Enablement, Written Description, and Indefiniteness After Amgen v. Sanofi: What Section 112 Demands

The expert was trying to help. Seagen's patent claimed a cancer drug whose linker contained a four-amino-acid chain made only of glycine and phenylalanine. The patent was filed in 2019, and to reach back before a competitor's 2015 public disclosure it needed support in an application filed in 2004. That older application never mentioned a chain limited to those two amino acids. So the expert told the jury it would be a "straightforward leap" for a skilled scientist to get there from what the 2004 application did describe.

The jury found for Seagen and awarded more than $41 million plus a running royalty. The Federal Circuit, the court that hears nearly all patent appeals, took the expert's own words and ended the case with them: "That which one must leap to is obviously not there." Seagen Inc. v. Daiichi Sankyo Co., No. 23-2424 (Fed. Cir. Dec. 2, 2025).

Section 112 of the Patent Act tends to work this way. It is the part of the statute about what the patent says, as opposed to what the prior art said before it, and for most of its history practitioners treated it as the second or third invalidity argument in the brief. It has moved up. In the space of eight months in 2025 and 2026, the Federal Circuit erased the Seagen verdict, affirmed the erasure of a $107.5 million verdict against AstraZeneca, and restored a verdict for Teva that a trial judge had thrown out on the same grounds. Those three opinions cannot be understood without the two sentences at their center, both written by Justice Gorsuch for a unanimous Court in 2023: "The more one claims, the more one must enable," and the instruction that a specification may not leave skilled workers with "little more than two research assignments."

The short answer

• Section 112 does four jobs. A patent must describe the invention (written description), teach how to make and use all of what it claims (enablement), and end with claims whose boundaries are reasonably certain (definiteness). If a claim element is written as a function rather than a structure, the claim covers only the structure the specification discloses and its equivalents (means-plus-function).
• Amgen enforced an old test rather than announcing a new one. The specification must enable the full scope of the claim, and when a claim is defined by what something does across a vast class, a recipe for screening candidates usually does not get there. The Wands factors, the Federal Circuit's traditional list for measuring undue experimentation, survive.
• Functional genus claims are the main casualty. A genus claim covers a whole class of things. Claims to every antibody, enzyme, or compound that achieves a result have fared badly when the specification offers only examples and a screening process, unless it identifies a quality common to the members that work.
• A claim to a method of using a known class is treated differently, for now. In Teva v. Lilly, the court held that a jury could find a method-of-treatment claim described and enabled where the class of antibodies was well known and all of it worked for the claimed use. Rehearing en banc was denied over a dissent, and the distinction is likely to be tested again.
• Written description is its own trap. A later claim can outrun an earlier application even when the application could fairly be said to enable it. A long continuation chain aimed at a competitor's product is where this most often bites.
• Definiteness asks about reasonable certainty, and words like "about" are not safe. Terms of degree can survive when the patent supplies an objective anchor. They fail when the intrinsic record is inconsistent, and they are especially exposed when the term was used to get around prior art.
• In software, the rule against claiming a result without an algorithm has two routes. Section 112(f) applies when words like "module" or "payment handler" work as placeholders, and without a disclosed algorithm the claim is indefinite. Claims that avoid 112(f) still face enablement and written description.
• The burden matters. In court, a challenger must prove invalidity by clear and convincing evidence, yet several of the largest recent losses for patent owners came as judgment as a matter of law (JMOL), entered or ordered against a favorable jury verdict.

What the statute says, and why it has four jobs in two sentences

The operative language is short. 35 U.S.C. § 112(a) requires "a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same," along with the best mode the inventor contemplates. Subsection (b) requires "one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention." Subsection (f) allows a claim element to be written "as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof," and then limits that element to "the corresponding structure, material, or acts described in the specification and equivalents thereof."

For patents from applications filed before September 16, 2012, you will see the same rules cited as § 112, paragraphs 1, 2, and 6. The America Invents Act relabeled them without any change the courts have treated as material to the issues in this article.

The specification is the descriptive body of the patent, and the claims are the numbered sentences at the end that define the legal right. If you need a primer on how those pieces fit together, start with How to Read a Patent: Claims, Specification, and Prosecution History.

Best mode deserves one sentence and no more. Since the America Invents Act, 35 U.S.C. § 282(b)(3)(A) provides that failure to disclose the best mode "shall not be a basis on which any claim of a patent may be canceled or held invalid or otherwise unenforceable." The duty stays on the books, and applicants should honor it, but it has lost its teeth in litigation.

Why the law insists on all this: the bargain

Everything in Section 112 comes from one idea, which the Supreme Court calls the patent bargain. In exchange for bringing new designs and technologies into the public domain through disclosure, an inventor receives a limited period of protection from competitive exploitation. Bonito Boats, Inc. v. Thunder Craft Boats, Inc., 489 U.S. 141 (1989). The disclosure is the price. Congress has charged that price since the Patent Act of 1790, which required a specification so particular as to distinguish the invention and "to enable a workman or other person skilled in the art or manufacture" to make and use it.

Each requirement protects a different customer in that bargain. Enablement protects the public after expiration, since it must be able to practice what the patent covered. Written description protects against a kind of time travel, in which an applicant watches the market develop and then claims what others invented. Definiteness protects competitors during the term, who need to know where they can safely design. Means-plus-function is a compromise that lets a drafter write in functional shorthand in return for tying the claim back to what was actually disclosed.

There is a real policy dispute inside this bargain. Broad claims encourage investment in pioneering work, and narrow claims can be designed around, sometimes by a single amino acid or a refactored module. Amgen raised exactly that objection, warning that its loss would destroy incentives for breakthrough inventions. The Court's answer was that striking that balance is a judgment for Congress, and that since 1790 Congress has made enablement one feature of the balance. Academic critics have pressed the other side too. Mark Lemley and Jacob Sherkow, in The Antibody Patent Paradox, 132 Yale L.J. 994 (2023), a paper the Court itself cited, examined how far broad functional antibody claims had drifted from what their patents actually taught. The law's direction since then has moved toward narrower claims.

From the telegraph to the antibody: the history of enablement

Morse, Edison, and the glue that worked like animal glue

Amgen rests on three old cases, and it pays to read them the way the Court did.

In O'Reilly v. Morse, 56 U.S. (15 How.) 62 (1854), Samuel Morse's patent had eight claims. The Court upheld the seven tied to the telegraph machinery he designed. The eighth claimed the use of electromagnetism "however developed" for printing intelligible characters at a distance, and Morse wrote that he did not intend to limit himself to the machinery he had described. The Court held that claim "too broad, and not warranted by law." If it could stand, the Court observed, there would be no need for any specification beyond saying that he had discovered that electromagnetism could print at a distance.

In the Incandescent Lamp Patent, 159 U.S. 465 (1895), the patent's owners, Sawyer and Man, had made a lamp with a carbonized paper conductor that failed commercially. They claimed every "fibrous or textile" conductor and sued Edison, whose men had searched the globe and found that a particular bamboo worked. Most fibrous and textile materials did not work. The Court sided with Edison in a sentence worth remembering: that paper happens to belong to the fibrous kingdom "did not invest" the patentees "with sovereignty over this entire kingdom." The Court left a door open, and it remains the most important door in this area. The broad claim might have been allowed if they had disclosed a quality common to the class that made its members peculiarly fit for the purpose.

In Holland Furniture Co. v. Perkins Glue Co., 277 U.S. 245 (1928), the patentee had invented a particular starch glue but also claimed any starch glue that, with about three parts or less of water, would have substantially the same properties as animal glue. The specification described the key input by its use rather than its properties. A person trying to use or avoid the claim could do so "only after elaborate experimentation."

The three cases share a shape. Someone found one thing that works, then claimed everything that works.

What the old cases allowed

The Court has never required a specification to spell out every embodiment. A patent on brickmaking survived in Wood v. Underhill, 46 U.S. (5 How.) 1 (1846), even though the proportions of coal dust had to be adjusted for different clays. In Minerals Separation, Ltd. v. Hyde, 242 U.S. 261 (1916), an ore separation process survived although preliminary tests were needed for each ore, because ores vary infinitely and the certainty the law requires is no greater than is reasonable. The rule that emerges is about the character of the remaining work. Adjusting a known process to local conditions is ordinary skill. Searching for what will work at all is invention.

Wands and the eight factors

The Federal Circuit put a framework around this in In re Wands, 858 F.2d 731 (Fed. Cir. 1988). The claims were to immunoassay methods using high-affinity monoclonal antibodies against hepatitis B surface antigen. The Board had rejected them because, in its view, producing such antibodies was unpredictable and unreliable, given how many hybridomas (the antibody-secreting cell lines) had to be screened. The court reversed, and listed the factors that bear on undue experimentation:

1. The quantity of experimentation necessary.
2. The amount of direction or guidance presented.
3. The presence or absence of working examples.
4. The nature of the invention.
5. The state of the prior art.
6. The relative skill of those in the art.
7. The predictability or unpredictability of the art.
8. The breadth of the claims.

The court's reason for reversing matters today. Practitioners in the field were prepared to screen negative hybridomas, the disclosure gave considerable guidance and working examples, skill was high, and the methods were known. An "experiment" in that art was the whole attempt to make a monoclonal antibody against an antigen, and screening was simply what the work consisted of. Wands was a patent owner's favorite for thirty-five years, and it remains good law. The Federal Circuit has since been careful to say that Wands does not hold that antibody screening never requires undue experimentation. Whether it does depends on the claim and the record, and the Wands claims used antibodies of a defined type and affinity against one antigen in an assay.

The Federal Circuit's quiet march to "full scope"

Amgen did not come out of nowhere. For about twenty-five years the Federal Circuit had been tightening the same screw, often in cases that had nothing to do with biology.

• Genentech, Inc. v. Novo Nordisk A/S, 108 F.3d 1361 (Fed. Cir. 1997), produced the line later courts quote constantly: "It is the specification, not the knowledge of one skilled in the art, that must supply the novel aspects of an invention." The familiar idea that a patent need not disclose what is well known is "merely a rule of supplementation, not a substitute for a basic enabling disclosure."
• Liebel-Flarsheim Co. v. Medrad, Inc., 481 F.3d 1371 (Fed. Cir. 2007), is the great cautionary tale. The patent owner deleted the pressure jacket from its claims during prosecution so that they would cover a competitor's jacketless injector, then won that claim construction on appeal. The specification taught only injectors with jackets, and the inventors had themselves failed to build one without. "The irony of this situation," the court wrote, is that Liebel won the battle and then could not show the broader claim was enabled. "The motto, 'beware of what one asks for,' might be applicable here."
• Automotive Technologies International, Inc. v. BMW of North America, Inc., 501 F.3d 1274 (Fed. Cir. 2007), held a side-impact sensor claim not enabled where the patent taught mechanical sensors in detail and gave little on electronic ones, the novel part of what was claimed.
• Sitrick v. Dreamworks, LLC, 516 F.3d 993 (Fed. Cir. 2008), is a software-side example. Claims to integrating a user's image into a pre-existing work covered both video games and movies, and the specification's techniques worked only for games. "Because the asserted claims are broad enough to cover both movies and video games, the patents must enable both embodiments."
• ALZA Corp. v. Andrx Pharmaceuticals, LLC, 603 F.3d 935 (Fed. Cir. 2010), rejected the idea that the knowledge of the skilled worker can substitute for missing information about a claimed dosage form.
• MagSil Corp. v. Hitachi Global Storage Technologies, Inc., 687 F.3d 1377 (Fed. Cir. 2012), dealt with an open-ended claim. The patent recited a resistance change of "at least 10%" and could not support everything above that number.
• Wyeth & Cordis Corp. v. Abbott Laboratories, 720 F.3d 1380 (Fed. Cir. 2013), and Enzo Life Sciences, Inc. v. Roche Molecular Systems, Inc., 928 F.3d 1340 (Fed. Cir. 2019), found functional classes not enabled because identifying the operative members required screening.
• Idenix Pharmaceuticals LLC v. Gilead Sciences Inc., 941 F.3d 1149 (Fed. Cir. 2019), held that claims to a method of using nucleosides to treat hepatitis C were not enabled where many thousands of candidates had to be screened for efficacy.
• Trustees of Boston University v. Everlight Electronics Co., 896 F.3d 1357 (Fed. Cir. 2018), reversed a jury verdict in a semiconductor case because a claim construed to cover a growth configuration the specification did not teach was not enabled as a matter of law.

The court also policed the other side of the line. Alcon Research Ltd. v. Barr Laboratories, Inc., 745 F.3d 1180 (Fed. Cir. 2014), holds that the Wands factors do not float free as a general test of breadth. The challenger must first put forward evidence that some experimentation is needed, and working examples are not required, since the burden is on the challenger to show that prophetic examples are not enabling (citing Atlas Powder Co. v. E.I. du Pont de Nemours & Co., 750 F.2d 1569 (Fed. Cir. 1984)). McRO, Inc. v. Bandai Namco Games America Inc., 959 F.3d 1091 (Fed. Cir. 2020), reversed a summary judgment because the challengers never identified with particularity a method within the claim that was not enabled. The Wands analysis, the court said, routinely involves "concrete identification of at least some embodiment or embodiments asserted not to be enabled," so that breadth is shown "concretely and not just as an abstract possibility." That sentence should be taped above the desk of anyone drafting an enablement contention.

Amgen v. Sanofi, read closely

The facts

LDL cholesterol is removed from the blood by LDL receptors, and a protein called PCSK9 degrades those receptors. An antibody that binds PCSK9 at the right place, a stretch of 15 amino acids out of 692 that the parties called the "sweet spot," blocks the degradation. Amgen and Sanofi each developed such a drug, Repatha and Praluent, and each patented its own antibody by amino acid sequence in 2011. Those patents were not the dispute.

The dispute concerned two later Amgen patents, issued in 2014. Their claims did not recite a sequence. They claimed the entire genus of antibodies that (1) bind to specific residues on PCSK9 and (2) block PCSK9 from binding LDL receptors. The specification gave the sequences of 26 antibodies and the three-dimensional structures of two. For everything else it offered two methods. The "roadmap" told scientists to generate antibodies, test which bind, test which of those bind the sweet spot, and test which of those block. "Conservative substitution" told them to start with an antibody that works, swap selected amino acids for similar ones, and test the result.

The genus held, on the record, at least millions of candidates. And Amgen's own expert conceded the science was unpredictable: translating a sequence into a known three-dimensional structure, he said, was not yet possible and would get somebody a Nobel Prize. The District of Delaware granted judgment as a matter of law, and the Federal Circuit affirmed in Amgen Inc. v. Sanofi, 987 F.3d 1080 (Fed. Cir. 2021).

The holding

Justice Gorsuch wrote for a unanimous Court in Amgen Inc. v. Sanofi, 598 U.S. 594 (2023). The rule is one the Court called a simple statutory command: "If a patent claims an entire class of processes, machines, manufactures, or compositions of matter, the patent's specification must enable a person skilled in the art to make and use the entire class." Or, as the Court compressed it, "The more one claims, the more one must enable." Amgen, 598 U.S. at 610.

The two methods, the Court said, "amount to little more than two research assignments." The roadmap "merely describes step-by-step Amgen's own trial-and-error method." Conservative substitution was no better, given the state of the art. They left scientists where Sawyer and Man left Edison, forced to engage in "painstaking experimentation" to see what works. "That is not enablement. More nearly, it is 'a hunting license,'" quoting Brenner v. Manson, 383 U.S. 519 (1966).

The Court's analogy, borrowed from an amicus brief, is worth holding on to. Imagine a combination lock with 100 tumblers, each with 20 positions. By trial and error an inventor finds 26 combinations that open it, then claims all successful combinations, and instructs others to try a large set at random and record which ones work. That roadmap would produce working combinations. It would not enable anyone to make and use them, only to discover them.

What the Court carefully did not say

Practitioners who stop at the slogan misread the case. Four limits are stated in the opinion.

Examples can be enough. A specification need not describe how to make and use every embodiment. "[I]t may suffice to give an example (or a few examples) if the specification also discloses 'some general quality . . . running through' the class that gives it 'a peculiar fitness for the particular purpose.'" This is the Incandescent Lamp door, and the Court expressly left open that a roadmap might suffice in another patent, perhaps where the inventor identifies a quality common to every functional embodiment.

Some experimentation is fine. A specification may call for a reasonable amount of experimentation, and what is reasonable depends on the nature of the invention and the art.

Cumulative effort is not the test. Amgen argued that the Federal Circuit had measured enablement by how long it would take to make every embodiment. The Court agreed that enablement "is not measured against the cumulative time and effort it takes to make every embodiment within a claim," and read the Federal Circuit not to have held otherwise. So a challenger's arithmetic on how many centuries it would take to exhaust the genus does not win by itself.

There is one standard. Amgen argued that the lower court raised the bar for functional genus claims. The Court agreed in principle that "[t]here is one statutory enablement standard," and said that all the lower court had recognized was the arithmetic relation between claiming and enabling. Whatever the practical effect, no doctrine singles out biotech or functional claims as such.

What the Federal Circuit has done with Amgen

Three years of decisions now give Amgen its working content. I have read them chronologically here, because the pattern is easier to see that way.

Baxalta: routine does not rescue a research assignment

Baxalta is the case that closed the most obvious escape. The patent owner argued that its process (immunize mice, make hybridomas, screen) was routine and reliably yielded at least one new antibody every time. The court accepted that for argument's sake and held it did not matter. Making candidates and screening them to see which perform the claimed functions, it said, "is the definition of trial and error." The patent disclosed no common structural or other feature delineating which antibodies would work, and did not explain why its eleven worked or why others failed.

The panel also answered the question everyone was asking about old law. It did not read Amgen to disturb Wands, and saw "no meaningful difference between Wands' 'undue experimentation' and Amgen's '[un]reasonable experimentation' standards."

Seagen: when functional language and an unlimited element combine

Beyond the written description ruling described at the top of this article, the panel in Seagen gave an instructive enablement analysis. The district court had construed "D is a drug moiety" to mean any drug, as the patent owner wanted. The claim also recited that the drug is intracellularly cleaved in a patient. Together those covered any conjugate with any drug that cleaves, and the inventor admitted a skilled person would need an assay to tell whether any given conjugate did. The court noted that "the scale of trial and error is so vast and the science so unpredictable." It is a reminder that winning a broad construction can be expensive, and that Liebel-Flarsheim now has company.

Brita: performance metrics are functional claims in disguise

It would be a mistake to file this area under biotechnology. Brita involved a gravity-fed water filter defined by a "Filter Rate and Performance" factor, computed from a formula with variables for media volume, filtration time, lead concentration, and claimed lifetime, with a claimed value of about 350 or less. The claim covered filter media generally. The specification's only embodiments that met the number were carbon blocks, its text called them "unique," it described the drawbacks of other media, and the inventors testified they had moved to carbon block to solve the problem. The Commission found no "road map" for other media, the variables were interrelated in a nonlinear way, and the Federal Circuit affirmed on both written description and enablement.

Brita's argument that water filtration is a mature, predictable art failed on its own record. Engineers should take the point. A parameter claim that looks structural because it has a number in it is still a claim to everything that reaches the number.

Wyeth: the dosing limitation that sank $107.5 million

Wyeth's claims covered treating resistant lung cancer by administering daily to the patient a "unit dosage" of an irreversible inhibitor that binds a particular residue. The court had adopted the specification's own definition of "unit dosage," a quantity "calculated to produce the desired therapeutic effect." The jury found infringement by Tagrisso and awarded $107.5 million. The district court granted JMOL, and in July 2026 the Federal Circuit affirmed.

The opinion is careful on one point that had the drug industry anxious. The panel recognized that it is accepted practice to claim a method of treatment reciting a range of doses without clinical data, leaving safety and efficacy to the FDA, and it cited United Therapeutics Corp. v. Liquidia Technologies, Inc., 74 F.4th 1360 (Fed. Cir. 2023), for that. The trouble was specific. These claims were limited to dosage forms given to patients, perhaps because of close prior art, as the panel itself speculated. The specification gave only in vitro data, general and "projected" ranges with no explanation of where they came from, and the sentence "The skilled artisan is aware of the effective dose for each patient." The inventors' own later paper and testimony showed that the concentrations required in the test tube could not be achieved in patients because of toxicity, and that all therapeutic doses of two of the three named compounds exceeded the tolerated maximum.

Because the claim class was functional (any compound that binds the residue), the patent had to teach dosing across that class, and it left that to the reader. The panel treated toxicity evidence as showing the absence of guidance, and said clearly that the claims carried no freestanding FDA safety requirement.

Teva: the method-of-use line, and why it is contested

Teva v. Lilly is the most important opinion in the group, and it points the other way. Teva's headache patents claimed treating headache by administering an effective amount of a humanized anti-CGRP antagonist antibody. The specification disclosed one humanized antibody and several mouse versions, and prior-art humanization methods. A jury found the claims valid and infringed, and the verdict, reported at about $177 million, was set aside by JMOL. The Federal Circuit reversed.

On written description, the court relied on a line running through Ajinomoto Co. v. International Trade Commission, 932 F.3d 1342 (Fed. Cir. 2019), and the predecessor court's decisions in In re Herschler and In re Fuetterer, for claims where the genus is well known and is not itself the invention but an ingredient in a different one. A reasonable jury could find such antibodies well known. Lilly's own statements in its earlier inter partes review petitions, calling the art "replete" with them and humanization "routine," were used against it. A skilled reader would understand from the specification that all such antibodies treat headache. The court distinguished University of Rochester v. G.D. Searle & Co., 358 F.3d 916 (Fed. Cir. 2004), and Ariad as cases that amounted to methods of doing X using something that does X. Here the claims were to treating headache with antibodies defined by a different function. The panel also read AbbVie Deutschland GmbH & Co. v. Janssen Biotech, Inc., 759 F.3d 1285 (Fed. Cir. 2014), not to require disclosure of something structurally similar to the accused product in every case.

On enablement, the court assumed in Lilly's favor that making every member of the genus would take undue work, and still reversed. Had the claims been to the antibodies themselves, the case would resemble Amgen. But these claims did not cover the antibodies for all purposes. The relevant research assignment was finding which antibodies treat headache, and that assignment was done, since the specification disclosed that they all do. Finding or making all of them was, in the court's memorable phrase, "more akin to extra credit than a necessary research assignment left to others to complete."

Read that against Idenix and Wyeth, and you can derive the current structure. For a method claim using a class:

• If the class is well known and the specification teaches that all its members work in the method, Teva offers real protection.
• If efficacy varies across the class and must be screened, Idenix controls.
• If the method claim imports a functional dosage limitation that the specification leaves to the reader, Wyeth is the closer guide.

Lilly petitioned for rehearing en banc, arguing the decision conflicted with Amgen, and amici reportedly included both Amgen and Sanofi. The court denied rehearing on September 30, 2026, with Judge Dyk noted as dissenting. A petition for certiorari remains possible as this article is dated. The honest characterization is that Teva is binding panel precedent resting on a jury-deferential record with a specific and somewhat unusual set of admissions. The limits of the "well-known genus" category will be fought over for years. What counts as well known when a claim says "humanized" and the prior art held none? The opinion answers that for this record. It does not answer it for all.

Satius: inoperable alternatives after Amgen

Atlas Powder tolerated some inoperative embodiments in a claim, reasoning that a skilled person would know how to avoid them, unless the number becomes significant and forces undue experimentation. Satius, decided in October 2026, marks an important boundary. The claim recited an apparatus for transmitting "electric or electromagnetic signals over air." It is undisputed physics that you cannot transmit electric signals over air in electric form. The district court called that indefinite. The Federal Circuit disagreed, because the claim's meaning was clear ("Indefiniteness is not the same as impossibility"), then invalidated it for lack of enablement anyway, reaching the issue for the first time on appeal because the answer was so plain.

The full-scope standard, the court said, "cannot possibly be met" when a claim includes "an express limitation that adds inoperable (and thus non-enabled) alternative embodiments." Drafters should see the practical meaning: an "or" in a claim is a promise about every branch. Contrast the nonprecedential Indect USA Corp. v. Park Assist, LLC, No. 24-1023 (Fed. Cir. Jan. 7, 2026), where the challenger's impossibility argument rested on reading method steps as simultaneous. No skilled person would read them that way, and the claim survived.

Evidence from after the filing date

Enablement is judged at the effective filing date, but that does not exclude all later evidence. Amgen Inc. v. Sanofi, Aventisub LLC, 872 F.3d 1367 (Fed. Cir. 2017), held it error to exclude post-priority evidence offered to show that the disclosure was not representative or that later efforts took lengthy experimentation. Pioneer Hi-Bred applied that against the patent owner's own later testing: of eight enzymes meeting the structural limits, only two showed the claimed function, and weakly.

In re Entresto marks the other edge. Novartis claimed a combination of valsartan and sacubitril; a later-discovered complex form of the two drugs was not known at filing. The district court had found the claims not described because they would cover complexes. The Federal Circuit reversed, explaining that later-existing technology cannot be used to reach back and invalidate a claim to the underlying invention, following In re Hogan, 559 F.2d 595 (C.C.P.A. 1977). Pioneer drew the line this way: Hogan and Entresto exclude technology that did not exist at the priority date, while evidence about embodiments that could have been made then is fair game even if tested later.

The USPTO's reading

The Patent Office responded with Guidelines for Assessing Enablement in Utility Applications and Patents in View of the Supreme Court Decision in Amgen Inc. et al. v. Sanofi et al., 89 Fed. Reg. 1563 (Jan. 10, 2024). The notice says it is not intended to announce any major changes in practice, and that examiners in all technologies will continue to use the Wands factors to decide whether the required experimentation is reasonable. It asks for adequate reasoning grounded in those factors. It does not address written description. Applicants therefore should not expect an examiner to say "Amgen" and stop; they can and should demand a factor-by-factor rejection, and the nonprecedential PEN decision shows what happens when an applicant answers by addressing only some embodiments.

Written description: possession, shown on the page

Ariad and the separate requirement

Is there really a written description requirement apart from enablement? The statutory sentence can be parsed either way, and the argument that there is not went to the en banc Federal Circuit in Ariad Pharmaceuticals, Inc. v. Eli Lilly & Co., 598 F.3d 1336 (Fed. Cir. 2010) (en banc). The court reaffirmed that there is, and held that the test is whether the disclosure "reasonably conveys to those skilled in the art that the inventor had possession of the claimed subject matter as of the filing date."

Then the court added something more useful than the test. "The term 'possession,' however, has never been very enlightening." What matters is that "the hallmark of written description is disclosure." So the inquiry is objective, confined to the four corners of the specification, viewed through the eyes of a skilled person. Lab notebooks do not count. No examples or reduction to practice are required, and the words need not be identical. But a description that merely renders the invention obvious is not enough, a point drawn from Lockwood v. American Airlines, Inc., 107 F.3d 1565 (Fed. Cir. 1997). The degree of detail varies with the nature and scope of the claims and the complexity and predictability of the technology.

Ariad's claims covered reducing the activity of a gene regulator (NF-kB) by any means, and the specification only hypothesized molecules that might do so. A jury verdict for the patent owner was reversed.

Why the two requirements are not the same

Writers often treat these as twins. Ariad itself said they often rise and fall together, and many recent decisions find both failures, but the difference has teeth.

Consider a claim added years after filing. A skilled person might be able to make and use it from the disclosure and general knowledge, which goes to enablement. But if the text never singled it out, the inventor has not shown possession. Seagen is the model: the problem was not that nobody could make a glycine-phenylalanine chain, but that the 2004 application pointed at 47 million tetrapeptides and never at those 81. Conversely, a specification might fully describe a thing that nobody can actually make without undue experimentation, which Satius illustrates in an extreme form.

Genus claims

Most of the hard cases concern a claim to a class. The governing formula from Ariad, drawing on Regents of the University of California v. Eli Lilly & Co., 119 F.3d 1559 (Fed. Cir. 1997), is that a genus can be described by "a representative number of species falling within the scope of the genus or structural features common to the members of the genus so that one of skill in the art can 'visualize or recognize' the members of the genus." A mere "wish" or "plan" for obtaining the invention will not do, as Fiers v. Revel, 984 F.2d 1164 (Fed. Cir. 1993), held.

Some landmark applications:

• Enzo Biochem, Inc. v. Gen-Probe Inc., 323 F.3d 956 (Fed. Cir. 2002): a function can describe the invention if it is correlated with structure known to the art, and a public deposit can do descriptive work.
• Capon v. Eshhar, 418 F.3d 1349 (Fed. Cir. 2005), and Falko-Gunter Falkner v. Inglis, 448 F.3d 1357 (Fed. Cir. 2006): the specification need not recite what skilled people already know, and the requirement does not demand a rote recital.
• Noelle v. Lederman, 355 F.3d 1343 (Fed. Cir. 2004): an antibody claimed by what it binds is described only if the antigen is itself sufficiently described.
• AbbVie: a genus of antibodies defined by binding functions was not described where the examples came from a single lineage and the accused antibody differed sharply.
• Juno Therapeutics, Inc. v. Kite Pharma, Inc., 10 F.4th 1330 (Fed. Cir. 2021): this is the highest-dollar illustration. The jury awarded a $585 million upfront payment plus a running royalty, and after enhancement for willfulness the judgment exceeded $1.1 billion, on a CAR-T patent claiming any single-chain antibody fragment that binds a selected target. The patent disclosed two such fragments and no sequences. Even if such fragments were known, the court held, the inventors had to convey possession of all of them, known and unknown, and the specification offered no means of telling which would bind which target.

Amgen's earlier appeal also retired a shortcut. In 872 F.3d 1367, the court rejected the "newly characterized antigen" test, under which describing an antigen was thought to describe every antibody against it, as inconsistent with Ariad.

Blaze marks and the priority trap

Priority chains are where written description becomes a fight about dates. To claim the benefit of an earlier application under 35 U.S.C. § 120, that application must disclose the invention in the manner § 112(a) requires. And when a patent owner loses that benefit, its own later-filed claim can be anticipated by whatever came in between, often the accused product. The Federal Circuit said in Seagen, drily, that one "cannot avoid the suspicion" the 2019 application was filed to encompass Enhertu, and that this is permissible only if the earlier application supported it.

The doctrinal image comes from In re Ruschig, 379 F.2d 990 (C.C.P.A. 1967). A trail through the woods needs blaze marks on particular trees, and a large number of unmarked trees is no help. Regents of the University of Minnesota v. Gilead Sciences, Inc., 61 F.4th 1350 (Fed. Cir. 2023), applied the same image to a subgenus of compounds in an IPR, where a priority loss made the claims vulnerable to prior art that is otherwise off limits. For how priority interacts with provisional filings, see Provisional Patent Applications: What They Protect, What They Do Not, and How to Keep the Priority Date, and for the intervening-art consequences, see Prior Art Under the America Invents Act: Section 102, the On-Sale Bar, Public Use, and the One-Year Grace Period.

Inventor testimony is often the most damaging proof. In Seagen, inventors testified that they first saw a glycine-phenylalanine chain in Enhertu or in the claims themselves. In Brita, they admitted they had not made non-carbon-block filters that met the metric. Nuvo Pharmaceuticals (Ireland) Designated Activity Co. v. Dr. Reddy's Laboratories Inc., 923 F.3d 1368 (Fed. Cir. 2019), says such testimony can illuminate the absence of critical description. The deposition preparation consequences are obvious, and the conduct rules are covered in Handling Objections, Coaching, and Misconduct at a Deposition: A Practical Guide.

Non-genus cases, and software

The requirement is not limited to chemistry. Ariad noted that it has been applied beyond chemical and biological inventions, and some of the clearest illustrations are mechanical.

In Gentry Gallery, Inc. v. Berkline Corp., 134 F.3d 1473 (Fed. Cir. 1998), a sectional sofa patent solved its problem by placing recliner controls on a console; later claims that freed the controls from the console went beyond the disclosure. In Tronzo v. Biomet, Inc., 156 F.3d 1154 (Fed. Cir. 1998), a generic claim to a hip prosthesis cup of any shape could not rest on a parent describing a conical one. ICU Medical, Inc. v. Alaris Medical Systems, Inc., 558 F.3d 1368 (Fed. Cir. 2009), found no description for spikeless claims where the specification described only a spike. And PowerOasis, Inc. v. T-Mobile USA, Inc., 522 F.3d 1299 (Fed. Cir. 2008), denied priority to continuation claims that reached beyond the machine-based interface originally described.

In software, the leading case is LizardTech, Inc. v. Earth Resource Mapping, Inc., 424 F.3d 1336 (Fed. Cir. 2005). The specification taught one way to produce a wavelet image transform without artifacts at tile boundaries, by maintaining updated sums of coefficients. A broader claim covered any method of doing it. The court offered an analogy that software lawyers should know: a patent that teaches one fuel-efficient engine does not thereby describe every possible type of fuel-efficient engine.

Silence and negative limitations present their own risk. In Novartis Pharmaceuticals Corp. v. Accord Healthcare Inc., 38 F.4th 1013 (Fed. Cir. 2022), the court said that "[s]ilence is generally not disclosure," so a later claim excluding a loading dose needed something in the specification to support the exclusion. And In re Xencor, Inc., No. 24-1870 (Fed. Cir. Mar. 13, 2025), held that the limiting preamble of a Jepson claim, the old form that recites the prior art and then the improvement, must itself be supported by adequate written description, with sufficiency depending on the art. Applicants cannot assume the preamble is someone else's problem.

Indefiniteness: reasonable certainty

Nautilus

For about a decade, the Federal Circuit asked whether a claim was "amenable to construction" and not "insolubly ambiguous," a test few claims failed. In Nautilus, Inc. v. Biosig Instruments, Inc., 572 U.S. 898 (2014), a case about a heart rate monitor with electrodes "in spaced relationship," the Supreme Court replaced it. A patent is invalid for indefiniteness "if its claims, read in light of the specification delineating the patent, and the prosecution history, fail to inform, with reasonable certainty, those skilled in the art about the scope of the invention."

The Court tied the test to incentives. Patent drafters have reason to inject ambiguity and to keep their options open, while the public needs clear notice and should not be exposed to a "zone of uncertainty" that experimentation may enter only at the risk of infringement claims, a phrase the Court drew from United Carbon Co. v. Binney & Smith Co., 317 U.S. 228 (1942). The standard "mandates clarity, while recognizing that absolute precision is unattainable." And it is measured when the application was filed, from the skilled person's perspective, not by a court looking at matters afterward. On remand, the Federal Circuit found the claim definite, Biosig Instruments, Inc. v. Nautilus, Inc., 783 F.3d 1374 (Fed. Cir. 2015).

The three common failures

Multiple measures with different answers. In Teva Pharmaceuticals USA, Inc. v. Sandoz, Inc., 789 F.3d 1335 (Fed. Cir. 2015), "molecular weight" could mean three things that gave different numbers, and nothing in the record chose. In Dow Chemical Co. v. Nova Chemicals Corp., 803 F.3d 620 (Fed. Cir. 2015), a "slope of strain hardening coefficient" could be measured by several methods, and the court set aside a damages judgment. In HZNP Medicines LLC v. Actavis Laboratories UT, Inc., 940 F.3d 680 (Fed. Cir. 2019), a property called "better drying time" had two test methods in the specification with inconsistent results. The rule that emerges for drafters is simple. If a number depends on how you measure it, the specification must name the method.

Subjective terms without an objective anchor. Datamize, LLC v. Plumtree Software, Inc., 417 F.3d 1342 (Fed. Cir. 2005), held, even under the older and more forgiving test, that "aesthetically pleasing" gave no objective standard. Interval Licensing LLC v. AOL, Inc., 766 F.3d 1364 (Fed. Cir. 2014), held the same for displaying content "in an unobtrusive manner that does not distract a user," which depended on the preferences of the particular user. The court said the claims must provide "objective boundaries," and that a claim is not definite merely because a court can give it some meaning.

Yet subjective-sounding terms can pass. Sonix Technology Co. v. Publications International, Ltd., 844 F.3d 1370 (Fed. Cir. 2017), upheld "visually negligible" where the specification gave examples, an expert's test results persuaded the examiner in reexamination, and the examiner and the parties had worked with the term without difficulty. Niazi Licensing Corp. v. St. Jude Medical S.C., Inc., 30 F.4th 1339 (Fed. Cir. 2022), upheld "resilient" and "pliable" for catheters on similar reasoning. Functional language is not itself the problem, as BASF Corp. v. Johnson Matthey Inc., 875 F.3d 1360 (Fed. Cir. 2017), and Cox Communications, Inc. v. Sprint Communication Co., 838 F.3d 1224 (Fed. Cir. 2016), make plain. Nevro Corp. v. Boston Scientific Corp., 955 F.3d 35 (Fed. Cir. 2020), adds that a term defined by a result, "paresthesia-free," can be definite where a skilled person can determine whether it is met.

Terms used to escape the prior art. This is the theme of 2026. Enviro Tech Chemical Services, Inc. v. Safe Foods Corp., No. 24-2160 (Fed. Cir. May 4, 2026), involved a poultry processing method with a pH "of about 7.6 to about 10." The claim had been amended up from about 7.3 to avoid art near 7.0. "About" is not inherently indefinite; words of approximation are often fine, as Ortho-McNeil Pharmaceutical, Inc. v. Caraco Pharmaceutical Laboratories, Ltd., 476 F.3d 1321 (Fed. Cir. 2007), recognized. But here the specification's experiments sometimes accepted deviations up to 0.3 and sometimes up to 0.5, including the largest run with 5.8 million chickens, and in prosecution the applicant was inconsistent about whether "about" mattered. With art so near, the court said the definiteness requirement necessitated "much more clarity," adding: "The prior art is almost 'about' a pH of 7.6."

The nonprecedential Network-1 Technologies, Inc. v. Google LLC, No. 24-1893 (Fed. Cir. Apr. 23, 2026), makes the same point in search software. The claims recited a "non-exhaustive search" of a reference data set. The phrase never appeared in the original application and was added years later, partly to overcome art. The specification contrasted "linear" search with other techniques without drawing a clear line for how much of a record a search must examine, and the academic literature offered no settled meaning for it. The panel's comment should travel well: "It does the patent system or the public no favor for this court to accept the distinguishing of prior art with vague language."

Compare Infinity Computer Products, Inc. v. Oki Data Americas, Inc., 987 F.3d 1053 (Fed. Cir. 2021), which held that inconsistent prosecution statements can create indefiniteness where the claim and specification leave uncertainty.

Other indefiniteness doctrines

A single claim that mixes an apparatus and a method of using it is indefinite, since it leaves unclear when infringement occurs. IPXL Holdings, L.L.C. v. Amazon.com, Inc., 430 F.3d 1377 (Fed. Cir. 2005). Software claims reciting a user's act inside a system claim are the usual occasion.

A claim that requires something internally impossible can be indefinite when a skilled person would see that the claim does not express what the inventor meant, as in Synchronoss Technologies, Inc. v. Dropbox, Inc., 987 F.3d 1358 (Fed. Cir. 2021). But Satius now says clearly there is no per se rule, and that impossibility is usually an enablement problem.

Antecedent basis slips are not automatically fatal. In the nonprecedential MPH Technologies Oy v. Apple Inc., No. 25-1069 (Fed. Cir. Aug. 3, 2026), the court walked through "the secure connection" step by step against the figures and reversed an indefiniteness ruling on a VPN claim.

The Patent Office applies a different standard

During examination, an applicant is in a position to fix the language, and the Office may reject a claim that "contains words or phrases whose meaning is unclear." In re Packard, 751 F.3d 1307 (Fed. Cir. 2014). MPEP Sections 2171 to 2173.05 collect the Office's practice. An issued claim that survived an examiner is therefore not safe, and a term that the applicant might have cleared up at modest cost during prosecution can later fail in litigation. Conversely, examiner understanding is evidence, as Sonix illustrates.

Means-plus-function: the bargain for functional drafting

How we got here

Pure functional claiming once had a bad name. In Halliburton Oil Well Cementing Co. v. Walker, 329 U.S. 1 (1946), and General Electric Co. v. Wabash Appliance Corp., 304 U.S. 364 (1938), the Court invalidated claims that recited a result at the point of novelty. In 1952 Congress allowed such elements on condition that they be construed narrowly. That is today's § 112(f): you may claim a combination element by function, and the element will reach only the corresponding structure in the specification and equivalents.

Williamson: the presumption loses its thumb

For years, the absence of the word "means" created a "strong" presumption against § 112(f). The en banc portion of Williamson v. Citrix Online, LLC, 792 F.3d 1339 (Fed. Cir. 2015), overruled that characterization, saying it placed "a thumb on what should otherwise be a balanced analytical scale" and had led to a proliferation of functional claiming "untethered" to the statute. The test now is whether the words of the claim "are understood by persons of ordinary skill in the art to have a sufficiently definite meaning as the name for structure."

The term in Williamson was "distributed learning control module." The court called "module" a well-known nonce word (a placeholder) and treated the prefix as adding no structure. The specification disclosed no algorithm for the coordinating function, so the claim failed.

The two-step and the algorithm rule

Once § 112(f) applies, a court identifies the claimed function and then the corresponding structure the specification clearly links to it. If there is none, the claim is indefinite under § 112(b). For a computer-implemented element, a general-purpose computer is not enough. WMS Gaming Inc. v. International Game Technology, 184 F.3d 1339 (Fed. Cir. 1999), established that the structure is the computer programmed to carry out the disclosed algorithm. Aristocrat Technologies Australia Pty Ltd. v. International Game Technology, 521 F.3d 1328 (Fed. Cir. 2008), rejected the argument that "appropriate programming" of a standard microprocessor was disclosure enough: "The question thus is not whether the algorithm that was disclosed was described with sufficient specificity, but whether an algorithm was disclosed at all."

An algorithm need not be source code. Finisar Corp. v. DirecTV Group, Inc., 523 F.3d 1323 (Fed. Cir. 2008), allows it to be expressed in any understandable form, including a formula, prose, or a flow chart. Typhoon Touch Technologies, Inc. v. Dell, Inc., 659 F.3d 1376 (Fed. Cir. 2011), accepted a description in words. But Noah Systems, Inc. v. Intuit Inc., 675 F.3d 1302 (Fed. Cir. 2012), rejected language that merely restated the function, and Blackboard, Inc. v. Desire2Learn, Inc., 574 F.3d 1371 (Fed. Cir. 2009), rejected an "access control manager" described as a black box. Expert testimony that a programmer could implement the function does not rescue the disclosure: the question concerns what the patent discloses.

There is a narrow exception. In re Katz Interactive Call Processing Patent Litigation, 639 F.3d 1303 (Fed. Cir. 2011), held that a general-purpose computer may suffice for functions such as processing, receiving, and storing, which any such computer does without special programming. EON Corp. IP Holdings LLC v. AT&T Mobility LLC, 785 F.3d 616 (Fed. Cir. 2015), confined the exception and described it as narrow.

Which words trigger it

TrackTime: the current state of the art

The most useful recent statement is TrackTime, LLC v. Amazon.com Services LLC, No. 24-1102 (Fed. Cir. July 2, 2026). The patent claimed "executable program code configured to facilitate annotation" of a transcript on a mobile device. The Delaware court held that generic and indefinite in 2021. Dyfan then came down, and the Federal Circuit vacated.

The opinion is a careful guide to the structure and function line:

• The proponent of § 112(f) treatment bears the burden, and any underlying fact must be shown by a preponderance. Extrinsic evidence of how the words are used in the field can be essential.
• Begin by parsing the phrase into the portion reciting function and the portion naming what performs it.
• The question is whether the claim recites sufficient structure for the claimed function, so the character of the function matters. The more general the function, the more the claim looks like a request for a result.
• A term can be structural even if its name reflects a function, and even a term that is non-structural in a vacuum can gain structure from surrounding language. In software, that happens when the claim calls for a particular operation, which "is more than just its function; it is how the function is achieved in the context of the invention."

The panel noted that code terms are "partly defined" by function, and that the claims here looked more about user-desired results than technological means. It did not resolve that; it remanded for evidence. What it does make clear is that neither "code" nor "configured to" is a magic word in either direction, and that a district court should not resolve the issue without looking at what practitioners would have understood.

Finally, Gramm v. Deere & Co., No. 24-1598 (Fed. Cir. Mar. 11, 2026), is a useful counterweight for hardware. A "control means" for raising and lowering a harvester header pointed to a "conventional" head controller in Deere combines. The district court concluded the corresponding structure must be a microprocessor and no algorithm was disclosed. The Federal Circuit reversed, because one commercially available model had no microprocessor and was properly corresponding structure: "A disclosure of an added function in the specification should not disqualify structure that meets the means clause's requirements." Commercial references can serve as structure if a skilled person would understand them, following Budde v. Harley-Davidson, Inc., 250 F.3d 1369 (Fed. Cir. 2001).

What is really happening when a software claim recites a result

Here I want to step away from the cases and talk as someone who has written software for a living.

A computer program implements a function in the mathematician's sense: it maps inputs to outputs. A claim saying "a module configured to select the delivery route that minimizes fuel use" specifies that mapping and says nothing about how it is computed. Any working engineer knows the size of the gap. The same input-output behavior can be produced by exhaustive enumeration, dynamic programming, a heuristic with local search, a mixed-integer solver, or a trained model. Their memory use and running time differ by many orders of magnitude, and whether any of them produces an answer in time for a truck to leave the depot may be the entire invention. If the problem is NP-hard, a claim to achieving the optimum at scale might recite something nobody knows how to do.

So what has the inventor taught a reader who has that claim and a block diagram with a box labeled "routing module"? The box is the problem restated. An interface definition, in other words, with no implementation behind it. It tells me what the method signature is. Calling it "configured to" tells me that somebody, somewhere, must write the body.

That framing also explains why Aristocrat was correct about "appropriate programming," even though engineers sometimes object that any competent programmer could write the code. Maybe so, for a conventional task. But the question the patent bargain asks is what the patentee contributed. If the claim is novel and non-obvious because of how the route is chosen, then "any competent programmer could figure it out" concedes either that the disclosure omitted the invention, or that the invention was obvious. A patent owner should think hard before making that argument.

The lock analogy, translated

Amgen's lock has a precise counterpart in software. A claim to every program that produces a given behavior is defined semantically, by what the program does, and not syntactically, by what it is. Computer scientists have a theorem about such sets. Rice's theorem says that every nontrivial property of what a program computes is undecidable in the general case: no procedure can take an arbitrary program and always decide whether it has that property. No court has rested a holding on this, and I offer it as an intuition, since real claims concern practical programs and practical testing. But it shows why "test it and see" has a deeper meaning for a claim class defined by behavior than for one defined by structure. There is often no way to recognize members of the class except by running them.

The analogy has limits that help patent owners. Software is usually a more predictable art than protein folding. A change to a function usually does what the programmer expects, which is one reason courts have long tolerated less detail in software specifications, and why Dyfan could rest on the availability of off-the-shelf code. If the function is conventional and the invention lies elsewhere in the combination, Teva's logic of a known component used as part of a different invention may have force in software too, though no precedential decision I found has yet said so in those terms.

The two routes to invalidity

A software drafter who avoids § 112(f) has not escaped. If the claim recites "a processor programmed to select the route that minimizes fuel use" and a court finds that recites sufficient structure, then the claim is not limited to the disclosed algorithm. Its scope is every way a processor can achieve that. Which means the full-scope requirements of § 112(a) now apply to every such way. LizardTech and Sitrick were decided on exactly this kind of reasoning. So was Vasudevan Software, Inc. v. MicroStrategy, Inc., 782 F.3d 671 (Fed. Cir. 2015), though there the court found fact issues on whether a claim to accessing "disparate databases" was enabled and described, which is a useful reminder that these cases can be won with real expert evidence of how skilled programmers would proceed.

The Patent Office said as much in its 2019 notice, Examining Computer-Implemented Functional Claim Limitations for Compliance With 35 U.S.C. 112, 84 Fed. Reg. 57 (Jan. 7, 2019). Under § 112(f), "the specification must disclose an algorithm for performing the claimed specific computer function." Under § 112(a), "The written description requirement is not met if the specification merely describes a 'desired result,'" and "It is not enough that one skilled in the art could theoretically write a program to achieve the claimed function." The Office republished that notice in a 2025 compilation of consolidated notices, which some commentary has mistaken for new guidance. It is the same notice.

So the functional software claim sits between two walls. Lean toward function and you trigger § 112(f), and your scope collapses to the disclosed algorithm or the claim fails. Lean toward structure in the claim and you may be limited in the marketplace. Claim the result without triggering § 112(f), and you must enable and describe all ways of reaching it. Section 101 adds a third wall: a result-focused claim is also exposed as abstract, as explained in Patent Eligibility Under Section 101: The Alice/Mayo Framework and How Courts Apply It.

A worked hypothetical: the routing claim

Suppose a startup called Ostrander Route Software files this claim: "A system comprising a processor and a routing module configured to receive delivery stops and vehicle telemetry and to select a route that minimizes fuel consumption." The specification spends three pages on the business problem and one figure with a box labeled "routing module." One paragraph says the module "may use any suitable optimization technique, such as a heuristic or neural network."

A competitor is sued. Its lawyers run the analysis in sequence.

First, § 112(f). "Module" is a nonce word under Williamson and Rain. Ostrander responds with an expert: under Dyfan and TrackTime, routing code was off the shelf in the field. The competitor replies that the claimed function is achieving a fuel minimum from telemetry, a goal rather than a known operation, and that the expert names no off-the-shelf product doing that. The court decides on the evidence, and Ostrander has the better of it only if the field really did have such tools.

Second, if § 112(f) applies, the specification must disclose an algorithm. "Any suitable optimization technique" is a statement that one exists, not a disclosure of one. The claim is indefinite under Aristocrat and Fintiv.

Third, suppose § 112(f) does not apply. Then the claim reaches any way of selecting the minimum. If the inventive contribution was a specific way of weighing telemetry, the claim has outrun it. LizardTech points toward a written description problem, and the competitor must, under McRO, identify a specific embodiment within the claim that the specification does not enable, say a reinforcement learning approach, and show the experimentation it would take. Here, predictability cuts for Ostrander and the absence of any guidance cuts against it.

A better application would have described the specific scoring function, the telemetry features, the update step, and its running-time advantage, with a dependent claim reciting them. Then the claim has a fallback, the § 101 argument has a technical improvement to point to, and any 112(f) construction has something to attach to.

Trained models and AI claims

Machine learning claims put these questions in a new and unsettled form. Consider a claim to "a neural network trained to detect defects in solder joints with at least 95% accuracy."

Think about what such a specification usually offers: a training set description, a loss function, an architecture family, and advice about tuning hyperparameters. What is training? It is an iterative search through a very large parameter space, followed by a test to see whether the result meets the metric, repeated across architectures and settings until something works. Stated that way, it strongly resembles a roadmap. Generate candidates, test, keep the ones that work. Add a claim defined by performance and you have the Brita problem. Add a claim to all models that reach the metric and you have the lock.

Yet there are real arguments the other way. Training is routine in the sense Wands meant: practitioners expect to tune models, and that is simply what the work consists of. If the disclosed pipeline reliably produces a working model, you could say the specification teaches how to make the invention as much as a recipe does. The challenge then is whether the whole claimed class can be reached that way, what share of settings work, and whether the patent identifies what makes them work, which is the common-quality question. As Baxalta warned, reliably producing at least one member does not mean you have enabled the class.

Two other developments frame the issue. The Federal Circuit held in Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025), that claims applying a generic machine learning technique in a particular environment are abstract under § 101. And the Patent Office's Appeals Review Panel in Ex parte Desjardins, Appeal No. 2024-000567 (Sept. 26, 2025, designated precedential Nov. 4, 2025), reversed a § 101 rejection of a machine learning training claim, reasoning that §§ 102, 103, and 112, not § 101, are the traditional tools for limiting scope. The Office's stated direction, in other words, moves pressure onto Section 112. As of this article's date, I did not find a precedential Federal Circuit decision applying Amgen squarely to a trained-model claim. Applicants should expect one, and draft accordingly. The ownership and inventorship questions in this area are separate, and are covered in AI-Generated Inventions: Who Owns What the Machine Creates?.

Suppose a hypothetical Kestrel Vision patent discloses one architecture, one training set of 40,000 labeled images, and the resulting 96% accuracy, and claims any model trained to detect defects at 95%. A competitor uses a wholly different architecture and a synthetic data pipeline. Kestrel's best argument is that the art is predictable and well developed. The competitor's best argument is Juno and Brita: the inventors had one thing and claimed everything that reaches the number. A dependent claim reciting the architecture and training features may survive where the independent claim does not, and that is why sound drafting provides a ladder of scope. The drafting steps are in the Drafting Patent Claims and Specifications That Survive Section 112: A Practical Guide.

Biotech genus claims in practice

What can a life sciences applicant actually do now? The cases suggest an order of preference.

Claim sequences and close variants. Sanofi and Amgen each obtained patents on their own antibodies in 2011 by sequence, and nobody questioned those. A narrow claim is easier to design around, and that is a price of the rule.

Claim a structural genus with a real common feature. Pioneer Hi-Bred shows the limits. A claim using "at least 85% identity" plus a motif looks structural, but it was held not enabled because structure was not shown to correlate with function, with a genus on the order of 2 × 10^106 species and two examples. The patent owner's own later testing showed that only two of eight variants worked. A sequence identity number with no evidence that it predicts function is not the quality Amgen had in mind. Though nonprecedential, the opinion is a good guide to what the Board and courts will ask.

Claim method of use with a known class. Teva makes this attractive, so long as the class really is known, the specification really does teach that it works generally, and the claim is not just a method of doing X by using something that does X.

Identify the mechanism. The door the Court left open needs a disclosure of why. Baxalta's patent did not explain why its eleven antibodies worked. An applicant that can say what structural feature drives the function, and show a correlation across multiple diverse examples, has the argument Amgen lacked.

A worked hypothetical: the heat-stable enzyme

Suppose Fennimore Enzymes finds, after screening 4,000 variants, that 12 stabilize a detergent protease at high temperature. It files claims to (1) any protease variant retaining at least 50% activity after an hour at 70 degrees Celsius; (2) any variant having at least 90% identity to its parent sequence and that activity; (3) variants with a substitution at any of three identified surface positions and that activity; and (4) method-of-washing claims using such variants in a known detergent formula.

Claim 1 is Holland Furniture with a modern thermostat. Screening is the only route, the hit rate is 0.3%, and nothing identifies a common quality. It should expect to fail under Amgen, Baxalta, and, for written description, Juno.

Claim 2 is Pioneer Hi-Bred. The identity number adds a structural label, but the question is correlation. If a 90% identity variant usually does not work, the number has not narrowed the search enough.

Claim 3 is where the application has a chance. If the specification explains that those surface positions form a salt bridge that accounts for the stability, and shows this across several diverse substitutions, that is close to what Incandescent Lamp described: a quality common to the class that gives it a peculiar fitness. The inquiry is fact-specific, but it is a real argument.

Claim 4 has a different profile. If stabilized proteases of this type were not known before Fennimore's work, then Teva does not help, because the genus was not well known. Compare Rochester, where no compound was known and the "method" framing did not save the claim. If the applicant discovers a new use for a well-known class, the analysis improves considerably.

That sorting is why experienced prosecutors file all of these claims, in a sequence of independent and dependent forms, with supporting disclosure for each.

Burdens, standards, and who decides what

Four procedural points deserve attention.

Clear and convincing is less protective than it sounds. Enablement's status as a question of law lets a court grant judgment despite a verdict, as in Wyeth, Seagen, and Trustees of Boston University. Yet Teva shows the reverse risk for challengers: the JMOL inquiry must account for the burden, and if a reasonable jury could have found the proof less than clear and convincing, the verdict stands. See Fed. R. Civ. P. 50.

An inter partes review cannot raise Section 112 directly. 35 U.S.C. § 311(b) limits such petitions to §§ 102 and 103 on patents and printed publications. Section 112 still matters there, because priority fights turn on written description (Regents of Minnesota) and because substitute claims offered by motion to amend must comply, as Medytox illustrates in the post-grant setting. Post-grant review under 35 U.S.C. § 321 may raise Section 112 within nine months of issue for patents subject to the first-inventor-to-file rules, on a preponderance standard, as in Pioneer Hi-Bred. For the wider context, see Patent Post-Grant Proceedings: IPR, PGR, and Reexamination at the PTAB.

Claim construction and validity meet. A broad construction is often what kills the claim, as in Liebel-Flarsheim, Seagen, and Boston University. Patent owners should test any proposed construction against § 112 before they win it. The process is described in Claim Construction and Markman Hearings: How Courts Decide What a Patent Actually Covers. Indefiniteness rulings are often entered at the claim construction stage, and Satius shows that an appellate panel may reach a § 112 ground that the trial court did not.

Admissions travel. Teva used IPR statements against Lilly in court. Challengers who call a field routine and predictable in an obviousness petition should expect to hear it when they later argue undue experimentation. Patent owners face the mirror image: calling the art unpredictable to defeat obviousness makes enablement harder, and Wyeth quoted the specification's own description of complexity against its owner. For a look at the first of those arguments in depth, see Overcoming Obviousness Rejections: A Comprehensive Guide to Section 103 Analysis and Inter Partes Review Strategy: Discretionary Denial, Estoppel, and the Parallel District Court Case.

Federal Circuit jurisdiction over these appeals rests on the arising-under statute discussed in Federal Subject-Matter Jurisdiction: Arising Under, Diversity, Supplemental Claims, and the Defect That Never Waives, and the expert fights that fill these records are governed by Fed. R. Evid. 702; see Expert Witnesses After the 2023 Amendment to Rule 702: Daubert, Disclosure, and Cross-Examination.

Where the law is unsettled

How far Teva reaches. Is the well-known genus category limited to situations with admissions like Lilly's? Does it apply to claims where a large fraction of the genus would not work, or to method claims whose class is only partly known? With rehearing denied over a dissent, and a certiorari window open as of this writing, this remains the largest open question.

What counts as a common quality. No decision applying Amgen has yet given a detailed account of a functional genus claim enabled through a disclosed common quality. We know what fails. The positive model has yet to be written.

How predictability applies to software and AI. Courts have treated software as a more predictable art, and Brita rejected a similar label on its record for water filters. No precedent yet applies Amgen to a trained model. TrackTime's remand invites evidence on what code terms conveyed in a given field at a given time, and that evidence will vary by decade.

Reliance on the "no bright line" language. Ariad says the number of species needed varies with each invention. That helps patent owners at trial and makes outcomes hard to predict.

The distinction between §§ 112(a) and 112(b) for impossible claims. Satius and Synchronoss now coexist. Satius suggests enablement is the normal home, while keeping open the possibility of a claim so nonsensical that a skilled reader cannot find its bounds.

Whether Congress will weigh in. Amgen's policy arguments were sent to Congress, and I am not aware of enacted legislation changing Section 112. Check current status before relying on this statement in any filing.

What this means for you

If you own a patent, read your independent claims and ask what is the largest thing each covers. Then ask where the specification teaches how to reach the far edge of it. If you are accused of infringing, identify concretely which embodiment is not enabled or described, tie it to your own product or a named alternative, and find out early what the inventors will say. If you are drafting, remember that a court ten years from now will read the claim against a page you can still edit.

The Section 112 Patent Review Checklist: A Practical Checklist turns this analysis into pre-filing and pre-suit tasks, and the Section 112 Toolkit: Claim Support Charts, Invalidity Contentions, and Expert Declarations gives working charts, contention language, and declaration outlines. Trade secret protection remains the alternative for technology that can be kept in-house, with its own hazards, some of which are discussed in Hiring From a Competitor: Inevitable Disclosure, Employee Raiding, and the New Employer's Trade Secret Exposure. If you want to research the opinions behind this article, UpLaw (uplaw.us) carries them. This article is general information and is not legal advice for a specific situation. If you have a claim set or a validity fight that turns on these questions, MC Law is glad to talk it through.

Related Documents

Related Articles

• How to Read a Patent: Claims, Specification, and Prosecution History: the anatomy of claims and specifications this article assumes.
• Claim Construction and Markman Hearings: How Courts Decide What a Patent Actually Covers: where indefiniteness and § 112(f) rulings usually arrive.
• Patent Eligibility Under Section 101: The Alice/Mayo Framework and How Courts Apply It: the companion doctrine that polices result-focused claims.
• Provisional Patent Applications: What They Protect, What They Do Not, and How to Keep the Priority Date: why a thin first filing can lose its priority date.
• Prior Art Under the America Invents Act: Section 102, the On-Sale Bar, Public Use, and the One-Year Grace Period: what intervening art does when priority fails.
• Patent Post-Grant Proceedings: IPR, PGR, and Reexamination at the PTAB: which forum can hear which Section 112 argument.
• Inter Partes Review Strategy: Discretionary Denial, Estoppel, and the Parallel District Court Case: how statements in a petition can return in court.
• Biologics and the BPCIA: The Patent Dance, Biosimilar Approval, and Exclusivity: the regulatory setting for many antibody disputes.
• Hatch-Waxman Litigation: The Orange Book, Paragraph IV Certifications, and the 30-Month Stay: the setting for many small molecule and dosing claims.
• Federal Subject-Matter Jurisdiction: Arising Under, Diversity, Supplemental Claims, and the Defect That Never Waives: why these cases reach federal court and the Federal Circuit.
• Hiring From a Competitor: Inevitable Disclosure, Employee Raiding, and the New Employer's Trade Secret Exposure: the trade secret alternative and its own exposure.

Related Guides

• Drafting Patent Claims and Specifications That Survive Section 112: A Practical Guide: the step-by-step companion to this article.
• Preparing for a Markman Hearing: A Practical Guide: preparing the construction and indefiniteness record.
• Handling Objections, Coaching, and Misconduct at a Deposition: A Practical Guide: inventor testimony is often decisive here.

Related Checklists

• Section 112 Patent Review Checklist: A Practical Checklist: pre-filing and pre-litigation review tasks.
• Claim Construction Checklist: A Practical Checklist: testing a construction before you win it.

Related Toolkits

• Section 112 Toolkit: Claim Support Charts, Invalidity Contentions, and Expert Declarations: claim support charts, Wands worksheets, and declaration outlines.
• Section 101 Toolkit: Claim Drafting, Motions to Dismiss, and Expert Support: the eligibility counterpart for software claims.