Full opinion text
Opinion by Judge BYBEE; Partial Concurrence and Partial Dissent by Judge ARNOLD; Partial Concurrence and Partial Dissent by Judge RAWLINSON. OPINION BYBEE, Circuit Judge, with whom ARNOLD, Circuit Judge, joins as to Parts I, II, IV.C, IV.D, V.A, and V.C, and with whom RAWLINSON, Circuit Judge, joins except as to Part V.C.2: As the district court aptly put it, these cases arise from the “continuing war over protection of the delta smelt.” San Luis & Delta-Mendota Water Auth. v. Salazar, 760 F.Supp.2d 855, 863 (E.D.Cal.2010). We are joined to the fray. The district court invalidated a biological opinion by the Fish and Wildlife Service that concluded that the Central Valley and State Water Projects jeopardize the continued existence of a three-inch fish and its habitat. We reverse in part and affirm in part. The Central Valley Project and the State Water Project, operated respectively by the Bureau of Reclamation (Reclamation) and the State of California, are perhaps the two largest and most important water projects in the United States. These combined projects supply water originating in northern California to more than 20,000,000 agricultural and domestic consumers in central and southern California. The source of this water, the estuary at the confluence of the San Francisco Bay and Sacramento-San Joaquin Delta (Bay-Delta), is also the lone habitat for the delta smelt, a threatened species under the Endangered Species Act. 16 U.S.C. § 1531 et seq. In 2008, Reclamation requested a biological opinion (BiOp) from the U.S. Fish and Wildlife Service (FWS), in accord with the Endangered Species Act (ESA), on whether its continued operations would jeopardize the smelt. In a more than 400-page opinion — described by the FWS as the most complex biological opinion ever prepared — the FWS concluded that the Central Valley operations would threaten the delta smelt and, as required by the Endangered Species Act, proposed “reasonable and prudent alternatives” that Reclamation should take to ameliorate the effect on the smelt. The alternatives recommended by the FWS would reduce the water exported from northern California to southern California through the Central Valley and State Water Projects. Reclamation has notified the FWS that it intends to operate the Projects in compliance with the biological opinion. The plaintiffs-appellees — various water districts, water contractors, and agricultural consumers — brought suit under the Administrative Procedure Act against various federal defendants, including Reclamation, the FWS, and the Secretary of the Interior, to prevent the federal defendants from implementing the biological opinion and its proposed alternatives. The district court, in a lengthy and comprehensive opinion, was deeply critical of the biological opinion and concluded that it was arbitrary and capricious. The court accused the FWS of repeatedly “ignoring [the] best science available” to reach a “results-driven choice.” 760 F.Supp.2d at 940, and “show[ing] no inclination' to fully and honestly address water supply needs beyond the species,” even as it “interdicts] the water supply for domestic human consumption and agricultural use for over twenty million people who depend on the Projects for their water supply,” id. at 956-57 (quoting the FWS). We are acutely aware of the consequences of this proceeding. As a court, however, we are limited in our review of matters within the expertise of an agency. We may review the FWS’s biological opinion and Reclamation’s implementation for arbitrariness, caprice, or actions otherwise not in accordance with law. 5 U.S.C. § 706(2)(A). Although the FWS must employ “the best scientific and commercial data available,” 16 U.S.C. § 1536(a)(2), it is “not required to support its finding that a significant risk exists with anything approaching scientific certainty,” Indus. Union Dep’t v. Am. Petroleum Inst., 448 U.S. 607, 656, 100 S.Ct. 2844, 65 L.Ed.2d 1010 (1980) (plurality opinion). And, “[w]hen examining this kind of scientific determination ... a reviewing court must generally be at its most deferential.” Baltimore Gas & Elec. Co. v. NRDC, 462 U.S. 87,103,103 S.Ct. 2246, 76 L.Ed.2d 437 (1983). For the reasons explained below, we conclude that the district court failed to observe these standards and we reverse its judgment. We recognize the enormous practical implications of this decision. But the consequences were prescribed when Congress determined that “these species of fish, wildlife, and plants are of esthetic, ecological, educational, historical, recreational, and scientific value to the Nation and its people.” 16 U.S.C. § 1531(a)(3). As the Supreme Court observed in Tennessee Valley Authority v. Hill: “It may seem curious to some that the survival of a relatively small number of three-inch fish ... would require the permanent halting of a virtually completed dam,” but “the explicit provisions of the Endangered Species Act require precisely that result.” 437 U.S. 153, 172-73, 98 S.Ct. 2279,' 57 L.Ed.2d 117 (1978). Such species have been “afforded the highest of priorities,” by Congress, even if it means “the sacrifice of the anticipated benefits of the project and of many millions of dollars in public funds.” Id. at 174, 98 S.Ct. 2279 (footnote omitted). The law prohibits us from making “such fine utilitarian calculations” to balance the smelt’s interests against the interests of the citizens of California. Id. at 187, 98 S.Ct. 2279. Consequently, any other “[rjesolution of these fundamental policy questions” about the allocation of water resources in California “lies ... with Congress and the agencies to which Congress has delegated authority, as well as with state legislatures and, ultimately, the populace as a whole.” Baltimore Gas & Elec., 462 U.S. at 97,103 S.Ct. 2246. I. FACTS AND PROCEEDINGS BELOW A. Background 1. The Sacramento-San Joaquin Delta “The history of California water development and distribution is a story of supply and demand. California’s critical water problem is not a lack of water but uneven distribution of water resources.” United States v. State Water Res. Control Bd., 182 Cal.App.3d 82, 98, 227 Cal.Rptr. 161 (1986). California’s Central Valley comprises some of the most productive farmland in the world. Extending 450 miles from north to south, and averaging 100 miles wide east to west, the Central Valley includes two principal rivers: The Sacramento River begins in the northern part of the valley, runs south past Sacramento, and is fed by the Feather and American Rivers. The San Joaquin River begins in the Sierra Nevadas, northeast of Fresno, runs west and northwest into the Central Valley, and is fed by smaller rivers, including the Calaveras, Chowchilla, Fresno, Rings, Merced, Mokelumne, Stan-islaus, and Tuolumne Rivers. The two rivers converge in the Sacramento-San Joaquin Delta and form an estuary that joins Suisun Bay, San Francisco Bay, and the Pacific Ocean. Although over 70 percent of California’s water originates north of Sacramento, more than 70 percent of the state’s demand is in the south. The water from this region supplies irrigation for seven million acres of agriculture and more than twenty million people, nearly half of California’s residents. See Where Does California’s Water Come From?, Aquafornia, The California Water News Blog, Aug. 13, 2008, 9:29 a.m., http://www. aquafornia.com/index.php/where-does-californias-water-come-from/. See generally Dugan v. Rank, 372 U.S. 609, 612-13, 83 S.Ct. 999, 10 L.Ed.2d 15 (1963); United States v. Gerlach Live Stock Co., 339 U.S. 725, 728-29, 70 S.Ct. 955, 94 L.Ed. 1231 (1950); San Luis & Delta-Mendota Water Auth. v. United States, 672 F.3d 676, 681—83 (9th Cir.2012); Westlands Water Dist. v. United States, 337 F.3d 1092, 1095-96 (9th Cir.2003); In re Bay-Delta Programmatic Envtl. Impact Report Coordinated Proceedings, 43 Cal.4th 1143, 77 Cal.Rptr.3d 578, 184 P.3d 709, 715-17 (2008); State Water Res. Control Bd., 182 Cal.App.3d at 97-100, 227 Cal.Rptr. 161. In an effort to manage the increasing and conflicting demands placed on the water flowing through the Sacramento-San Joaquin Delta region, California and the United States have embarked on two massive projects. First, in 1933, California proposed the Central Valley Project (CVP), a plan to transfer water from the Sacramento River to water-deficient areas in the San Joaquin Valley and from the San Joaquin River to the southern regions of the Central Valley. State Water Res. Control Bd., 182 Cal.App.3d at 98-100, 227 Cal.Rptr. 161. Reclamation took over the project in 1935, and it is now “the largest federal water management project in the United States.” Central Delta Water Agency v. United States, 306 F.3d 938, 943 (9th Cir.2002). The CVP consists of a series of dams, including Shasta, Folsom, and Nimbus Dams; 21 reservoirs; 11 hy-dropower plants; and 500 miles of canals and aqueducts. In re Bay-Delta, 184 P.3d at 716 n. 1. In 1992, Congress adopted the Central Valley Project Improvement Act (CVPIA), Pub.L. No. 102-575, 106 Stat. 4706, which Congress described as designed “to achieve a reasonable balance among competing demands for use of Central Valley Project water, including the requirements of fish and wildlife, agricultural, municipal and industrial and power contractors.” CVPIA, § 3402(f), 106 Stat. at 4706. In 1951, California approved what is known as the State Water Project (SWP), the largest state-built water project in the United States. San Luis & Delta-Mendota Water Auth., 672 F.3d at 683. Managed by the California Department of Water Resources (DWR), “[t]he SWP serves the domestic water needs of approximately two-thirds of all Californians,” principally in Southern California. In re Bay-Delta, 184 P.3d at 716. SWP consists of “21 dams and reservoirs, ... five power plants, 16 pumping plants, and 662 miles of aqueduct.” Id. at 716 n. 2. In 1994, eight state agencies and 10 federal agencies formed the CALFED Bay-Delta Program (CALFED) to address comprehensively the challenges of managing the Bay-Delta estuary. Id. at 717. The CVP and SWP each operate a major station for pumping water from the Bay-Delta to canals and aqueducts that will carry the water to the south. Both plants are located near Tracy, California, and together they reverse the natural flow of the southern part of the Bay-Delta through two distributaries of the San Joaquin, Old, and Middle Rivers, referred to as “OMR.” San Luis & Deltob-Mendota, 760 F.Supp.2d at 863. The CVP operates the Jones Pumping Plant, capable of diverting 4,600 cubic feet per second (cfs). Nearby, the SWP operates the Harvey O. Banks Pumping Plant, with a capacity of 10,300 cfs, although it generally operates at or below 6,680 cfs. BiOp at 82, 108, 159-60. The plants have been constructed with louvers that allow water to pass through into the pumping plant, but will prevent most fish from entering the plants. The process of the fish entering the plants, known as entrainment, traps some 52 different species of fish. BiOp at 67. The salvaged fish are hauled in trucks injected with oxygen and released at sites on the Sacramento and San Joaquin Rivers. BiOp at 67, 145. Over a recent 15-year period, more than 110 million fish were salvaged from the Jones and Banks facility. BiOp at 160. This number, however, greatly underestimates the number of fish actually entrained at the facilities, because fish less than 30 mm (1.2 inches) are not efficiently collected at the louvers. BiOp at 160-61. Smaller fish, especially those in the juvenile or larval stage, are killed in the pumps. BiOp at 210. Those that are salvaged frequently do not survive the salvage process. BiOp at 338. The Colorado River and the SWP have historically been the major supply of water for southern California. As the result .of an interstate. agreement signed in 2003, California will receive less water from the Colorado River. Quantification Settlement Agreement, San Diego Water Authority, http://www.sdcwa.org/quantification-settlement-agreement (last visited July 29, 2013). As a consequence, southern California has sought more water from SWP. BiOp at 89-90. Increased demand for water from SWP has had a predictable effect on the water flowing into the Bay-Delta. As more water is diverted from the rivers that feed the Delta into the southern Central Valley, the salinity of the Delta and its estuaries increases along with the threat to the species that thrive there. The Delta generally describes a large lowland area with a labyrinth of natural channels in and around the confluence of the Sacramento and San Joaquin Rivers.... The major factor affecting water quality in the Delta is saltwater intrusion. Delta Lands, situated at or below sea level, are constantly subject to ocean tidal action. Salt water entering from San Francisco Bay extends well into the Delta, and intrusion of the saline tidal waters is checked only by the natural barrier formed by fresh water flowing out from the Delta. State Water Resources Control Bd., 182 Cal.App.3d at 107, 227 Cal.Rptr. 161. Since the 1970s, Reclamation and a raft of state agencies have worked to mitigate the effects of increased water salinity on Sui-sun Bay resulting from the upstream diversion of water that would otherwise naturally flow through Suisun. BiOp at 112-13. Salinity levels in Suisun Bay are highly sensitive to diversion from the Delta. Two related standards are used to describe the salinity of the Bay-Delta. The first is the Low Salinity Zone or LSZ. BiOp at 147. The LSZ is the transition point between the freshwater of the inland rivers and brackish water flowing eastward from San Francisco Bay and the Pacific Ocean, and includes water ranging in salinity from 0.5 parts per thousand to six parts per thousand. BiOp at 191. The second is referred to as X2. X2 represents the point in the Bay-Delta at which the salinity is less than two parts per thousand. See Westlands Water Dist. v. Dep’t of Interior, 376 F.3d 853, 876 (9th Cir.2004). The LSZ, which encompasses a larger region of the Bay-Delta, is generally centered around X2. Together, these regions are largely determined by Bay-Delta outflow, which is the difference between the inflow and the water exported. The agencies use X2 as a marker for the LSZ as well as a habitat indicator for fish and as a regulatory standard. BiOp at 149-50, 236; San Luis & Delta-Mendota, 760 F.Supp.2d at 864 & nn. 4-5. They express the location of X2 as its distance in kilometers east of the Golden Gate Bridge, Westlands Water Dist., 376 F.3d at 876. 2. The delta smelt The delta smelt (Hypomesus transpaci-ficus) is a small, two-to-three inch species of fish endemic to the San Francisco Bay/Sacramento-San Joaquin Delta Estuary. BiOp at 140-41. Once an abundant species in the Bay-Delta ecosystem, the delta smelt is now in imminent danger of extinction. In March 1993, the species was listed as threatened under the ESA, and the FWS designated the BayDelta system a critical habitat for the delta smelt in 1994. 50 C.F.R. § 17.11; BiOp at 140. Yet, over the past decade, the delta smelt population has been decimated even relative to these depleted levels, with a measured decline since 2000 of up to three orders of magnitude below historic lows. San Luis & Delta-Mendota, 760 F.Supp.2d at 866. As a consequence, the FWS announced in 2010 that reclassifying the delta smelt from a threatened to an endangered species was warranted but precluded by higher priority listings. Id. The ESA provides “both substantive and procedural provisions designed to protect endangered species and their habitat.” Am. Rivers v. Nat’l Marine Fisheries Serv., 126 F.3d 1118, 1121 (9th Cir.1997). One such protection, § 7(a)(2) of the ESA, requires federal agencies to “insure that any action authorized, funded, or carried out by such agency ... is not likely to jeopardize the continued existence of any endangered species or threatened species or result in the destruction or adverse modification of habitat of such species.” 16 U.S.C. § 1536(a)(2). Should the agency find that its proposed action may affect a listed species or critical habitat, it must formally or informally consult with the Secretary of the Interior, or his or her delegee. 16 U.S.C. § 1536(a)(4); 50 C.F.R. § 402.14(a); see Am. Rivers, 126 F.3d at 1122. If no effect is found, consultation is not required. 50 C.F.R. § 402.14. Formal consultation is required when the acting agency or consulting agency determines that the proposed action is likely to adversely affect a listed species or critical habitat. 50 C.F.R. §§ 402.13, 402.14. Formal consultation requires the consulting agency, here the FWS, to issue a biological opinion stating whether the proposed action is likely to jeopardize such species or habitat. 16 U.S.C. § 1536(b); 50 C.F.R. § 402.14. Should the action jeopardize the species or habitat, the consulting agency must suggest any “reasonable and prudent alternatives” (RPA) that would allow the projects to continue operation without causing jeopardy to the species'or adverse modification to its critical habitat. 16 U.S.C. § 1536(b)(3)(A). Once it receives the "BiOp, the acting agency “shall determine whether and in what manner to proceed with the action in light of its section 7 obligations and the [FWS’s] biological opinion.” 50 C.F.R. § 402.15(a). If, after consultation, thé agency determines that it cannot comply with § 7(a)(2), it may apply for an exemption, which can only be authorized by the Endangered Species Committee, an ad hoc panel composed of executive branch members and at least one appointee from the state in which the project is to occur. ■ 16 U.S.C. § 1536(e); 50 C.F.R. §§ 402.15(c), 451. B. Proceedings Leading To The Present Controversy 1. The FWS’s 2008 Biological Opinion Reclamation sought a biological opinion from the FWS as part of its continued long-term' operation of the CVP and its coordinated operations with state agencies of the SWP. BiOp at 1, 8. Following § 7(a)(2) review and a subsequent formal consultation, the FWS issued a biological opinion in 2005 (2005 BiOp). The 2005 BiOp found that the proposed coordinated operations of the CVP and SWP would not have an adverse effect on the continued existence and recovery of the delta smelt and its critical habitat. San'Luis & Delia-Mendota, 760 F.Süpp.2d at 863. The Natural Resources Defense Council — defendants-intervenors-appellants in the present case — challenged the FWS’s conclusion in U.S. District Court for the Eastern District of California, and the court found the 2005 BiOp arbitrary and capricious. Kempthome, 506 F.Supp.2d at 387. After conducting an extensive evidentiary hearing, the district court issued an interim remedial order and findings of fact and conclusions of law, which covered, among other things, the effects on delta smelt of negative flows in OMR. See San Luis & Delta-Mendota, 760 F.Supp.2d at 864. The district court ordered Reclamation and DWR to implement a winter “pulse flow” in OMR of no more negative than - 2,000 cfs, and to “operate the CVP and SWP to achieve a daily average net upstream (reverse) flow in the OMR not to exceed -5,000 cfs on a seven-day running average during a defined period in the spring.” Id.; Int. Rem. Order at 5-7. The district court also ordered the FWS to complete a new BiOp in just nine months, a deadline that it would ultimately extend to one year. San Luis & Delta-Mendota, 760 F.Supp.2d at 865; Int. Rem. Order at 2. The FWS issued a new delta smelt biological opinion on the deadline, December 15, 2008 (2008 BiOp or BiOp). BiOp at 1-396. In stark contrast to the 2005 BiOp, the 2008 BiOp concluded that the “coordinated operations of the CVP and SWP, as proposed, are likely to jeopardize the continued existence of the delta smelt” and “adversely modify delta smelt critical habitat.” BiOp at 276-78. With respect to the delta smelt, the FWS entered five findings of fact: (1) “Diversions of water from the Delta have increased since 1967 when the SWP began operation in conjunction with the CVP.” BiOp at 276. The CVP/SWP operations have entrained smelt, including adults, larvae, and juveniles, at the Banks and Jones facilities; reduced smelt habitat; and reduced the Delta outflows, altering the location of the LSZ; (2) “The delta smelt is currently at its lowest level of abundance since monitoring began in 1967”; (3) “Under the proposed CVP/SWP operations, inflows to the Delta are likely to be further reduced, as water demands upstream of the Delta increase, most notably on the American River.” BiOp at 276. This is likely to “cause increased relative entrainment of adult delta smelt in the winter and spring, and of larval and juvenile delta smelt in the spring”; (4) “Other baseline stressors will continue to adversely affect the delta smelt, such as contaminants, microcystis, aquatic macrophytes, and invasive species”; (5) To recover, delta smelt will need a more abundant adult population, an increase in the quality and quantity of spawning, rearing, and migratory habitat, a reduction in contaminants and pollutants, a reduction in exposure to disease and toxic algal blooms, and a reduction in entrainment at water-diversion facilities in the Bay-Delta. BiOp at 276-77. With respect to delta smelt critical habitat, the FWS found that “past and present operations of the CVP/SWP have degraded these habitat elements” such that they are “insufficient to support successful delta smelt recruitment at levels that will provide for the species’ conservation.” BiOp at 278. The FWS recommended five components and listed six separate actions as “reasonable and prudent alternatives” (RPA): RPA Component 1 (Actions 1 and 2). Component 1 protects the adult delta smelt life stage by controlling OMR flows during the vulnerable December to March period. It has two proposed actions. Action 1 is “designed to protect upmigrating delta smelt” and describes the two periods when delta smelt are most vulnerable to entrainment: in December and when the first flush appears. BiOp at 280-81. Action 1 therefore proposes limiting the negative flows at OMR based on a “daily salvage index.” Id. In effect, this means that when the “daily salvage index” reaches a critical point (“the risk threshold”), the Projects have to reduce their diversion for 14 days. During that period, OMR flows can be “no more negative than -2,000 cfs” for a 14-day running average and “no more negative than -2,500 cfs” for a 5-day running average. BiOp at 281, 329. Action 2 follows from Action 1 but covers the period from December through March, when- pre-spawning adult delta smelt are vulnerable to entrainment. BiOp at 352. During that period, OMR flows can be no more negative than -5,000 cfs, although the FWS expected that flows would generally be in the range of -2,000 cfs to -3,500 cfs. RPA Component 2 (Action 3). Component 2 protects larval and juvenile delta smelt by limiting OMR flows following the completion of Component 1 when the Bay-Delta water temperatures reach 12°C, or when a spent female smelt is detected in trawls at Jones or Banks or is found in the salvage facilities. Action 3 requires the CVP/SWP projects to maintain their average OMR flows between -1,250 and -5,000 cfs until June 30, or until the mean water temperature reaches a target level, whichever occurs earlier. BiOp at 290, 357-58. RPA Component 3 (Action 4). Component 3 improves smelt habitat by increasing Bay-Delta outflow during the fall. Action 4 requires that in September and October, in years when the precipitation and runoff is defined as “wet or above normal,” Reclamation and DWR must provide sufficient Delta outflow to maintain X2 no more eastward than 74 km from the Golden Gate in wet years and 81 km in above-normal years. BiOp at 282-83, 369. RPA Component k (Action 6 ). Component 4 restores habitat by establishing a program to create or restore intertidal and associated subtidal habitat to the Bay-Delta and Suisun Marsh. Action 6 requires DWR to create or restore at least 8,000 acres in the Delta and Suisun Marsh. BiOp at 283. RPA Component 5. Component 5 monitors and reports on the implementation, success, and possible improvements of Components 1-4. Finally, the FWS issued an “incidental take statement” (ITS) in accord with 50 C.F.R. § 402.02. For purposes of the ITS, the FWS presumed that its reasonable and prudent alternatives would be implemented. Based on that premise, the FWS found that, as a result of CVP/SWP operations, there would be a take of the delta smelt, and that although the- extent of the take would be difficult to estimate, smelt entrainment would be minimized when OMR flows were regulated according to the FWS’s proposed RPA. BiOp at 285-86. As a consequence, the FWS concluded that “this level of anticipated take is not likely to result in jeopardy to the species or destruction or adverse modification of critical habitat when the RPA is implemented.” BiOp at 293. 2. The present case The first of six complaints challenging the FWS’s 2008 BiOp was filed in March 2009. San Luis & Delta-Mendota, 760 F.Supp.2d at 865. “Plaintiffs moved for a preliminary injunction ... to prevent Reclamation from implementing Component 2 of the RPA, alleging that FWS violated the National Environmental Policy Act (“NEPA”), and the ESA.” Id. The district court granted, the motion in part, finding that plaintiffs-appellees were likely to succeed on the merits of their NEPA claim, and requiring the FWS to make specific written findings to justify weekly decisions regarding OMR flow restrictions. Id. Plaintiffs-appellees sought a preliminary injunction against the implementation of RPA Component 3. Id. Following an evi-dentiary hearing, the district court issued a preliminary injunction confirming that plaintiffs-appellees had succeeded on their NEPA claims and finding that plaintiffs-appellees were likely to succeed on the merits of their ESA claim. Id. In December 2010, the district court entered final judgment on the primary claims in a 115-page opinion. Id. at 967-70. Although the FWS’s 2008 BiOp reached antipodal conclusions to the 2005 BiOp — which reached a no jeopardy conclusion and was found arbitrary and capricious — the district court once again found the 2008 BiOp to be arbitrary and capricious under the ESA and the APA and remanded the BiOp, its RPA, and Reclamation’s provisional acceptance of the RPA to the agency. San Luis & Delta-Mendo-ta, 760 F.Supp.2d at 855, 970. The court’s remand required the completion of yet a third BiOp analyzing the impact of CVP and SWP operations on the delta smelt. San Luis & Delta-Mendota, 760 F.Supp.2d at 870. In March 2011, the district court entered final judgment on all remaining claims. Although the district court accepted the BiOp’s central conclusion that “Project operations are likely to jeopardize the continued existence and/or adversely modify the critical habitat of the delta smelt,” id. at 969, the district court determined that there were a number of specific flaws with the BiOp, id. at 967-70. We will briefly set forth the district court’s principal objections here — which are highly technical and somewhat obtuse out of context — and explain them in more detail in the discussion section. First, the district court found the BiOp’s reliance on analyses using raw salvage figures — e.g., those calculations that incorporated the absolute “raw” number of smelt entrained in pumping stations, as opposed to the smelt entrained as a percentage of the total population — to be arbitrary and capricious and not the result of the best available science. Id. at 968. These calculations significantly influenced the upper and lower OMR flow limits in Actions 1, 2, and 3. Second, the district court found that the BiOp’s use of two different models, CAL-SIM II and DAYFLOW, that predict the location of X2, introduced bias requiring a corrective calibration or, at the very least, explanation. Id. The district court also found that the bias produced by the comparison of CALSIM II to DAYFLOW tainted the BiOp’s justification for Action 4, which involves the management of X2. In addition, the district court found that the BiOp did not sufficiently explain why it is essential in Component 3 (Action 4) to maintain X2 at the specific locations of 74 km upstream from the Golden Gate Bridge following “wet years” and 81 km following “above-normal years.” Id. at 969. Third, the district court found that the BiOp did not sufficiently explain why different data sets were used to calculate the incidental take limit for juvenile and for adult smelt, and why these limits were calculated using an average of previous years’ smelt salvage (which would be expected to be exceeded in 50% of all future years). Fourth, the district court found the BiOp did not adequately support its conclusions that Project operations are reasonably certain to indirectly affect the delta smelt by limiting delta smelt food supply, by increasing harmful pollution and contaminants, and by increasing the detrimental impact of the “other stres-sors” of predation, macrophytes, and mi-crocystis on delta smelt. Id. Fifth, the district court held that the BiOp failed to analyze economic feasibility, consistency with the purpose of the action, and consistency with the action agency’s authority, as required by § 402.02. See 5 U.S.C. § 551 et seq; 50 C.F.R. § 402.02. Id. at 969-70. Federal Defendants and the NRDC have timely appealed, urging reversal of the district court’s remand. We address their claims in Part IV. The San Luis & Delta-Mendota Water Authority and other appellees timely cross-appealed, arguing that the district court did not go far enough. They raise three claims: First, that the FWS violated the ESA by not separating out nondiscre-tionary actions from discretionary actions in setting an environmental baseline; second, that Reclamation acted arbitrarily and capriciously in adopting the flawed BiOp; and, third, that the FWS failed to conduct the review required by NEPA. We address their claims in Part V. This tortured procedural history has extended over seven years, and has led to five fully consolidated suits and one partially consolidated suit brought by various groups who use water supplied by the CVP and SWP, as well as to the completion of two extensively researched BiOps— with a third currently in progress. All the while, the delta smelt has teetered on the brink of extinction. We agree with Federal Defendants and the NRDC and reverse the district court’s remand of the BiOp. We affirm the district court’s judgment that Reclamation failed to comply with NEPA before implementing the FWS’s BiOp. II. STANDARDS OF REVIEW We review the district court’s ruling on summary judgment de novo. McFarland v. Kempthome, 545 F.3d 1106, 1110 (9th Cir.2008). Neither the ESA nor NEPA supply a separate standard for our review, so we review claims under these Acts under the- standards of the APA. Bennett v. Spear, 520 U.S. 154, 174, 117 S.Ct. 1154, 137 L.Ed.2d 281 (1997); Oregon Natural Desert Ass’n v. Bureau of Land Mgmt., 625 F.3d 1092, 1109 (9th Cir.2010); Pyramid Lake Paiute Tribe of Indians v. U.S. Dept. of Navy, 898 F.2d 1410, 1414 (9th Cir.1990). Section 706(2) of the APA provides that an agency action must be upheld on review unless it is “arbitrary, capricious, an abuse of discretion, or otherwise not in accordance with law.” 5 U.S.C. § 706(2)(A). As a reviewing court, we “must consider whether the decision was based on a consideration of the relevant factors and whether there has been a clear error of judgment.” Citizens to Preserve Overton Park, Inc. v. Volpe, 401 U.S. 402, 416, 91 S.Ct. 814, 28 L.Ed.2d 136 (1971), abrogated in part on other grounds as recognized in Califano v. Sanders, 430 U.S. 99, 105, 97 S.Ct. 980, 51 L.Ed.2d 192 (1977). Although our inquiry must be thorough, the standard of review is highly deferential; the agency’s decision is “entitled to a presumption of regularity,” and we may not substitute our judgment for that of the agency. Id. at 415-16, 91 S.Ct. 814. Where the agency has relied on “relevant evidence [such that] a reasonable mind might accept as adequate to support a conclusion,” its decision is supported by “substantial evidence.” Bear Lake Watch, Inc. v. FERC, 324 F.3d 1071, 1076 (9th Cir.2003). Even “[i]f the evidence is susceptible of more than one rational interpretation, [the court] must uphold [the agency’s] findings.” Id. Under the ESA, the agency must base its actions on evidence supported by “the best scientific and commercial data available.” 50 C.F.R. § 402.14(g)(8); 16 U.S.C. § 1536(a)(2). The determination of what constitutes the “best scientific data available” belongs to the agency’s “special expertise.... When examining this kind of scientific determination, as opposed to simple findings of fact, a reviewing court must generally be at its most deferential.” Baltimore Gas & Elec. Co., 462 U.S. at 103, 103 S.Ct. 2246. “Absent superior data[,] occasional imperfections do not violate” the ESA best available standard. Kern Cnty. Farm Bureau v. Allen, 450 F.3d 1072, 1080-81 (9th Cir.2006). “The best available data requirement ‘merely prohibits [an agency] from disregarding available scientific evidence that is in some way better than the evidence [it] relies on.’ ” Kern Cnty. Farm Bureau, 450 F.3d at 1080 (emphasis added) (quoting Sw. Ctr. for Biological Diversity v. Babbitt, 215 F.3d 58, 60 (D.C.Cir. 2000)). “Essentially, FWS ‘cannot ignore available biological information.’ ” Id. at 1080-81 (quoting Conner v. Burford, 848 F.2d 1441, 1454 (9th Cir.1988)). Thus, “insufficient ... [or] incomplete information ... does not excuse [an agency’s] failure to comply with the statutory requirement of a comprehensive biological opinion using the best information available” where there was some additional superior information available. Conner, 848 F.2d at 1454-55. On the other hand, where the information is not readily available, we cannot insist on perfection: “[T]he ‘best scientific ... data available,’ ” does not mean “the best scientific data possible.” Building Indus. Ass’n v. Norton, 247 F.3d 1241, 1246 (D.C.Cir. 2001). III. THE SCOPE OF THE RECORD Our review is limited to “the administrative record already in existence, not some new record made initially in the reviewing court.” Camp v. Pitts, 411 U.S. 138, 142, 93 S.Ct. 1241, 36 L.Ed.2d 106 (1973). If the record before the agency does not support the agency action, if the agency has not considered all relevant factors, or if the reviewing court simply cannot evaluate the challenged agency action on the basis of the record before it, the proper course, except in rare circumstances, is to remand to the agency for additional investigation or explanation. The reviewing court is not generally empowered to conduct a de novo inquiry into the matter being reviewed and to reach its own conclusions based on such an inquiry.... The factfinding capacity of the district court is thus typically unnecessary to judicial review of agency decisionmak-ing. Florida Power & Light Co. v. Lorion, 470 U.S. 729, 744, 105 S.Ct. 1598, 84 L.Ed.2d 643 (1985). See also Citizens to Preserve Overton Park, 401 U.S. at 419-20, 91 S.Ct. 814. There is a danger when a reviewing court goes beyond the record before the agency. “When a reviewing court considers evidence that was not before the agency, it inevitably leads the reviewing court to substitute its judgment for that of the agency.” Asarco, Inc. v. EPA 616 F.2d 1153, 1160 (9th Cir.1980). See Sw. Ctr. for Biological Diversity v. U.S. Forest Serv., 100 F.3d 1443, 1450 (9th Cir.1996) (“Judicial review of an agency decision typically focuses on the administrative record in existence at the time of the decision and does not encompass any part of the record that is made initially in the reviewing court.”). Accordingly, we do not review “the evidence to determine the correctness or wisdom of the agency’s decision ... even if the court has also examined the administrative record.” Asar- co, 616 F.2d at 1160. If the reviewing court cannot find substantial evidence in the record, it should “not compensate for the agency’s dereliction by undertaking its own inquiry into the merits,” id., but should remand to the agency for further proceedings, see INS v. Ventura, 537 U.S. 12, 16, 123 S.Ct. 353, 154 L.Ed.2d 272 (2002). “We have, however, crafted narrow exceptions to this general rule.” Lands Council v. Powell, 395 F.3d 1019,1030 (9th Cir.2005). As we have explained, [w]e allow expansion of the administrative record in four narrowly construed circumstances: (1) supplementation is necessary to determine if the agency has considered all factors and explained its decision; (2) the agency relied on documents not in the' record; (3) supplementation is needed to explain technical terms or complex subjects; or (4) plaintiffs have shown bad faith on the part of the agency. Fence Creek Cattle Co. v. U.S. Forest Serv., 602 F.3d 1125, 1131 (9th Cir.2010). Keeping in mind the Supreme Court’s concerns with reviewing court factfinding, we have approached these exceptions with caution, lest “the exception ... undermine the general rule.” Lands Council, 395 F.3d at 1030. We have serious concerns that the district court failed to observe these rules. First, the district court appointed four experts to aid it in understanding the technical and scientific aspects of the BiOp. These experts were appointed consistent with Federal Rule of Evidence 706. See supra Note 13. Having read the BiOp, we are sympathetic to the district court’s need for a scientific interpreter. No party has objected on appeal to the district court’s appointments, and we can see no reasonable objection to the use of experts to explain the highly technical material in the BiOp. The district court, however, did not limit itself to the court-appointed experts. Over the vigorous objection of the appellants, the court admitted multiple declarations from multiple experts hired by the appellees, even though the party-appointed experts addressed many of the same issues being addressed by the court-appointed experts. By the government’s count — which we have not sought to verify, although we have examined the declarations in the record — the district court admitted more than forty expert declarations from the appellees; once the court denied the appellant’s motion to exclude the declarations, the appellants submitted their own experts’ declarations. Yet, we cannot see what the parties’ experts added that the court-appointed experts could not have reasonably provided to the district court. See Animal Def. Council v. Hodel, 840 F.2d 1432, 1437 (9th Cir.1988) (plaintiff failed to show why the district court needed to go outside the administrative record). The effect of this was, predictably, to create a battle of the experts. Moreover, it gave the proceedings in the district court the appearance that the administrative record was open and that the proceedings were a forum for debating the merits of the BiOp. See Ariz. Cattle Growers’ Ass’n v. U.S. Fish & Wildlife, 273 F.3d 1229, 1245 (9th Cir.2001) (“Considering evidence outside the record would render the extraordinarily complex consultation process ... meaningless.”). Just as we will not allow the agency to supply post-hoc rationalizations for its actions, so “post-decision information ... may not be advanced as a new rationalization either for sustaining or attacking an agency’s decision.” Sw. Ctr. for Biological Diversity, 100 F.3d at 1450. As the Court cautioned, “[w]hen specialists express conflicting views, an agency must have discretion to rely on the reasonable opinions of its own qualified experts even if, as an original matter, a court might find contrary views more persuasive.” Marsh v. Ore. Natural Res. Council, 490 U.S. 360, 378, 109 S.Ct. 1851, 104 L.Ed.2d 377 (1989). Even a quick review of the district court’s opinion shows that the appearance of an open record was the reality. The district court relied extensively on the declarations of the parties’ experts-as-advocates as the basis for rejecting the BiOp. See, e.g., San Luis & Delta-Mendota, 760 F.Supp.2d at 877-79, 881-84, 889-90, 894-97, 903-07, 912, 922. In places, the district court pits the experts against each other and resolves their contrary positions as a matter of scientific fact. E.g., San Luis & Delta-Mendota, 760 F.Supp.2d at 884, 904-07. In effect, the district court opened the BiOp to a post-hoc notice-and-comment proceeding involving the parties’ experts, and then judged the BiOp against the comments received. The ESA consultation process is not a rulemaking proceeding, but a request.from one agency for the expertise of a second agency. Although we may review each of those proceedings under the APA, the agency’s obligations under each is slightly different, and we must account for that difference in our review. . Our opinion in Asarco v. EPA addressed the question of using experts in a technical case and is on point here. In that case, the district court had a technical matter before it — involving a 1,000 foot stack at a smelter — and permitted the parties to call their experts. The court held four days of hearings in which at least ten witnesses testified. We held that.some of the testimony, mainly from experts who provided “background material” on the smelter’s operations, was - appropriate; the rest, of it was not appropriate, even if it came from experts and “served some marginal purpose”: Most of the expert testimony, however, should not have been admitted or at least not have been considered for the purpose of judging the wisdom of the EPA’s stack-testing requirement. This technical testimony was plainly elicited for the purpose of determining the scientific merit of the EPA’s decision. Although the testimony may have served some marginal purpose in allowing the district court to evaluate the EPA’s court of inquiry, we can only conclude that the extent of the scientific inquiry undertaken at trial necessarily led the district court to substitute its judgement for that of the agency. 616 F.2d 1153,1160-61 (9th Cir.1980). The district court similarly overstepped its bounds here. Because we review the court’s judgment de novo, however, we can confine our own scope of review to the administrative record, plus that evidence that satisfies the standards we have set forth here. See Dow AgroSciences v. Nat’l Marine Fisheries Serv., 707 F.3d 462, 469 (4th Cir.2013) (confining review to the administrative record and disregarding an affidavit submitted to the district court). We will consider the BiOp and any other evidence in the record submitted by the parties that the FWS considered in making its decision. We will also consider the testimony of the four experts' the district court appointed pursuant to Rule 706, as to whom there is no objection. IV. MERITS OF THE BiOp Before we consider the challenges to the BiOp, we have some preliminary observations. First, the BiOp is a bit of a mess. And not just a little bit of a mess, but, at more than 400 pages, a big bit of a mess. And the FWS knew it. In November 2008, shortly before the FWS submitted its final draft, the California Department of Water Resources commented on a portion of the FWS’s draft. Under the title “The Document is Confusing and Disorganized,” the DWR advised that “the document is especially unrefined. The text is often out of logical order ..., many actions are fairly vague, and the document is unpolished.” Some portions, DWR wrote, were “largely unintelligible.” The FWS also submitted its draft to a peer review panel, coordinated by an outside engineering firm. One of the questions the FWS asked the panel was whether the BiOp was “organized in a manner that is clear, concise and complete (i.e., is it understandable)?” The panel pulled no punches and responded as follows: “The Panel’s response to this question is ‘no.’ ■ The version of the [Effects Analysis] provided to the Panel was a draft and had not been adequately edited for general organization, consistency across sections in how analyses were described and reported, and for redundancies.” The bottom line was that “most readers would have a difficult time.” Both the California DWR and the peer review panel commented on the FWS’s imposing schedule for issuing the BiOp. DWR observed that the “FWS was under an exceptionally difficult deadline to get this document in.” In' a comment letter sent from the DWR to Reclamation just two weeks before' the deadline, the DWR reflected on the pressures the FWS labored under: “Our concerns with the Draft [BiOp] are extensive but are correctable before -the December 15 court ordered [deadline] for its finalization. We are willing to work closely with the FWS to address these concerns quickly....” And the peer review panel complained that its own “review was conducted in a four-day period under a tight schedule.” See Ass’n of Pac. Fisheries v. EPA, .615 F.2d 794, 811 (9th Cir.1980) (“The Agency itself recognized that its -data collection was not as thorough as it otherwise would have been: ‘The time constraints imposed by the statutory deadlines precluded the Agency from conducting an exhaustive sampling program.’”). This challenging deadline was not the fault of the agency, but was set by the same district court that would later, hold that the FWS’s rushed BiOp was arbitrary and capricious. See San Luis & Delta-Mendota, 760 F.Supp.2d at 865 (ordering the FWS to produce a revised BiOp in just nine months — a deadline that would later be extended by three months to give the FWS just one year to produce a new BiOp). We concur in these assessments. The BiOp is a jumble of disjointed facts and analyses. It appears to be the result of exactly, what we would imagine happens when an agency is ordered to produce an important opinion on an extremely complicated and technical subject matter covering multiple federal and state agencies and affecting millions of acres of land and tens of millions of people. We expect that the document was patched together from prior documents, assembled quickly by individuals working independent of each other,- and not edited for readability, redundancies, and flow. It is a ponderous, chaotic document, overwhelming in size, and without the kinds of signposts and roadmaps that even trained, intelligent readers need in order to follow the agency’s reasoning. We wonder whether anyone was ultimately well-served by the imposition -of tight deadlines in a matter of such consequence. Deadlines become a substantive constraint on what an agency can reasonably do. In this case, the FWS not only had to write and compile the report — a substantial task in and of itself — but was under pressure to, among other things, produce a reliable population estimate of the delta smelt. Kempthome, 506 F.Supp.2d at 373 (faulting the 2005 BiOp for failing to produce such an estimate, and noting that “[t]he viability of Delta smelt has been under scrutiny for over ten years. No party has shown that producing a reliable population estimate is scientifically unfeasible”). Such scientific tasks may not be as well suited to deadlines as producing written copy; the final product will necessarily reflect the time allotted to the agency. The FWS is currently preparing its third BiOp, again under the orders of the district court and on the district court’s deadlines. Although we ultimately conclude that we can discern the agency’s reasoning and that the FWS’s 2008 BiOp is adequately supported by the record and not arbitrary and capricious, we also recognize that Reclamation has continuing responsibilities under CVP and SWP and that this is likely not the last BiOp that the FWS will issue with respect to the delta smelt, nor is this the last legal challenge that we will hear. Future analy-ses should be given the time and attention that these serious issues deserve. A. The 2008 BiOp’s Reliance on Raw Salvage Figures to Set the Upper and Lower OMR Flow Limits Was Not Arbitrary and Capricious Under normal pumping operations, enough water is pumped from the OMR that the river’s flow reverses. BiOp at 159-60. As vast quantities of water are pumped from the river, smelt and other fish are entrained in the pumps, where they are captured and counted in fish salvage facilities. Id. Not all smelt are salvaged, however, as juvenile smelt smaller than 20 mm (0.79 inches), smelt still in the larval stage, and some percentage of adult delta smelt, are killed by the pumps. BiOp at 338. The prospects for salvaged smelt are also grim, as smelt rarely survive the salvage process, id.; up to sixty percent of the smelt population is lost each year at the pumping plants, BiOp at 210. Unsurprisingly, the 2008 BiOp found that as the OMR flows became increasingly negative, the entrainment risk and accompanying population loss increased accordingly, thereby threatening the smelt population’s continuing viability. BiOp at 163. To mitigate this effect, the BiOp RPA imposed pumping limits expressed in terms of negative OMR flows, ranging from -1,250 cubic feet per second (cfs) to - 5,000 cfs. BiOp at 280. Which limit is applicable is determined by the location of the smelt population, water turbidity, and a range of other factors. BiOp at 280. In determining the OMR flow limits, the FWS relied, in part, on the number of delta smelt salvaged from the fish screening facilities. Although it acknowledged that the number of smelt salvaged only represents “a small percentage of the actual number entrained,” and “is not a good estimate of actual delta smelt mortality through entrainment,” BiOp at 338, the FWS relied on this information because “[d]ata on the salvage of delta smelt is typically used to provide an index of entrainment into the diversion pumps,” BiOp at 145. The OMR flow limit has a great practical significance, not merely to the delta smelt but to Californians, as it represents the ultimate limit on the amount of water available to sustain California’s millions of urban and agricultural users. Appellees challenge the -5,000 cfs upper limit on OMR flows, claiming that the FWS based its calculation of the flow limit solely on information contained in two figures that show a significant increase in the number of smelt salvaged at approximately -5,000 cfs: Figures B-13 and B-14. See BiOp at 347, 350. These figures, appellees assert, failed to justify the -5,000 cfs limit because the number of delta smelt salvaged (raw salvage) was not adjusted for the smelt’s estimated population for that year (normalized). Appellees’ objection to the BiOp’s use of raw salvage data, rather than normalized data, in calculating appropriate OMR flow limits is essentially an omitted variable bias argument; appellees assert that the number of smelt salvaged in a year is highly influenced by the total population of smelt that year, and that therefore the BiOp’s failure to account for this relation renders its calculated flow limits unreliable. Put another way, according to appellees, when there are more smelt in the BayDelta, more smelt are salvaged. Any apparent relationship between OMR flows and smelt salvage, therefore, may actually be a relationship between smelt population size and smelt salvage. Failing to account for this will skew, if not invalidate entirely, the analysis. The district court agreed, finding that “[t]he use of raw salvage data, as opposed to salvage data scaled to population size, is problematic because raw salvage figures do not account for the size (or relative size) of the smelt population.” San Luis & Delta-Mendota, 760 F.Supp.2d at 889. The district court further found that “the use of normalized salvage data rather than gross salvage data is the standard accepted scientific methodology.” Id. It concluded that the FWS’s use of raw salvage data in Figures B-13 and B-14 to determine restrictions on OMR flows was “scientifically unacceptable.” Id. at 891. Because the FWS failed to use the best available scientific data, its -5,000 cfs flow limit in Actions 1, 2 and 3 was arbitrary and capricious. Id.; see id. at 891-94. These objections, however, suffer from several problems. First, the FWS appropriately relied on Figures B-13 and B-14 to justify its -5,000 cfs flow limits. In Figures B-13 and B14, the FWS acknowledged the uncertainty inherent to modeling the relation between OMR flows and smelt and chose a conservative model, a choice that is within the FWS’s discretion to make. See Nw. Coal, for Alts, to Pesticides v. EPA 544 F.3d 1043, 1050 (9th Cir.2008). Second, the -5,000 cfs flow limit prescribed in the RPA was not solely determined from Figures B-13 and B-14, and therefore flaws in those figures do not necessarily doom the BiOp’s conclusions. Third, the BiOp’s OMR flow limits work in tandem with the incidental take statement (ITS), which accounts for population-level impacts. That the FWS could have done more in determining OMR flow limits is uncontro-verted. This, however, is not to say that the FWS acted arbitrarily and capriciously; we hold, contrary to the district court, that the FWS’s OMR flow limits are supported by substantial evidence. 1. The FWS’s choice of a more conservative model to calculate flow limits in Figures B-13 and B-14 was supported by substantial evidence Figure B-13 in the BiOp is a graph showing the relationship between salvaged adult delta smelt and the OMR flow, measured in cfs. The graph shows a positive correlation between salvaged smelt and the reverse flow of the river. That is, the greater the water pumped through the Jones and Banks pumping stations, the greater the count of smelt salvaged at those stations. The FWS noted that the graph is upward sloping and linear in the lower half of the curve. At about -5,000 cfs there is a “break” in the data, and for flows more negative than -5,000 cfs (meaning more water exported from the Bay-Delta), the upward slope increases at an increasing rate. The FWS sought to verify whether the break in the data was actual. It conducted additional analyses of the data to verify that there was not a natural break at any other point and that any error in the OMR flow rates or salvage could not have caused the break. BiOp at 349-51. The FWS concluded that with “flows more negative than -1683, salvage increased.” BiOp at 351. The FWS stated that “[a] major assumption of this analysis is that as the population of Delta smelt declined, the number of fish at risk of entrainment remained constant.” BiOp at 349. What B-13 did was compare actual salvage numbers with OMR flow. What the FWS did not do — and what the appellees and the district court claim the FWS should have done — was prepare an additional figure in which it compared “normalized” salvage numbers with OMR flows. Normalized salvage would be the measure of the salvaged smelt divided by the total population of smelt, effectively yielding figures showing what percentage of smelt each year were salvaged at the pumps. In fact, the FWS itself had stated that it could verify its conclusion “by normalizing the salvage data by the estimated population size based on the [Fall Midwater Trawl] data.” BiOp at 349. The peer review had similarly suggested normalizing the data: The Panel suggests that the use of predicted salvage of adult smelt should be normalized for population size.... One way to normalize salvage for population size is to divide by the previous fall Midwater Trawl (MWT) index. A similar regression model to the one fitted to salvage would relate the normalized salvage to Old and Middle River (OMR) flows.... Expressing salvage as a normalized index may help remove some of the confounding of the temporal trends during the baseline period.... As the district court’s experts acknowledged, the FWS faces significant practical challenges in setting OMR flow rates to minimize delta smelt entrainment. For example, day-to-day variations in OMR flows and “noise” in smelt sampling used to establish abundance and distribution of the delta smelt are significant confounding factors in determining appropriate OMR flow rates, as is the distribution of the delta smelt population in relation to the pumps. BiOp at 165, 331, 353-55. A lack of real-time information and variations inherent to environmental systems make precision virtually impossible. BiOp at 165, 331, 353-55. Yet, as even the district court recognized, population numbers of the delta smelt are perilously low, San Luis & Deltar-Mendota, 760 F.Supp.2d at 866, and entrainment by the pumping plants has a “sporadically significant influence on population dynamics,” id. at 877 (emphasis added). In such circumstances, the FWS’s decision to use raw salvage data rather than population-adjusted salvage data reasonably protects the delta smelt population without regard to year-by-year fluctuation in population size. The BiOp notes that this decision was motivated by a concern for the absolute number of smelt entrained in the pumps, not the relative number of smelt: “The current population cannot tolerate direct mortality through adult entrainment at levels approaching even ‘moderate’ take as observed through the historic record of recent decades.” BiOp at 287. Thus, the RPA is designed to “reduce entrainment of pre-spawning adult delta smelt during December to March” and to “Minimize the number of larval delta smelt entrained at the facilities” by controlling OMR flow from March to June. BiOp at 280, 357. The analytical approach preferred by appellees and the district court is best gauged to measure the number of smelt entrained at the pumps relative to the population size. This may be a more accurate reflection of the relative impact of OMR flows on the smelt population, but it is not tailored to protect the maximum absolute number of individual smelt, as the BiOp’s approach is; the process of adjusting raw salvage for the smelt population size results in normalized numbers, but it does so at the potential cost of minimizing the impact of each individual smelt lost. Our deference to agency determinations is at its greatest when that agency is choosing between various scientific models, as the FWS did in the present instance. See Nw. Coal, for Alts, to Pesticides, 544 F.3d at 1050. Facing great measurement uncertainty and a smelt population whose existence is threatened, the FWS chose to be conservative in setting the flow limits in Actions 1, 2, and 3. This choice was well within its discretion; the Supreme Court has held that an agency may choose to “counteract the uncertainties” inherent in its scientific analyses by “overestimat[ing]” known parameters without being unreasonable, Baltimore Gas & Elec. Co., 462 U.S. at 103, 103 S.Ct. 2246, and we have upheld an agency’s reliance on models that “yield conservative data because the models incorporate the higher of [known potential values] in assessing the overall risk,” Nw. Coal, for Alts, to Pesticides, 544 F.3d at 1050. Likewise, we give the FWS great deference in its choice of scientific tools, and, in these circumstances, hold that the FWS did not act arbitrarily or capriciously in choosing an analytical tool that resulted in greater protections for the imperiled smelt population. 2. The BiOp’s determination of OMR flow limits was influenced by more than Figures B-13 and B-14 The district court’s experts concluded that it was the BiOp’s apparent exclusive reliance on Figures B-13 and B-14 that was problematic: as they acknowledged, the figures could be useful in tandem with other analyses and data. The district court therefore based its invalidation of the -5,000 cfs OMR flow limit on its finding that the limit “depend[s] so heavily” on Figures B-13 and B-14. San Luis & Del-tar-Mendota, 760 F.Supp.2d at 968. After conducting an independent review of the record, however, we hold that the BiOp’s d