The two-step framework
Every software claim is evaluated under the two-step Mayo/Alice framework as it is actually applied by the USPTO and Federal Circuit in 2026:
Step 2A Prong One
Does the claim RECITE a judicial exception (abstract idea, mental process, math)? Naming a formula recites it. Describing the operation functionally only INVOLVES it. Per the USPTO August 2025 memo, large-scale neural-network operations are categorically outside the mental-process category.
Step 2A Prong Two
Is the exception INTEGRATED into a practical application? This is where most software claims live or die. The integration must be a real-world technical action or a technical improvement to computer/network/hardware functioning - not just "applying" the exception in some field.
Step 2B
If not integrated under Prong Two, is there an "inventive concept" - significantly more than well-understood, routine, conventional activity? Berkheimer makes this a question of fact. The unconventional element must be IN the claim, not just in the spec.
Three winning templates we look for
We score patents against three claim shapes that have actually survived 101 challenges at the Federal Circuit or won at the USPTO Appeals Review Panel. When your claim structurally matches one of these templates, you can cite the precedent directly during prosecution.
Desjardins
Training/architecture improvement
Improvements to ML training or architecture that solve a NAMED technical failure mode (catastrophic forgetting, mode collapse, KV-cache pressure, gradient underflow, cold-start drift).
The USPTO Appeals Review Panel made Desjardins precedential on November 4, 2025, and MPEP 2106 was updated December 5, 2025 to incorporate it. Training-regime improvements with named failure modes are now a documented safe harbor.
Required claim elements
- Name the technical problem in the claim itself
- Recite the specific architectural or training step that solves it
- State the causal link with "thereby" or "so that"
- Tie to a concrete technical action or measurable outcome
Contour
Coordinated system
Coordinated systems where the coordination itself is the technical effect, even when the individual components are conventional.
Contour IP v. GoPro (Fed. Cir. Sept. 2024) was the first Fed Cir reversal of software-patent ineligibility since Cosmokey (2021). Specific, coordinated, technologically concrete means produce a practical application that survives Alice Step 2A Prong Two.
Required claim elements
- A first subsystem with a specific technical operation
- A second subsystem operating in parallel or coordination with the first
- A coordination/synchronization component
- A named technical effect that neither subsystem alone provides
Example 47
AI tied to a downstream technical action
AI or ML output that triggers a concrete downstream action in a technical field - dropping packets, reallocating GPU memory, rerouting traffic, administering a drug dose, adjusting hardware state.
USPTO 2024 AI SME Update (89 FR 58128) Example 47 specifically allowed an ANN claim that added "responsive to detection, drop the anomalous packets" to an otherwise-ineligible classification step. The downstream technical action is what integrates the abstract idea into a practical application.
Required claim elements
- Receive a technical input
- Process through a specific ML architecture (named structural detail)
- Produce a technical output
- Take a concrete technical action in response
Three kill zones we flag
When we see these patterns, we flag them with low Alice scores and recommend reframing before filing. Filing into any of these is a near-certain rejection.
Recentive
"Apply ML to X" claims
Claims that "do no more than apply established methods of machine learning to a new data environment" - recommendation engines, pricing, matching, scheduling, prediction with generic ML.
Recentive Analytics v. Fox Corp. (Fed. Cir. April 2025, cert denied December 2025) is the ceiling case. Iterative training, dynamic adjustment, and improved accuracy are NOT technical improvements - they are within the nature of machine learning.
Business method
Business process dressed in software
Computer-implemented methods of human business tasks - KYC workflows, AML processes, trading logic, ops procedures.
Alice itself was a business method. Of 22 precedential 101 decisions in 2024, claims were held eligible in only 1 case (Contour). Business methods need a genuinely new technical substrate (novel cryptographic primitive, new protocol, new data structure) - the business rule on top is incidental.
Mental process
Could be done in a human mind
Claims reciting "analyzing", "comparing", "determining" at a conceptual level that a person could in principle do mentally.
Per the USPTO August 4, 2025 memo, the mental-process category has limits - operations that cannot practically be performed in a human mind (NN with >100 parameters, backprop at scale, multi-dim tensor ops) are categorically outside it. But naïve "analyze X to determine Y" framings still get rejected.
Authoritative sources
The doctrine module is built on these cases and guidance docs. Each is citeable in office action responses. We update this list whenever new precedent shifts the framework.
Alice Corp. Pty. Ltd. v. CLS Bank Int'l, 573 U.S. 208 (2014)
Established the two-step Mayo/Alice framework for patent eligibility under 35 USC 101. Killed an entire category of business-method software patents.
U.S. Supreme Court - decided 2014-06-19
Mayo Collaborative Servs. v. Prometheus Labs., 566 U.S. 66 (2012)
Established the "inventive concept" requirement at Step 2B of what became the Alice framework.
U.S. Supreme Court - decided 2012-03-20
Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025)
Claims that "do no more than apply established methods of machine learning to a new data environment" are patent ineligible. Iterative training and dynamic adjustment are not technical improvements - they are within the nature of ML. SCOTUS denied cert December 2025.
Federal Circuit - decided 2025-04-18
Contour IP Holding LLC v. GoPro, Inc., 113 F.4th 1373 (Fed. Cir. 2024)
Specific, coordinated, technologically concrete means constitute a practical application even where individual components are conventional. First Fed Cir software-patent reversal of ineligibility since Cosmokey (2021).
Federal Circuit - decided 2024-09-09
Ex parte Desjardins, Appeal 2024-000567 (P.T.A.B. Sept. 26, 2025) (precedential Nov. 4, 2025)
Improvements to ML training that solve named technical failure modes (here: catastrophic forgetting via weight-importance regularization) are eligible because they improve how the ML system operates. MPEP 2106 updated December 5, 2025 to incorporate.
USPTO Appeals Review Panel - decided 2025-11-04
In re McFadden (Fed. Cir. Sept. 2025)
Reversed PTAB's "software per se" 101 rejection by reframing under 112(f). Signals Fed Circ disfavor of pure-abstraction rejections when a 112(f) means-plus-function construction is available.
Federal Circuit - decided 2025-09-15
Berkheimer v. HP Inc., 881 F.3d 1360 (Fed. Cir. 2018)
Whether a claim element is "well-understood, routine, conventional" is a question of fact, not law. Preserves factual disputes about unconventionality at Alice Step 2B. Still binding; some district courts now sending Step 2 to juries.
Federal Circuit - decided 2018-02-08
2024 Guidance Update on Patent Subject Matter Eligibility, Including on Artificial Intelligence, 89 Fed. Reg. 58128 (July 17, 2024)
Sharpened Step 2A Prong Two analysis for AI inventions via Examples 47-49. Generic "apply ML to domain X" is ineligible; specific architectural/training/pipeline detail producing a concrete downstream technical action is eligible.
USPTO - decided 2024-07-17
USPTO Memorandum, "Reminders on Evaluating Subject Matter Eligibility" (Aug. 4, 2025)
Materially raised the bar for examiner 101 rejections. Reject only when "more likely than not" ineligible. Mental-process grouping has limits - operations that cannot practically be performed in a human mind are categorically outside it.
USPTO - decided 2025-08-04
USPTO Subject Matter Eligibility Declaration (SMED) Memos (Dec. 4, 2025)
New Rule 132 evidentiary pathway. Standalone declarations from inventor, co-worker, or independent expert can assert as fact: (a) steps cannot be practically performed mentally, (b) invention materially improves computer/technical field, (c) unconventional arrangement of conventional components.
USPTO - decided 2025-12-04
Patent Eligibility Restoration Act of 2025, S.1546, 119th Cong. (2025)
Would eliminate all judicial exceptions to patent eligibility and replace them with narrow statutory exclusions (math formulas per se, mental processes done entirely in a human mind, unmodified natural materials, "non-technological" economic/social processes). Software explicitly eligible if it requires a machine or transformation. PENDING as of April 2026 - no floor vote.
U.S. Senate (pending) - decided 2025-05-01
What we publish vs. what stays proprietary
We publish the doctrine, the case law, the framework, and the philosophy. The implementation details - specific prompts, scoring weights, gating thresholds, anchor extraction techniques, redraft prompt patterns - stay internal. This is how we keep the methodology citeable while preserving Patentable's edge.
If you want to see AliceScore applied to your invention, sign up at patentable.com/get-started. If you're a patent attorney evaluating the methodology, contact us about the attorney-grade dashboard.