules for appeals from decisions concerning arbitration. Under 9 U.S.C. § 16(b) any order by a district *980court directing the parties to arbitrate is non-appealable, but § 16(a)(1)(E) provides that any order “modifying, correcting, or vacating an award” may be appealed immediately. Courts of appeals routinely assume, though without discussion, that an order vacating an arbitrator’s decision but remanding for additional arbitration is ap-pealable under § 16(a)(1)(E), rather than non-appealable under § 16(b). See, e.g., Aircraft Braking Systems Corp. v. Automobile Workers, 97 F.3d 155 (6th Cir.1996). Aurora Equipment supports the conclusion that a remand should not be treated like an original order to arbitrate, so this circuit, too, presumably will entertain jurisdiction under § 16(a)(1)(E) when the time comes. Thus whether we think of remands to arbitrators as equivalent to remands to agencies (the rationale of Much and Aurora Equipment) or as orders vacating awards, the remands would today be appealable, which offers further support for the conclusion that § 1291 permits appeal of a remand to an ERisa administrator.
Permitting appeals from remand orders does carry a cost—not only because it sets the stage for successive appeals if the decision on remand should be contested, but also because it causes the attorneys’ fees dispute to come to a head prematurely. How is one to tell whether Perlman is a “prevailing party” entitled to fees until we know the final outcome? It is tempting to treat the remand as non-final simply in order to postpone the fee question until the outcome is known.
Once again, however, ERisa remands are functionally identical to Social Security remands. Prevailing parties obtain fees in ERisa cases if the opponent’s position was not substantially justified; the Equal Access to Justice Act, 28 U.S.C. § 2412, applies the same standard to Social Security cases. See Bittner v. Sadoff & Rudoy Industries, 728 F.2d 820, 829-31 (7th Cir.1984); Continental Can Co. v. Chicago Truck Drivers Pension Fund, 921 F.2d 126 (7th Cir.1990). A series of decisions from the Supreme Court tells us that obtaining a sentence-four remand makes the claimant a “prevailing party” without regard to what happens on the remand. Shalala v. Schaefer, 509 U.S. 292, 300-02, 113 S.Ct. 2625, 125 L.Ed.2d 239 (1993); Melkonyan v. Sullivan, 501 U.S. 89, 111 S.Ct. 2157, 115 L.Ed.2d 78 (1991); Sullivan v. Hudson, 490 U.S. 877, 109 S.Ct. 2248, 104 L.Ed.2d 941 (1989). In Schaefer the Court linked the attorneys’ fees and appealability issues together, so that an appealable order triggers an entitlement to fees (and a need to make a prompt application for them). The Court could have held that a final decision on the disability claim is essential to prevailing-party status, which would have negated appealability, but it did not. Thus the fee issues do not justify treating the remand as a non-final order; it is the Supreme Court’s decision to deem remands final that creates perplexing fee problems, but their existence does not justify departing from the approach marked out in Finkelstein, Forney, Schaefer, Melkonyan, and Hudson. All four of the cross-appeals are within our jurisdiction.
Before turning to the merits, we must decide how searching judicial review is to be. Decisions of ERisa plan administrators presumptively receive de novo review, see Firestone Tire & Rubber Co. v. Bruch, 489 U.S. 101, 109 S.Ct. 948, 103 L.Ed.2d 80 (1989), but if the plan establishes discretionary authority then review will be deferential. Id. at 115, 109 S.Ct. 948. The district court concluded that language in Swiss Bank’s plan and unum’s policy gave unum enough discretion in administering benefits that deferential review is called for. 979 F.Supp. at 729, relying on Patterson v. Caterpillar, Inc., 70 F.3d 503, 505 (7th Cir.1995). But the court then put a thumb on the scale against unum, reasoning that its self-interest as an insurer of long-run benefits created at least a potential for bias in making decisions.
*981Although a plan administrator’s self-interest does not affect the standard of review, Firestone calls it a “factor” to “weigh” in evaluating the decision. 489 U.S. at 115, 109 S.Ct. 948. How much does it weigh? Judge Boudin recently concluded that it can’t weigh very much without exceeding the judicial capacity to tailor standards of review. Judges understand deferential and non-deferential review, but intermediate variations blur into one another without promoting understanding or consistent adjudication. Doe v. Travelers Insurance Co., 167 F.3d 53, 57 (1st Cir.1999). Panels have on occasion suggested that a conflict should lead to judicial review with “more bite,” see Chojnacki v. Georgia-Pacific Corp., 108 F.3d 810, 815 (7th Cir.1997); Van Boxel v. Journal Co. Employees’ Pension Trust, 836 F.2d 1048, 1052-53 (7th Cir.1987), but have never come up with an operational definition of “more bite” or a specification of the appropriate circumstances for mastication. Perhaps this shows only that review is easier to perform than to articulate. If the first-level decisionmaker has an interest in the outcome, this potential for bias is bound to affect the mind set of the reviewing court. Gallo v. Amoco Corp., 102 F.3d 918, 921 (7th Cir.1996). But it is unsound for the judiciary automatically to impute the plan administrator’s position to the person who decides on its behalf.
“When the administrator is a large corporation, the firm has a financial interest, but the award in any one case will have only a trivial effect on its operating results. Corporations act through agents, and these agents usually lack any stake in the outcome. Getting employees to act as if shareholders’ welfare were their own is a daunting challenge for any corporation. See Candice Prendergast, The Provision of Incentives in Firms, 37 J. Econ. Lit. 7 (1999). Nothing in our record suggests that unum has even tried. Large businesses such as Swiss Bank want to maintain a reputation for fair dealing with their employees. They offer fringe benefits such as disability plans to attract good workers, which they will be unable to do if promised benefits are not paid. We have no reason to think that unum’s benefits staff is any more “partial” against applicants than are federal judges when deciding income-tax cases.
A further problem impedes treating unum’s self-interest as a strike against its decision: it may not have any stake in the decision. Although it insures the payment of the long-term benefits, the record does not reveal the terms on which the insurance was written. For many large firms, health and disability insurance on their ' labor forces is retrospectively rated. This means that the employer agrees to reimburse the insurer for all outlays, plus a loading charge and administration fee. Insurance is valuable to workers, who know that even if the employer encounters financial difficulties their benefits will not cease; but as long as the employer continues in business, it bears the full economic cost of its fringe-benefit promises. Perlman has not asked us to probe the terms on which unum’s insurance was written, or to investigate the compensation and promotion opportunities of the benefits staff, just to determine whether to “weigh” an additional “factor” in the calculus of judicial review. Thus we have no reason to think that the actual decisionmakers at unum approached their task any differently than do the decisionmakers at the Social Security Administration, and ordinarily deferential review is the order of the day.
It follows from the conclusion that review of unum’s decision is deferential that the district court erred in permitting discovery into unum’s decision-making. There should not have been any inquiry into the thought processes of unum’s staff, the training of those who considered Perl-man’s claim, and in general who said what to whom within unum—all of which Perl-man was allowed to explore at length by depositions and interrogatories, and on some of which the district judge relied. Deferential review of an administrative de-*982cisión means review on the administrative record. We have allowed parties to take discovery and present new evidence in erisa cases subject to de novo judicial decisions, see Casey v. Uddeholm Corp., 32 F.3d 1094, 1098-99 & n. 4 (7th Cir.1994); accord, Kinstler v. First Reliance Standard Life Insurance Co., 181 F.3d 243 (2d Cir.1999), but never where the question is whether a decision is supported by substantial evidence, or is arbitrary and capricious. Six courts of appeals have held that when review under erisa is deferential, courts are limited to the information submitted to the plan’s administrator. Wilkins v. Baptist Healthcare System, Inc., 150 F.3d 609, 617-20 (6th Cir.1998); De-Felice v. American International Life Assurance Co., 112 F.3d 61, 65 (2d Cir.1997); Donatelli v. Home Insurance Co., 992 F.2d 763, 765 (8th Cir.1993); Quesinberry v. Life Insurance Co. of North America, 987 F.2d 1017, 1021-27 (4th Cir.1993) (en banc); Sandoval v. Aetna Life & Casualty Insurance Co., 967 F.2d 377, 380 (10th Cir.1992); Luby v. Teamsters Health, Welfare, and Pension Trust Funds, 944 F.2d 1176, 1184-85 (3d Cir.1991). One court has gone the other way, Wildbur v. ARCO Chemical Co., 974 F.2d 631, 636-42 (5th Cir.1992), but we believe that the majority has this right. Perhaps the disagreement is more apparent than real. Courts in the majority have allowed, what Wildbur stressed, that discovery may be appropriate to investigate a claim that the plan’s administrator did not do what it said it did—that, for example, the application was thrown in the trash rather than evaluated on the merits. But when there can be no doubt that the application was given a genuine evaluation, judicial review is limited to the evidence that was submitted in support of the application for benefits, and the mental processes of the plan’s administrator are not legitimate grounds of inquiry any more than they would be if the decisionmaker were an administrative agency.
At last we reach the question that occupied central ground in the district court: did unum make an arbitrary or capricious decision? Both Perlman and the district court treat the central issue as whether unum correctly understood her abilities in relation to the demands of her job. If Swiss Bank had told Perlman that her performance was unsatisfactory, then these would indeed be the right questions; unum would have needed to determine whether the shortcomings were caused by medical conditions (and thus established disability) or had a non-medical source. But Swiss Bank did not fire Perlman, transfer her to a less demanding and less interesting position (which might have implied constructive discharge), or even suggest that her performance needed improvement. Perlman’s successful performance of the job between 1992 and 1994 shows that the 1988 accident did not prevent her from doing an attorney’s work. And Swiss Bank’s disability plan differs from Social Security to the extent that Swiss Bank does not have any version of “listed impairments”—conditions so serious that a finding of disability follows without any effort to determine whether the applicant is employable. Cf. Cleveland v. Policy Management Systems Corp., — U.S. -, 119 S.Ct. 1597, 143 L.Ed.2d 966 (1999). A person who can do the job at Swiss Bank is not “disabled” as the program uses that term. This meant that unless Perlman’s condition (physical or mental) deteriorated during 1993 and 1994, she is not entitled to disability benefits.
None of the medical evidence Perlman submitted to unum showed that she had taken a turn for the worse in recent years. Her application contained a physician’s description of her limitations and conditions, but not any effort to evaluate these in relation to her job or to assess how they had changed over time (if they have changed). In response to unum’s request for additional information, Perlman submitted about 800 pages of her personal *983diary and her physician’s notes. None of these materials offers medical support for a change of condition. During the appeals process, in the district court, and on this appeal Perlman has stood pat on the proposition that unum must assess her abilities in relation to the job without regard to her (apparent) ability to perform to Swiss Bank’s satisfaction. She did not provide UNUM (or for that matter the district court) with any evidence that her condition has worsened, unum’s insistence that she demonstrate a change in condition cannot be called arbitrary, capricious, or otherwise unreasonable.
We can imagine circumstances under which it would be unreasonable to demand proof of a change in condition. Some disabled people manage to work for months, if not years, only as a result of superhuman effort, which cannot be sustained. Sometimes work beyond one’s limitations is essential to maintain one’s standard of living, or is the result of an effort to disguise one’s limitations from friends and coworkers. Reality eventually prevails, however, and limitations that have been present all along overtake even the most determined effort to keep working. We can imagine a lawyer trying to carry on, to defy her limitations, in order to avoid abandoning the profession in which so much of her human capital and spirit is invested. But Perlman herself was willing to take long periods off from work between 1988 and 1992, and she did not argue to unum (or in this court) that her performance between 1992 and 1994 reflects a dogged effort that could not be continued. If Perlman were devoted to the workforce beyond her abilities, then likely she would have asked Swiss Bank for an accommodation under the Americans with Disabilities Act, but she did not do this. A responsible administrator of Swiss Bank’s disability plan therefore was entitled to deny Perlman’s claim, without obtaining additional medical evaluations, for the simple reason that Perlman has never argued that she was less able to do the job at the end of 1994 (or in 1995) than in the middle of 1993.
Because the district judge should have rejected Perlman’s challenge to unum’s decision, Perlman is not entitled to attorneys’ fees. None of the parties’ other arguments requires discussion in light of our conclusion that unum’s decision is reasonable, on the evidence that Perlman placed before it. The judgment is vacated, and the case is remanded with directions to enter judgment in defendants’ favor.