University Of Tennessee Research Foundation v. Anthropic PBC

I. Executive Summary and Procedural Information

• Parties & Counsel:
  • Plaintiff: University of Tennessee Research Foundation (Tennessee)
  • Defendant: Anthropic PBC (Delaware)
  • Plaintiff’s Counsel: Farnan LLP
• Case Identification: 1:26-cv-00887, D. Del., 07/20/2026
• Venue Allegations: Venue is alleged to be proper in the District of Delaware because the Defendant is organized and existing under the laws of the State of Delaware.
• Core Dispute: Plaintiff alleges that Defendant’s Claude Code agentic software architecture infringes two patents related to the construction, configuration, and operation of neuroscience-inspired artificial neural networks.
• Technical Context: The technology relates to dynamic and adaptive artificial neural network architectures, a foundational area for creating more sophisticated, scalable, and efficient artificial intelligence systems.
• Key Procedural History: The complaint notes that in a separate copyright infringement case, a court found that Anthropic had downloaded and kept millions of pirated books, which the complaint uses to characterize Anthropic’s "cavalier approach to others' intellectual property rights."

Case Timeline

Date	Event
2013-10-16	Earliest Priority Date for ‘470 and ‘718 Patents
2018-07-10	‘470 Patent Issued
2018-10-09	‘718 Patent Issued
2023-09-01	YouTube Video: Looking Inside Neural Networks
2023-01-01	Publication: Towards Monosemanticity (Date estimated from "2023")
2024-05-21	Website Article: Mapping the Mind of a Large Language Model
2024-12-19	Website Article: Building effective agents
2025-06-23	Court Order in Bartz, et al. v. Anthropic PBC
2025-08-21	YouTube Video: Building and prototyping with Claude Code
2025-01-01	Publication: On the Biology of a Large Language Model (Date estimated from "2025")
2026-01-05	YouTube Video: Claude Agent SDK [Full Workshop]
2026-02-19	YouTube Video: Head of Claude Code: What Happens After Coding Is Solved
2026-03-04	YouTube Video: Building Claude Code with Boris Cherny
2026-03-15	Website Article: How we built Claude Code auto mode
2026-04-14	ArXiv Paper: Dive into Claude Code
2026-05-08	YouTube Video: The Expanding Toolkit and Memory and Dreaming
2026-05-20	YouTube Video: Claude Code Head Boris Cherny and Build a Proactive Agent Workflow
2026-07-20	Complaint Filing Date

II. Technology and Patent(s)-in-Suit Analysis

U.S. Patent No. 10,019,470, "Method and Apparatus for Constructing, Using and Reusing Components and Structures of an Artificial Neural Network," Issued Jul. 10, 2018

The Invention Explained

• Problem Addressed: The patent's background section notes that while biological neural networks possess desirable characteristics like robustness and adaptability, conventional artificial neural networks (ANNs) face challenges in achieving similar capabilities, particularly with fixed architectures. (Compl. ¶17; ’470 Patent, col. 1:10-16). Researchers encountered difficulties in designing ANNs that could adapt, reuse learned functionality, and improve performance without excessive manual intervention. (Compl. ¶18).
• The Patented Solution: The invention provides a method for constructing and evolving ANNs by identifying, extracting, and reusing "useful substructures" (Compl. ¶25; ’470 Patent, abstract). This process allows for dynamic adaptation of the network architecture, where successful or unsuccessful components can be implanted into or removed from an ANN to improve its performance on a given task. (’470 Patent, abstract; ’470 Patent, col. 16:30-41). The system may also incorporate "affective" networks that can regulate parameters like neuron firing thresholds, analogous to neurotransmitters in the brain. (’470 Patent, col. 17:55-65).
• Technical Importance: This approach to dynamically constructing and optimizing neural networks represented a move toward more biologically plausible, modular, and evolvable AI systems. (Compl. ¶27).

Key Claims at a Glance

• The complaint asserts at least dependent claim 10, which depends from independent claim 1. (Compl. ¶70).
• Essential elements of independent claim 1 ("A method of constructing a neuromorphic network...") include:
  • constructing a neuromorphic network for a process control application upon prediction of a component failure.
  • the network comprising a reconfigurable structure of components, including a two-dimensional array of neuron and synapse circuit elements.
  • the structure having at least one addressably configured input neuron and one output neuron at an edge of the array.
  • addressably configuring a synapse to connect two neuron circuit elements to form an initial neuromorphic network.
  • the synapse having a programmable delay or distance parameter.
  • a signal processor processing input signals.

U.S. Patent No. 10,095,718, "Method and Apparatus for Constructing a Dynamic Adaptive Neural Network Array (DANNA)," Issued Oct. 9, 2018

The Invention Explained

• Problem Addressed: The technical problems are identical to those described in the ’470 Patent, focusing on the limitations of conventional, fixed-architecture ANNs. (Compl. ¶17; ’718 Patent, col. 1:10-16).
• The Patented Solution: The invention describes a specific hardware-oriented architecture for a "dynamic adaptive neural network array" (DANNA). Its core is a configurable "circuit element" that can be programmed to function as either a neuron or a synapse. (’718 Patent, abstract). These elements are arranged in a multi-dimensional array, allowing for the construction of complex, interconnected networks where signals can propagate between neurons via these configurable synapse pathways. (’718 Patent, col. 29:1-10).
• Technical Importance: This patent provides a blueprint for a scalable and flexible hardware implementation of a neuromorphic computing system, enabling the creation of specialized processing arrays for AI tasks. (Compl. ¶26).

Key Claims at a Glance

• The complaint asserts at least independent claim 35. (Compl. ¶109).
• Essential elements of independent claim 35 ("A method for constructing a neuromorphic network...") include:
  • the network comprising a special purpose computer processor with an ANN configuration structure for addressably configuring a multi-dimensional array of circuit elements.
  • an interface and control structure for connecting the configured array to an external process.
  • the special purpose computer being responsive to input data from the external process for configuring the array.
  • at least one circuit element configured as an input neuron and at least one as an output neuron.
  • one to multiple circuit elements configured as a neuron or synapse forming a pathway between the input and output neuron.

III. The Accused Instrumentality

Product Identification

• The accused products are "Anthropic's Claude Code and its underlying agentic software architecture" (Compl. ¶36; Compl. ¶77).

Functionality and Market Context

• The complaint alleges the accused products constitute a "multi-agent system" built on a large language model (LLM) that functions as a "software-based neuromorphic network" (Compl. ¶79; Compl. ¶81). Key components are described by their alleged source code filenames, including a "background execution scheduling system (cronScheduler.ts)" and a "background memory consolidation engine (autoDream.ts)" (Compl. ¶38). This "dreaming" process is described as a batch asynchronous task that reviews agent sessions to find patterns and mistakes, producing "organized and up-to-date memory content" (Compl. ¶40).
• The architecture allegedly includes a "coordinator system" that configures "agent workers as discrete nodes (neurons)" and orchestrates tasks (Compl. ¶82; Compl. ¶90). The complaint alleges the system uses routing and orchestrator patterns where a central LLM breaks down tasks and delegates them to worker LLMs. (Compl. ¶90).

IV. Analysis of Infringement Allegations

10,019,470 Infringement Allegations

Claim Element (from Independent Claim 1)	Alleged Infringing Functionality	Complaint Citation	Patent Citation
a method of constructing a neuromorphic network... comprising a reconfigurable structure of components...	Anthropic's Claude Code and its agentic software architecture are alleged to be a "neuromorphic network" (Compl. ¶36). The structure is allegedly reconfigured when the Yoga Flexbox layout engine recomputes the layout tree. (Compl. ¶56).	¶36; ¶56	col. 21:1-5
the structure of components comprising... a two-dimensional array of neuron and synapse circuit elements...	The accused products allegedly map input and output boundary nodes (neurons) to a two-dimensional flexbox rendering grid. (Compl. ¶51).	¶51	col. 21:49-54
addressably configuring one synapse circuit element to connect one addressably configured neuron circuit element to another... to form an initial addressably configured array...	Synapses are allegedly implemented as parent-child tree node references and JSON-RPC channels. Layout nodes are connected via "insertChild" and "removeChild" to form an initial grid. (Compl. ¶55).	¶55	col. 22:5-10
the addressably configured synapse circuit element having a programmable delay or distance...	Programmable delay is allegedly implemented as "jittered backoff intervals, network connection retries, and tool timeouts." A programmable "distance" parameter allegedly maps to scope-level tracking constraints. (Compl. ¶64).	¶64	col. 22:1-4
a signal processor for processing input signals comprising one of a signal generated or sampled in the external automated process...	The complaint does not provide sufficient detail for analysis of the "signal processor" element, but alleges the system receives inputs from external components. (Compl. ¶47).	¶47	col. 21:61-64

• Identified Points of Contention:
  • Scope Question: A central issue will be whether software constructs within a large language model architecture, such as a "flexbox rendering grid" (Compl. ¶51) and inter-process communication channels (Compl. ¶55), meet the patent's definition of a "two-dimensional array of neuron and synapse circuit elements." The patent's language appears to contemplate a more explicit, spatially organized structure.
  • Technical Question: The complaint alleges that predicting "context window saturation or memory limit saturation" constitutes the claimed step of "controlling the external... device upon prediction of a component failure" (Compl. ¶66; '470 Patent, cl. 10). A court may need to determine if these software optimization routines are functionally equivalent to the "component failure" prediction described in the patent.

10,095,718 Infringement Allegations

Claim Element (from Independent Claim 35)	Alleged Infringing Functionality	Complaint Citation	Patent Citation
...constructing a neuromorphic network comprising at least one artificial neural network, the artificial neural network comprising a special purpose computer processor...	The accused products are alleged to perform a method of constructing a neuromorphic network, with the Claude LLM being the artificial neural network. (Compl. ¶80; Compl. ¶81).	¶80; ¶81	col. 40:50-53
...an artificial neural network configuration structure for addressably configuring the multi-dimensional array of addressably configurable circuit elements...	The "QueryEngine" message submission logic and the main loop in "query.ts" are alleged to act as the configuration structure, adjusting prompt contexts and token budgets to reconfigure the LLM. (Compl. ¶87).	¶87	col. 40:54-57
...an interface and control structure for connecting the addressably configured and connected multi-dimensional array... to the external process...	The interface is allegedly implemented in "toolExecution.ts" and "query.ts", which "hook the model's outputs... to external OS processes." (Compl. ¶87).	¶87	col. 40:58-61
...at least one circuit element of the addressably configured multi-dimensional array addressably configured as an input neuron...	User input prompts and text attachments are alleged to represent input neurons that initialize attention weights inside the transformer network. (Compl. ¶92).	¶92	col. 41:22-25
...at least one circuit element of the addressably configured multi-dimensional array addressably configured as an output neuron.	The output generated by the accused products (tokens, text block completions, tool use JSON schemas) is alleged to represent output neurons. (Compl. ¶94).	¶94	col. 41:31-34
...one to multiple circuit elements... forming a pathway between the input neuron... and the output neuron...	In "coordinator mode," the accused products allegedly configure pathways using multiple agents and tool execution loops. AgentTool (neurons) communicate via XML (synapses) to form pathways. (Compl. ¶97). The complaint also points to a diagram of an "evaluator-optimizer workflow" as an example of a feedback pathway. (Compl. ¶105). This diagram, from an Anthropic website article, shows a loop where an "LLM Call Generator" produces a solution that is then evaluated by an "LLM Call Evaluator," with feedback for rejected solutions. (Compl. p. 30).	¶97; ¶105	col. 41:38-44

• Identified Points of Contention:
  • Scope Question: A key dispute may arise over whether the high-dimensional vector/token space of an LLM (Compl. ¶81) constitutes a "multi-dimensional array of addressably configurable circuit elements" as required by the claim. The defense may argue that the patent's language implies a physically or structurally defined array, not an abstract mathematical space.
  • Technical Question: The complaint alleges that software modules like "QueryEngine" and control loops like "auto mode" constitute the claimed "special purpose computer processor" and its "configuration structure" (Compl. ¶85; Compl. ¶87; Compl. ¶88). The court may examine whether these general software routines running on general-purpose hardware meet the claim requirement of a "special purpose" processor and its associated structures.

V. Key Claim Terms for Construction

U.S. Patent No. 10,019,470

• The Term: "central pattern generator"
• Context and Importance: This term is the central component of the asserted method claim. The complaint alleges that Anthropic’s background scheduling ("cronScheduler.ts") and memory consolidation ("autoDream.ts") systems together form a "central pattern generator" (Compl. ¶38). The case's outcome for this patent likely hinges on whether these software processes can be defined as a CPG.
• Intrinsic Evidence for Interpretation:
  • Evidence for a Broader Interpretation: The patent does not appear to provide an explicit definition in the specification. A party could argue that any system producing "cyclic or regenerative behavior" (language from claim 8, from which claim 10 depends) meets the functional definition, allowing software loops to qualify.
  • Evidence for a Narrower Interpretation: The term "central pattern generator" has a specific meaning in neuroscience related to rhythmic motor patterns. The patent's focus on "neuroscience-inspired" design may support an interpretation tied to this specific biological analog, which a generic software scheduler may not satisfy. (’470 Patent, title; ’470 Patent, abstract).

U.S. Patent No. 10,095,718

• The Term: "circuit element"
• Context and Importance: Practitioners may focus on this term because the infringement theory depends on mapping abstract software concepts (e.g., LLM features, agent workers) to this term. The complaint alleges that "agent workers" are "discrete nodes (neurons)" (Compl. ¶82) and that LLM "features" are "neuron[s]" (Compl. ¶83). The viability of the infringement case rests on whether the term "circuit element" can be construed to encompass these software abstractions.
• Intrinsic Evidence for Interpretation:
  • Evidence for a Broader Interpretation: The patent states that an element may be "analog or digital" and that the DANNA may be implemented in software, suggesting a non-physical interpretation is possible. (’718 Patent, abstract).
  • Evidence for a Narrower Interpretation: The detailed description and figures provide specific hardware implementations. Figure 9A, for example, shows a detailed schematic for a single array element with components like a MUX, latch, and comparator, suggesting a structural, hardware-based definition. (’718 Patent, FIG. 9A). The abstract also describes the element as part of a "multi-dimensional dynamic adaptive neural network array," which could imply a physical or grid-like structure.

VI. Other Allegations

• Indirect Infringement: The complaint alleges that Anthropic actively induces infringement by providing "documentation and training materials" that instruct customers and end-users to use the accused products in an infringing manner. (Compl. ¶73; Compl. ¶112).
• Willful Infringement: The complaint alleges knowledge of the patents "since at least service of this Complaint or shortly thereafter" (Compl. ¶72; Compl. ¶111). While it does not use the word "willful," it requests that the court find the case "exceptional" and award attorneys' fees under 35 U.S.C. § 285. (Compl. p. 33).

VII. Analyst’s Conclusion: Key Questions for the Case

1. A core issue will be one of definitional scope: can terms rooted in neuroscience and hardware design, such as "central pattern generator" and "circuit element", be construed broadly enough to cover the high-level, abstract software components of a modern, LLM-based agentic architecture? The resolution of this question will likely determine whether the patents, filed in 2014, can read on technology developed nearly a decade later.
2. A second central question will be one of technical mapping: does the complaint provide sufficient evidence that Anthropic's software systems functionally operate in the specific manner claimed by the patents? For example, does the "dreaming" process for memory consolidation (Compl. ¶40) perform the same steps as the claimed method for reconfiguring a network in response to a predicted "component failure"? (Compl. ¶66). The case may turn on whether the plaintiff can prove an operational correspondence beyond a high-level analogy.