Prosecution Insights
Last updated: November 20, 2026
Application No. 18/669,366

Policy-Based Machine Learning Monitoring Using Intermediate Results

Non-Final OA §101§103§112§DOUBLEPATENT
Filed
May 20, 2024
Examiner
KWON, JUN
Art Unit
2127
Tech Center
2100 — Computer Architecture & Software
Assignee
HiddenLayer, Inc.
OA Round
6 (Non-Final)
42%
Grant Probability
Moderate
6-7
OA Rounds
2y 2m
Est. Remaining
83%
With Interview

Examiner Intelligence

Grants 42% of resolved cases
42%
Career Allowance Rate
35 granted / 83 resolved
-12.8% vs TC avg
Strong +40% interview lift
Without
With
+40.5%
Interview Lift
resolved cases with interview
Typical timeline
4y 8m
Avg Prosecution
26 currently pending
Career history
127
Total Applications
across all art units

Statute-Specific Performance

§101
29.0%
-11.0% vs TC avg
§103
45.7%
+5.7% vs TC avg
§102
9.5%
-30.5% vs TC avg
§112
14.9%
-25.1% vs TC avg
Black line = Tech Center average estimate • Based on career data from 83 resolved cases

Office Action

§101 §103 §112 §DOUBLEPATENT
Detailed Action This Office Action is in response to the remarks entered on 07/23/2026. Amended claims 1, 8, 20-21, 28-29 and 32 have been entered. Claims 1, 3-15 and 17-32 are presently pending. Notice of Pre-AIA or AIA Status The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA . Claim Rejections - 35 USC § 112 The following is a quotation of 35 U.S.C. 112(d): (d) REFERENCE IN DEPENDENT FORMS.—Subject to subsection (e), a claim in dependent form shall contain a reference to a claim previously set forth and then specify a further limitation of the subject matter claimed. A claim in dependent form shall be construed to incorporate by reference all the limitations of the claim to which it refers. The following is a quotation of pre-AIA 35 U.S.C. 112, fourth paragraph: Subject to the following paragraph [i.e., the fifth paragraph of pre-AIA 35 U.S.C. 112], a claim in dependent form shall contain a reference to a claim previously set forth and then specify a further limitation of the subject matter claimed. A claim in dependent form shall be construed to incorporate by reference all the limitations of the claim to which it refers. Claim 15 is rejected under 35 U.S.C. 112(d) or pre-AIA 35 U.S.C. 112, 4th paragraph, as being of improper dependent form for failing to further limit the subject matter of the claim upon which it depends, or for failing to include all the limitations of the claim upon which it depends. Claim 15 fails to further limit the subject matter of the claim 1 which it depends, because claim 1 already recites “whether the prompt seeks to cause the GenAI model to behave in an undesired manner, the one or more prohibited actions including seeking one or more of: a response in a non-approved spoken or written language, computer code, sensitive information, or a response to an encrypted prompt;” while claim 15 recites “The method of claim 1, wherein the one or more prohibited actions include seeking one or more of: a response in a non-approved spoken or written language, computer code, sensitive information, a response to an encrypted prompt, a prompt containing two or modalities, or a prompt in a first modality obfuscating information in a second modality” Applicant may cancel the claim(s), amend the claim(s) to place the claim(s) in proper dependent form, rewrite the claim(s) in independent form, or present a sufficient showing that the dependent claim(s) complies with the statutory requirements. Amended claims were received on 07/23/2026. 35 U.S.C. 112(b) rejections have been withdrawn. Claim Rejections - 35 USC § 101 Amended claim 32 was received on 07/23/2026. 35 U.S.C. 101 rejections have been withdrawn. Claim Rejections - 35 USC § 103 The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action: A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made. Claims 1, 3 and 14-15 are rejected under 35 U.S.C. 103 as being unpatentable over Goswami (US 20190238568 A1, hereinafter ‘Goswami’) in view of Shi et al. (“PL-Transformer: a POS-aware and layer ensemble transformer for text classification”, Published online 2022, hereinafter ‘Shi’) and further in view of Xu et al. (“Uncovering Safety Risks in Open-source LLMs through Concept Activation Vector”, 18 Apr 2024, hereinafter ‘Xu’). Regarding claim 1, Goswami teaches: A computer-implemented method comprising: receiving, by a computing system, data [Goswami, 0108, Fig. 9] discloses operating the trained DNN 820 on new data) capturing, during inference of the [Goswami, 0108, Fig. 9] The new network activation data 920 which is the means for the intermediate layers of the DNN 820 is input to the SVL classifier 870. And then, the SVL classifier 870 which is trained to detect distortion in the data is used to classify the data. The DNN is interpreted as the GenAI model, and it is more clearly disclosed by the secondary reference Shi) determining, based on the intermediate result and using a machine learning-based classifier trained using that maps intermediate results to one or more prohibited actions by the GenAI model, whether the prompt seeks to cause the GenAI model to behave in an undesired manner; ([Goswami, Fig. 9 and 0108] discloses generating an output 930 indicating whether or not the SVM classifier 870 has determined that the new input data 910 is distorted or undistorted. [0110-0111] collectively discloses the training process of the SVM classifier which performs selective dropout logic to selectively dropout filters (i.e., policy that maps result to action). [0122] discloses in response to the output of the SVM classifier, a responsive operation may be performed such as alerting contingent systems or operating or shutting down the downstream computing system) and providing, by the computing system over a network, data characterizing the determination to a consuming application or process, the consuming application or process (i) allowing an output of the GenAI model to be provided to a requestor when it is determined that the prompt does not seek to cause the GenAI model to behave in an undesired manner, and (ii) initiates at least one remediation action preventing the GenAI model from behaving in an undesired manner when it is determined that the prompt seeks to cause the GenAI model to behave in an undesired manner. ([Goswami, Fig. 9 and 0108] discloses generating an output 930 indicating whether or not the SVM classifier 870 has determined that the new input data 910 is distorted or undistorted. [0122] discloses if the input image data is determined to not be adversarial, the image processing is performed (i.e., allowing the output to be provided). Otherwise, in response to the output of the SVM classifier, a responsive operation may be performed such as alerting contingent systems or operating or shutting down the downstream computing system, or dropout the adversarial image data (i.e., remediation action)) Goswami does not specifically disclose: receiving, by a computing system, data characterizing a prompt for ingestion by a generative artificial intelligence (GenAI) model, the GenAI model having an input layer, an output layer, and a plurality of intermediate transformer layers positioned immediately after the input layer and immediately before the output layer; inputting, by the computing system, the prompt into the GenAI model; capturing, during inference of the GenAI model and using runtime hooks applied to one or more transformer layers of the GenAI model, an intermediate result comprising activations in residual streams derived from one or more of the intermediate transformer layers of the GenAI model; determining, based on the intermediate result and using a machine learning-based classifier trained using data set generated according to a policy … the one or more prohibited actions including seeking one or more of: a response in a non-approved spoken or written language, computer code, sensitive information, or a responsive to an encrypted prompt; Shi teaches: receiving, by a computing system, data characterizing a prompt for ingestion by a generative artificial intelligence (GenAI) model, the GenAI model having an input layer, an output layer, and a plurality of intermediate transformer layers positioned immediately after the input layer and immediately before the output layer; ([Shi, page 1974, Fig. 1] and [Shi, page 1975, left col, 3.2 Embedding layer, lines 1-4] show that the embedding layer obtains the input sequence of sentence representation, which are stream input. Fig.1 further shows that the model having an embedding layer (i.e., input layer), an output layer (i.e., softmax layer), and a plurality of intermediate transformer layers (i.e., PL transformer layer with a plurality of C-Encoder layers). The transformer layer is the GenAI model) inputting, by the computing system, the prompt into the GenAI model; ([Shi, page 1974, Fig. 1] and [Shi, page 1975, left col, 3.2 Embedding layer, lines 1-4] shows that the embedding layer obtains the input sequence of sentence representation, which are stream input) capturing, during inference of the GenAI model and using runtime hooks applied to one or more transformer layers of the GenAI model, an intermediate result comprising activations in residual streams derived from one or more of the intermediate transformer layers of the GenAI model; ([Shi, page 1975, left col, 3.2 Embedding layer, lines 1-4] shows that the embedding layer obtains the input sequence of sentence representation, which are stream input. [Shi, page 1974, Fig. 1] and [page 1976, left col, 3.3.2 Layer ensemble mechanisms, line 1 – right col, line 13] collectively discloses capturing PL-Transformer intermediate layers (C-Encoder Layers, which are interpreted as transformer layers) using hooks, and then ensembles the captured layer outputs h1(1)h2(1)…hn(1) and h1(2)h2(2)…hn(2) … h1(l)…hn(l). The transformer layer outputs are combined and sent to Linear Layer and Softmax to generate the final text classification output) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami and Shi to use a method of capturing an intermediate result of the transformer model of Shi to implement the anomaly detection method of Goswami. The suggestion and/or motivation for doing so is to improve the performance of the model by providing more minority class data from the trained latent space. However, Goswami in view of Shi do not specifically disclose: determining, based on the intermediate result and using a machine learning-based classifier trained using data set generated according to a policy … the one or more prohibited actions including seeking one or more of: a response in a non-approved spoken or written language, computer code, sensitive information, or a responsive to an encrypted prompt; Xu teaches: determining, based on the intermediate result and using a machine learning-based classifier trained using data set generated according to a policy … the one or more prohibited actions including seeking one or more of: a response in a non-approved spoken or written language, computer code, sensitive information, or a responsive to an encrypted prompt; ([Xu, page 4, 3.2 LLMs Attack Using SCAV] The LLM is composed of L Transformer layers. CAV are extracted by processing collected datasets. [page 6, 5.1 Experimental Setup, lines 1-4] A set of positive dataset contains harmful instructions, and the negative dataset contains harmless instructions that LLMs should not give responses, which indicates that both dataset are generated according to a policy. [page 4, last para, SCAV Extraction, lines 1-5] finds vector v in the space of activations of each layer k. The difference in activation vectors triggered by benign versus malicious instructions are calculated. [page 5, lines 1-5] A binary linear classifier is trained to distinguish between the layer activation outputs of the positive example and the negative example (i.e., if the input example causes prohibited actions). [page 12, A.3 HE, lines 1-33] The method is evaluated based on if the model generated harmful text) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi and Xu to use the method of determining, based on the intermediate result and using a classifier trained using data set generated according to a policy, the one or more prohibited actions of Xu to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the accuracy of the existing LLM prompt evaluation system [Xu, Abstract, lines 13-17]. Regarding claim 3, Goswami in view of Shi teaches: wherein the GenAI model comprises a plurality of transformer layers and the intermediate result comprises per-layer activations captured from residual streams generated by one or more of the transformer layers via the runtime hooks during a forward pass. ([Shi, page 1974, Fig. 1] and [page 1976, left col, 3.3.2 Layer ensemble mechanisms, line 1 – right col, line 13] collectively discloses capturing PL-Transformer intermediate layers (C-Encoder Layers, which are interpreted as transformer layers) using hooks, and then ensembles the captured layer outputs h1(1)h2(1)…hn(1) and h1(2)h2(2)…hn(2) … h1(l)…hn(l). The transformer layer outputs are combined and sent to Linear Layer and Softmax to generate the final text classification output) Regarding claim 14, Goswami teaches: wherein a policy maps intermediate residual stream activations to one or more prohibited actions ([Goswami, 0108] The SVL classifiers 870 is trained to detect adversarial input image data. [Goswami, 0113 and 0115] collectively disclose that the method is not limited to image data and the input data may be textual data (residual stream)). Regarding claim 15, Goswami in view of Shi and further in view of Xu teaches: The method of claim 1, wherein the one or more prohibited actions include seeking one or more of: a response in a non-approved spoken or written language, computer code, sensitive information, a response to an encrypted prompt, a prompt containing two or modalities, or a prompt in a first modality obfuscating information in a second modality ([Xu, page 4, 3.2 LLMs Attack Using SCAV] The LLM is composed of L Transformer layers. CAV are extracted by processing collected datasets. [page 4, last para, SCAV Extraction, lines 1-5] finds vector v in the space of activations of each layer k. The difference in activation vectors triggered by benign versus malicious instructions are calculated. [page 5, lines 1-5] A binary linear classifier is trained to distinguish between the layer activation outputs of the positive example and the negative example (i.e., if the input example causes prohibited actions). [page 12, A.3 HE, lines 1-33] The method is evaluated based on if the model generated harmful text) Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over Goswami in view of Shi in view of Xu and further in view of Pardeshi (US 20210407051 A1). Regarding claim 4, Goswami in view of Shi teaches the method of claim 1. Goswami in view of Shi does not specifically disclose: wherein the GenAI model comprises a mixture of experts (MoE) model and the intermediate result comprises outputs from at least a subset of experts in the MoE model. Pardeshi teaches: wherein the GenAI model comprises a mixture of experts (MoE) model and the intermediate result comprises outputs from at least a subset of experts in the MoE model ([Pardeshi, 0050] The GAN contains a different trained variational autoencoder VAE for each of a set of object classes, and these VAEs, which corresponds to the subject of experts, are considered experts for different objects classes in a mixture of experts-based approach. [Pardeshi, 0051] An appropriate VAE is selected based on outputs of the VAEs (intermediate result).). Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi and Pardeshi to use the mixture of experts (MoE) model of Pardeshi to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the performance of the machine learning system by utilizing machine learning models specialized for different tasks (i.e., different types of subjects, responses, etc.). Claims 5-7 are rejected under 35 U.S.C. 103 as being unpatentable over Goswami in view of Shi in view of Xu and further in view of Kuan (US 20230027149 A1, hereinafter ‘Kuan’). Regarding claim 5, Goswami in view of Shi and further in view of Xu teaches the method of claim 1. Goswami in view of Shi and further in view of Xu do not specifically disclose wherein the consuming application or process flags the prompt based on the determination. Kuan teaches: wherein the consuming application or process flags the prompt based on the determination. ([Kuan, 0076] Malicious data 704 is identified by the anomaly prediction system 114. The identification process corresponds to the ‘flags’, as the flags provided to the prompt merely identifies which data contains anomaly) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi and Kuan to use the method of flagging the prompt based on the determination of Kuan to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the performance of the machine learning system by allowing the machine learning model to quickly determine which input data is determined to contain malicious data. Regarding claim 6, Goswami in view of Shi and further in view of Xu teaches the method of claim 1. Goswami in view of Shi and further in view of Xu do not specifically disclose: wherein the consuming application or process automatically blocks, at a network gateway, an internet protocol (IP) address of a requester of the prompt based on the determination. Kuan teaches: wherein the consuming application or process automatically blocks, at a network gateway, an internet protocol (IP) address of a requester of the prompt based on the determination. ([Kuan, 0080] IP addresses associated with the malicious data 704 identified by the anomaly prediction system 114 is blocked by the control action 708 performed by the system) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi, Xu and Kuan to use the method of blocking the IP address of a requester based on the determination of Kuan to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the security of the machine learning system by blocking the prompt from the potentially malicious IP address. Regarding claim 7, Goswami in view of Shi in view of Xu and further in view of Kuan teaches: The method of claim 1, wherein the consuming application or process causes subsequent prompts from an internet protocol (IP) address of a requester of the prompt to be programmatically modified based on the determination and causes the modified prompt to be ingested by the GenAI model. ([Kuan, 0076] Malicious data 704 is identified by the anomaly prediction system 114. [Kuan, 0088] The system removes the corresponding malicious data 704 from the queue, and not issue any control action against the malicious data again. Removal of the malicious data is interpreted as the process of modifying prompt based on the determination.) Claims 8 and 12-13 are rejected under 35 U.S.C. 103 as being unpatentable over Goswami in view of Shi in view of Xu and further in view of Elthakeb et al. (“Divide and Conquer: Leveraging Intermediate Feature Representations for Quantized Training of Neural Networks”, 2020, hereinafter ‘Elthakeb’). Regarding claim 8, Goswami in view of Shi and further in view of Xu teaches: wherein the intermediate result is captured by a proxy of the GenAI model … transformer layers in the GenAI model ([Shi, page 1975, left col, 3.2 Embedding layer, lines 1-4] shows that the embedding layer obtains the input sequence of sentence representation, which are stream input. [Shi, page 1974, Fig. 1] and [page 1976, left col, 3.3.2 Layer ensemble mechanisms, line 1 – right col, line 13] collectively discloses capturing PL-Transformer intermediate layers (C-Encoder Layers, which are interpreted as transformer layers) using hooks, and then ensembles the captured layer outputs h1(1)h2(1)…hn(1) and h1(2)h2(2)…hn(2) … h1(l)…hn(l). The transformer layer outputs are combined and sent to Linear Layer and Softmax to generate the final text classification output) However, Goswami in view of Shi and further in view of Xu do not specifically disclose: wherein the intermediate result is captured by a proxy of the GenAI model that executes on distinct computing hardware and mirrors a structure of the transformer layers in the GenAI model, the proxy comprising a model that mimics some or all of the behavior of the GenAI model and whose output or intermediate result is used with the classifier. Elthakeb teaches: wherein the , the proxy comprising a model that mimics some or all of the behavior of the GenAI model and whose output or intermediate result is used with the classifier. ([Elthakeb, page 3, 2.1. Matching Activations for Intermediate Layers, lines 1-23], [Figure 3] collectively discloses that the quantized models have the same topology and the produce intermediate streams compatible with the full precision network (the GenAI model). [page 3, 2.2. Splitting, Training and Merging, lines 1-4] discloses that the models are trained in isolation and in parallel. [page 8, right col, 4.2. Lipschitz Constants in Classification Networks, lines 1-22] discloses the classifier which uses intermediate streams to classify the data) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi, Xu and Elthakeb to use the method of quantizing the generative model of Elthakeb to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to increase the performance of the system, as quantizing a model makes running inference faster and cheaper by reducing the file size of model weight files. Regarding claim 12, Goswami in view of Shi in view of Xu and further in view of Elthakeb teaches: The method of claim 8, wherein the proxy is a quantized version of the GenAI model having a same topology and producing intermediate residual streams compatible with the GenAI model. ([Elthakeb, page 3, 2.1. Matching Activations for Intermediate Layers, lines 1-23], [Figure 3] collectively discloses that the quantized models have the same topology and the produce intermediate streams compatible with the full precision network (the GenAI model)) Regarding claim 13, Goswami in view of Shi in view of Xu and further in view of Elthakeb teaches: The method of claim 1 further comprising: quantizing the GenAI model prior to capturing the intermediate result. ([Elthakeb, page 3, 2.1. Matching Activations for Intermediate Layers, lines 1-23], [Figure 3] collectively discloses that the quantized models have the same topology and the produce intermediate streams compatible with the full precision network (the GenAI model)) Claims 9-10 are rejected under 35 U.S.C. 103 as being unpatentable over Goswami in view of Shi in view of Xu in view of Elthakeb and further in view of Spalt (US 20200409323 A1). Regarding claim 9, Goswami in view of Shi in view of Xu and further in view of Elthakeb teaches the method of claim 8. Goswami in view of Shi in view of Xu and further in view of Elthakeb does not specifically disclose: wherein the consuming application or process prevents the prompt from being input into the GenAI model based on the determination. Spalt teaches: wherein the consuming application or process prevents the prompt from being input into the GenAI model based on the determination ([Spalt, 0100] The matrix generator determines that whether the input matrix can be input to the generator machine learning model. Based on the result of the determination, incomplete input matrix (anomalous prompt) is blocked.). Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi, Elthakeb, and Spalt to use the method of blocking the prompt based on the determination of Spalt to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the security of the machine learning system by blocking the prompt from the potentially malicious IP address. Regarding claim 10, Goswami in view of Shi in view of Xu and further in view of Elthakeb and further in view of Spalt teaches: wherein the consuming application or process allows the prompt to be input into the GenAI model based on the determination ([Spalt, 0100] The matrix generator determines that whether the input matrix can be input to the generator machine learning model. Based on the result of the determination, incomplete input matrix (anomalous prompt) is blocked. However, sometimes the input matrix generator 212 can pass the incomplete input matrix to the machine learning model generator 214 (allows the prompt to be input into the generative AI model).). Claim 11 is rejected under 35 U.S.C. 103 as being unpatentable over Goswami in view of Shi in view of Xu in view of Elthakeb and further in view of Kuan. Regarding claim 11, Goswami in view of Shi in view of Xu and further in view of Elthakeb does not specifically disclose: wherein the consuming application or process modifies the prompt based on the determination and causes the modified prompt to be ingested by the GenAI model Kuan teaches: wherein the consuming application or process modifies the prompt based on the determination and causes the modified prompt to be ingested by the GenAI model ([Kuan, 0076] Malicious data 704 is identified by the anomaly prediction system 114. [Kuan, 0088] The system removes the corresponding malicious data 704 from the queue, and not issue any control action against the malicious data again. Removal of the malicious data is interpreted as the process of modifying prompt based on the determination.). Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi, Xu, Elthakeb and Kuan to use the method of blocking the IP address of a requester and modifies the prompt based on the determination of Kuan to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the security of the machine learning system by blocking the prompt from the potentially malicious IP address. Claims 17-19 are rejected under 35 U.S.C. 103 as being unpatentable over Goswami in view of Shi in view of Xu and further in view of Brodie (US 20240005690 A1). Regarding claim 17, Goswami in view of Shi and further in view of Xu teaches the method of claim 1. Goswami in view of Shi and further in view of Xu do not specifically disclose wherein the GenAI model is a state-based text analysis model. Brodie teaches: wherein the GenAI model is a state-based text analysis model ([Brodie, 0035] The neural network can include a deep neural network, a convolutional neural network, a recurrent neural network (e.g., an LSTM), a graph neural network, a transformer neural network, or a generative adversarial neural network. Upon training, such a neural network may become a large language model. The instant specification [0055] explains that the state-based text analysis models include LSTMs) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi, Xu and Brodie to use a large language model which includes LSTMs of Brodie to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the performance of the model in long-term dependencies in sequential prompt data. Regarding claim 18, Goswami in view of Shi in view of Xu and further in view of Brodie teaches: wherein the state-based text analysis model comprises: a large language model. ([Brodie, 0035] The neural network can include a deep neural network, a convolutional neural network, a recurrent neural network (e.g., an LSTM), a graph neural network, a transformer neural network, or a generative adversarial neural network. Upon training, such a neural network may become a large language model. The instant specification [0055] explains that the state-based text analysis models include LSTMs) Regarding claim 19, Goswami in view of Shi in view of Xu and further in view of Brodie teaches: wherein the state-based text analysis model comprises: a long short-term memory model (LSTM) ([Brodie, 0035] The neural network can include a deep neural network, a convolutional neural network, a recurrent neural network (e.g., an LSTM), a graph neural network, a transformer neural network, or a generative adversarial neural network. Upon training, such a neural network may become a large language model. The instant specification [0055] explains that the state-based text analysis models include LSTMs). Claims 20 is rejected under 35 U.S.C. 103 as being unpatentable over Goswami in view of Shi in view of O’CONNOR et al. (US 20220092404 A1, hereinafter ‘O’CONNOR’) and further in view of Elthakeb. Regarding claim 20, Goswami teaches: A computer-implemented method comprising: receiving, by a computing system, data c[Goswami, 0108, Fig. 9] The new network activation data 920 which is the means for the intermediate layers of the DNN 820 is generated in response to the new input data and input to the SVL classifier 870. And then, the SVL classifier 870 which is trained to detect distortion in the data is used to classify the data) capturing, during inference of the , the capturing comprising applying the runtime hooks to each of the intermediate ; ([Goswami, 0105] The forward-pass is a common neural network operation performed to generate inferences. The paragraph and Fig. 8 shows generating network activation data 830 by processing original input data 810 (forward pass). [0108, Fig. 9] The new network activation data 920 which is the means for the intermediate layers of the DNN 820 is input to the SVL classifier 870. And then, the SVL classifier 870 which is trained to detect distortion in the data is used to classify the data. [Claims 1 and 6] indicates that the output value from the intermediate layers are vectors (array of numerical data)) determining, using a machine learning-based classifier trained using a policy that maps intermediate residual stream activations to acceptable and undesired actions and based on the intermediate result, whether the prompt elicits undesired actions by the GenAI model, the classifier being trained using a data set generated according to a policy which maps acceptable and undesired actions to intermediate results; and ([Goswami, Fig. 9 and 0108] discloses generating an output 930 indicating whether or not the SVM classifier 870 has determined that the new input data 910 is distorted or undistorted. [0122] discloses in response to the output of the SVM classifier, a responsive operation may be performed such as alerting contingent systems or operating or shutting down the downstream computing system) returning, automatically by the computing system in response to a determination the the prompt does not elicit undesired actions by the GenAI model, an output of the GenAI model to a requestor; and ([Goswami, Fig. 9 and 0108] discloses generating an output 930 indicating whether or not the SVM classifier 870 has determined that the new input data 910 is distorted or undistorted. [0122] discloses in response to the output of the SVM classifier, a responsive operation may be performed such as alerting contingent systems or operating or shutting down the downstream computing system, or dropout the adversarial image data) initiating, automatically by the computing system in response to a determination that the prompt elicits undesired actions by the GenAI model, at least one remediation action preventing the output as generated by the GenAI model from being returned to the requestor. ([Goswami, Fig. 9 and 0108] discloses generating an output 930 indicating whether or not the SVM classifier 870 has determined that the new input data 910 is distorted or undistorted. [0122] discloses in response to the output of the SVM classifier, a responsive operation may be performed such as alerting contingent systems or operating or shutting down the downstream computing system, or dropout the adversarial image data) Goswami does not specifically disclose: receiving, by a computing system, data characterizing a prompt redirecting, by the computing system, the prompt from the GenAI model to a quantized version of the GenAI model having an input layer, an output layer, and a plurality of intermediate transformer layers positioned immediately after the input layer and immediately before the output layer, the quantized version of the GenAI model executing on separate computing resources and being different than the GenAI model; capturing, during inference of the quantized version of the GenAI model, an intermediate result from residual streams derived from one or more of the intermediate transformer layers of the quantized version of the GenAI model, the capturing comprising applying the runtime hooks to each of the intermediate transformer layers and, after completion of a forward pass responsive to the prompt, querying each runtime hook to retrieve the actions and storing the retrieved activations in an array of floating-point values having a consistent shape, regardless of a length or content of the prompt, for input to the classifier; Shi teaches: receiving, by a computing system, data characterizing a prompt ([Shi, page 1974, Fig. 1] and [Shi, page 1975, left col, 3.2 Embedding layer, lines 1-4] show that the embedding layer obtains the input sequence of sentence representation, which are stream input. Fig.1 further shows that the model having an embedding layer (i.e., input layer), an output layer (i.e., softmax layer), and a plurality of intermediate transformer layers (i.e., PL transformer layer with a plurality of C-Encoder layers)) capturing, during inference of the , the capturing comprising applying the runtime hooks to each of the intermediate transformer layers and, after completion of a forward pass responsive to the prompt, querying each runtime hook to retrieve the actions and storing the retrieved activations in an array of floating-point values having a consistent shape, regardless of a length or content of the prompt, for input to the classifier; ([Shi, page 1975, left col, 3.2 Embedding layer, lines 1-4] shows that the embedding layer obtains the input sequence of sentence representation, which are stream input. The stored values are floating values because the C-Encoder Layer inherently involves Attention softmax and layer normalization as shown in the equation (7) which requires floating-point operations. [Shi, page 1974, Fig. 1] and [page 1976, left col, 3.3.2 Layer ensemble mechanisms, line 1 – right col, line 13] collectively discloses capturing PL-Transformer intermediate layers (C-Encoder Layers, which are interpreted as transformer layers) using hooks, and then ensembles the captured layer outputs h1(1)h2(1)…hn(1) and h1(2)h2(2)…hn(2) … h1(l)…hn(l). The transformer layer outputs are combined and sent to Linear Layer and Softmax to generate the final text classification output) Goswami in view of Shi does not specifically disclose: redirecting, by the computing system, the prompt from the GenAI model to a quantized version of the GenAI model having an input layer, an output layer, and a plurality of intermediate transformer layers positioned immediately after the input layer and immediately before the output layer, the quantized version of the GenAI model executing on separate computing resources and being different than the GenAI model; O’CONNOR teaches: redirecting, by the computing system, the prompt from the GenAI model to a quantized version of the GenAI model having an input layer, an output layer ([O’CONNOR, 0011], [0041] and [0043-0046] The student model 106 are interpreted as the ‘quantized version’ of the teacher model as it is trained based on the teacher model 105. [O’CONNOR, Fig. 4 and Fig. 7] The teacher network 105 receives input data 103 (Step S220), processes the input data 103 (Step S230), and then transmits the input data to the student neural network 106 (Step S200). [0051] shows that the input data are prompts (words spoken by speaker)) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi and O’CONNOR to use the method of redirecting the output of the GenAI model (teacher model) to a quantized model (student model) of O’CONNOR to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to increase the efficiency and performance of the quantized model by updating the simplified quantized (student) model based on the original GenAI model (teacher model) with higher complexity. However, Goswami in view of Shi and further in view of Li does not specifically disclose: the quantized version of the GenAI model executing on separate computing resources Elthakeb teaches: the quantized version of the GenAI model executing on separate computing resources ([Elthakeb, page 3, 2.1. Matching Activations for Intermediate Layers, lines 1-23], [Figure 3] collectively discloses that the quantized models have the same topology and the produce intermediate streams compatible with the full precision network (the GenAI model). [page 3, 2.2. Splitting, Training and Merging, lines 1-4] discloses that the models are trained in isolation and in parallel) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi, O’CONNOR and Elthakeb to use the method of quantizing the generative model of Elthakeb to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to increase the performance of the system, as quantizing a model makes running inference faster and cheaper by reducing the file size of model weight files. Claims 21-25 and 28 are rejected under 35 U.S.C. 103 as being unpatentable over Goswami in view of Shi and further in view of Inan et al. (“Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations”, 2023, hereinafter ‘Inan’). Regarding claim 21, Goswami teaches: A computer-implemented method comprising: receiving, by a monitoring computing environment from a model computing environment executing a [Goswami, 0108, Fig. 9] The new network activation data 920 which is the means for the intermediate layers of the DNN 820 is generated in response to the new input data and input to the SVL classifier 870. The DNN is interpreted as the GenAI, and the GenAI is further disclosed in Shi. And then, the SVL classifier 870 which is trained to detect distortion in the data is used to classify the data) determining, by the monitoring computing environment and based on the activations and using a machine learning-based classifier trained using a policy mapping residual stream activations to prohibited actions, whether the prompt seeks to cause the manner, ; and ([Goswami, Fig. 9 and 0108] discloses generating an output 930 indicating whether or not the SVM classifier 870 has determined that the new input data 910 is distorted or undistorted. [0109] discloses the pre-training process of the SVM classifier 870 using the undistorted data) returning, automatically by the monitoring computing environment in response to a determination the the prompt does not elicit undesired actions by the [Goswami, Fig. 9 and 0108] discloses generating an output 930 indicating whether or not the SVM classifier 870 has determined that the new input data 910 is an distorted or undistorted. [0122] discloses in response to the output of the SVM classifier, a responsive operation may be performed such as alerting contingent systems or operating or shutting down the downstream computing system) initiating, automatically by the monitoring computing environment in response to a determination that the prompt elicits undesired actions by the GenAI model, at least one remediation action preventing the output as generated by the GenAI model from being returned to the requestor. ([Goswami, Fig. 9 and 0108] discloses generating an output 930 indicating whether or not the SVM classifier 870 has determined that the new input data 910 is an distorted or undistorted. [0122] discloses in response to the output of the SVM classifier, a responsive operation may be performed such as alerting contingent systems or operating or shutting down the downstream computing system, or dropout the adversarial image data) Goswami does not specifically disclose: receiving, from a model computing environment executing a generative artificial intelligence (GenAI) model, activations in residual streams between transformer layers of the GenAI model generated during inference of the GenAI model in response to a prompt, the activations being captured via runtime hooks during a forward pass; the classifier being a multi-class classifier that outputs a label identifying a type of the prohibited action, and the policy specifying a remedial action for at least one prohibited action such that the label output by the classifier is used to select the at least one remediation action; Shi teaches: receiving, from a model computing environment executing a generative artificial intelligence (GenAI) model, activations in residual streams between transformer layers of the GenAI model generated during inference of the GenAI model in response to a prompt, the activations being captured via runtime hooks during a forward pass; ([Shi, page 1975, left col, 3.2 Embedding layer, lines 1-4] shows that the embedding layer obtains the input sequence of sentence representation, which are stream input. [Shi, page 1974, Fig. 1] and [page 1976, left col, 3.3.2 Layer ensemble mechanisms, line 1 – right col, line 13] collectively discloses capturing PL-Transformer intermediate layers (C-Encoder Layers, which are interpreted as transformer layers) using hooks, and then ensembles the captured layer outputs h1(1)h2(1)…hn(1) and h1(2)h2(2)…hn(2) … h1(l)…hn(l). The transformer layer outputs are combined and sent to Linear Layer and Softmax to generate the final text classification output. The Transformer Layer is the GenAI model) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami and Shi to use a method of capturing an intermediate result of the transformer model of Shi to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the performance of the model by providing more minority class data from the trained latent space. However, Goswami in view of Shi do not specifically disclose: the classifier being a multi-class classifier that outputs a label identifying a type of the prohibited action, and the policy specifying a remedial action for at least one prohibited action such that the label output by the classifier is used to select the at least one remediation action Inan teaches: the classifier being a multi-class classifier that outputs a label identifying a type of the prohibited action, and the policy specifying a remedial action for at least one prohibited action such that the label output by the classifier is used to select the at least one remediation action ([Inan, page 2, 2 Safety Risk Taxonomy, line 1 – page 4, line 7] The automated input-output safeguards relies on classifiers to make decisions about content in real time and have a Taxonomy (classes), and Risk guidelines (policies). The paragraph 2.1 shows that the classifier is a multi-class classifier. [page 3, last paragraph, The output format, line 1 – page 4, line 7] shows the classifier’s output format, for example ‘unsafe O3’, wherein the ‘unsafe’ is the binary classification output and O (taxonomy category) followed by the 1-based category index (policy) is the multi-label classification output) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi and Inan to use a multi-class classifier that outputs a label identifying a type of prohibited action and a policy specifying a remedial action of Inan to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the security of the model by preventing the malicious prompt to be ingested by the system or model. Regarding claim 22, Goswami in view of Shi teaches: The method of claim 21, wherein capturing the activations comprises applying runtime hooks to every transformer layer and, after completion of a forward pass responsive to the prompt, querying each hook to retrieve activations from residual streams and storing the retrieved activations for the prompt in a single array of floating-point values. ([Shi, page 1974, Fig. 1] and [page 1976, left col, 3.3.2 Layer ensemble mechanisms, line 1 – right col, line 13] collectively discloses capturing PL-Transformer intermediate layers (C-Encoder Layers, which are interpreted as transformer layers) using hooks, and then ensembles the captured layer outputs h1(1)h2(1)…hn(1) and h1(2)h2(2)…hn(2) … h1(l)…hn(l). The transformer layer outputs are combined and sent to Linear Layer and Softmax to generate the final text classification output. [page 1975, right col, 3.3.1 C-encoder layer, lines 1-26] The stored values are floating values because the C-Encoder Layer inherently involves Attention softmax and layer normalization as shown in the equation (7) which requires floating-point operations.) Regarding claim 23, Goswami teaches: The method of claim 21, wherein determining whether the [Goswami, Fig. 9 and 0108] discloses generating an output 930 indicating whether or not the SVM classifier 870 has determined that the new input data 910 is an distorted or undistorted. [0122] discloses in response to the output of the SVM classifier, a responsive operation may be performed such as alerting contingent systems or operating or shutting down the downstream computing system, or dropout the adversarial image data. [Claims 1 and 6] indicates that the output value from the intermediate layers are vectors (array of numerical data)) Goswami does not specifically disclose: wherein determining whether the prompt Shi teaches: wherein determining (classifying) whether the prompt seeks… ([Shi, page 1974, Fig. 1] and [page 1976, left col, 3.3.2 Layer ensemble mechanisms, line 1 – right col, line 13] collectively discloses capturing PL-Transformer intermediate layers (C-Encoder Layers, which are interpreted as transformer layers) using hooks, and then ensembles the captured layer outputs h1(1)h2(1)…hn(1) and h1(2)h2(2)…hn(2) … h1(l)…hn(l). The transformer layer outputs are combined and sent to Linear Layer and Softmax to generate the final text classification output) Regarding claim 24, Goswami in view of Shi and further in view of Inan teaches: The method of claim 21, wherein the classifier is a multi-class classifier executed by an analysis engine in a monitoring environment and configured to output a label identifying exactly one prohibited action selected from: a response in a non-approved spoken or written language or computer code. ([Inan, page 2, 2 Safety Risk Taxonomy, line 1 – page 4, line 7] The automated input-output safeguards relies on classifiers to make decisions about content in real time and have a Taxonomy (classes), and Risk guidelines (policies). The paragraph 2.1 shows that the classifier is a multi-class classifier. [page 3, last paragraph, The output format, line 1 – page 4, line 7] shows the classifier’s output format, for example ‘unsafe O3’, wherein the ‘unsafe’ is the binary classification output and O (taxonomy category) followed by the 1-based category index (policy) is the multi-label classification output) Regarding claim 25, Goswami teaches: The method of claim 21, wherein a policy maps intermediate results to prohibited actions and specifies a remedial action for at least one prohibited action, and wherein the classifier output is used to select the remedial action. ([Goswami, 0090 and 0092] discloses the selective dropout logic based on the classification by the SVM classifier. The selective dropout selects which portion of the input data to dropout, which corresponds to the selection of remedial action) Claim 26 is rejected under 35 U.S.C. 103 as being unpatentable over Goswmai in view of Shi in view of Inan and further in view of Kuan. Regarding claim 26, Goswami in view of Shi and further in view of Inan teaches: The method of claim 21. However, Goswami in view of Shi and further in view of Inan do not specifically disclose: wherein the consuming application or process, based on the determination, causes external remediation resources to block, at a network gateway, subsequent requests from an internet protocol (IP) address of a requester of the prompt. Kuan teaches: wherein the consuming application or process, based on the determination, causes external remediation resources to block, at a network gateway, subsequent requests from an internet protocol (IP) address of a requester of the prompt. ([Kuan, 0080] IP addresses associated with the malicious data 704 identified by the anomaly prediction system 114 is blocked by the control action 708 performed by the system) Claim 27 is rejected under 35 U.S.C. 103 as being unpatentable over Goswami in view of Shi in view of Inan and further in view of Rand et al. (US 20070204341 A1, hereinafter ‘Rand’) Regarding claim 27, Goswami teaches: The method of claim 21, wherein the consuming application or process, based on the determination, causes a local remediation engine in a model environment to [Goswami, Fig. 9 and 0108] discloses generating an output 930 indicating whether or not the SVM classifier 870 has determined that the new input data 910 is distorted or undistorted. [0122] discloses in response to the output of the SVM classifier, a responsive operation may be performed such as alerting contingent systems or operating or shutting down the downstream computing system, or dropout the adversarial image data) However, Goswami in view of Shi and further in view of Inan do not specifically disclose: delete portions of subsequent prompts from an internet protocol (IP) address of a requester of the prompt and to cause a modified prompt to be ingested by the GenAI model Rand teaches: delete portions of subsequent prompts from an internet protocol (IP) address of a requester of the prompt and to cause a modified prompt to be ingested by the [Rand, 0034] discloses modifying malicious or prohibited activities from specific IP addresses to be ingested by the system) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi, Inan and Rand to use a method of modifying prompts from an IP address to be ingested by the system of Rand to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the security of the model by preventing the malicious prompt to be ingested by the system or model. Claims 28-29 are rejected under 35 U.S.C. 103 as being unpatentable over Goswami in view of Shi in view of Inan and further in view of Elthakeb. Regarding claim 28, Goswami teaches: The method of claim 21, wherein the activations are captured by a proxy of the GenAI model that executes in a monitoring environment and, based on an output or an intermediate result of the proxy, controls whether the GenAI model ingests the prompt ([Goswami, 0107; Fig. 8 and Fig. 9] The detection module 840 comprises layer-wise comparison logic 844 which captures every network activation data 830 and 860 to generate distance metrics for each of the intermediate layers of the DNN 820. [0109] discloses that the classifier 870 discards the input data if the input is detected to be adversarial) However, Goswami in view of Shi and further in view of Inan does not specifically disclose: the proxy comprising a model that mimics some or all of the behavior of the GenAI model. Elthakeb teaches: the proxy comprising a model that mimics some or all of the behavior of the GenAI model. ([Elthakeb, page 3, 2.1. Matching Activations for Intermediate Layers, lines 1-23], [Figure 3] collectively discloses that the quantized models have the same topology and the produce intermediate streams compatible with the full precision network (the GenAI model)) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi, Inan and Elthakeb to use a method of utilizing a quantized version of the GenAI model having a same topology with the GenAI model of Elthakeb to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the efficiency of the model by offering additional speedup through enabling parallelization [Elthakeb, ABSTRACT]. Regarding claim 29, Goswami in view of Shi in view of Inan and further in view of Elthakeb teaches: wherein the proxy is a quantized version of the GenAI model having a same topology and producing intermediate residual streams compatible with the GenAI model, and wherein quantization preserves access to residual stream activations via the runtime hooks. ([Elthakeb, page 3, 2.1. Matching Activations for Intermediate Layers, lines 1-23], [Figure 3] collectively discloses that the quantized models have the same topology and the produce intermediate streams compatible with the full precision network (the GenAI model)) Claim 30 is rejected under 35 U.S.C. 103 as being unpatentable over Goswami in view of Shi in view of Inan and further in view of Schmidtler et al. (US 20160335435 A1, hereinafter ‘Schmidtler’) Regarding claim 30, Goswami in view of Shi and further in view of Inan teaches: The method of claim 21, [Goswami, 0107; Fig. 8 and Fig. 9] The detection module 840 comprises layer-wise comparison logic 844 which captures every network activation data 830 and 860 to generate distance metrics for each of the intermediate layers of the DNN 820. [0109] discloses that the classifier 870 discards the input data if the input is detected to be adversarial. [0110] discloses that the SVM classifier is trained using adversarial data and based on the responses from intermediate layers) However, Goswami in view of Shi and further in view of Inan does not specifically disclose: wherein the policy used to train the classifier defines that encrypted instructions or instructions to encrypt an output constitute a prohibited action Schmidtler teaches: wherein the policy used to train the classifier defines that encrypted instructions or instructions to encrypt an output constitute a prohibited action ([Schmidtler, 0009] discloses that the classifier is trained from a collection of data comprising known malicious files, and the classifier is trained such that it can handle encrypted and/or compressed files without decrypting it) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi, Inan and Schmidtler to use a method of train the classifier defines that encrypted instructions or instructions to encrypt of Schmidtler to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the security of the model by preventing attackers from accessing the output data and/or input data. Claim 31 is rejected under 35 U.S.C. 103 as being unpatentable over Goswami in view of Shi in view of Inan and further in view of Cai et al. (US 20060026675 A1, hereinafter ‘Cai’). Regarding claim 31, Goswami teaches: The method of claim 21, wherein the policy used to train the classifier [Goswami, Fig. 9 and 0108] discloses generating an output 930 indicating whether or not the SVM classifier 870 has determined that the new input data 910 is distorted or undistorted. [0122] discloses in response to the output of the SVM classifier, a responsive operation may be performed such as alerting contingent systems or operating or shutting down the downstream computing system, or dropout the adversarial image data) Goswami in view of Shi and further in view of Inan does not specifically disclose: the classifier defines detection of obfuscation patterns including one or more of base64, hexadecimal, and rot13 encodings Cai teaches: the classifier defines detection of obfuscation patterns including one or more of base64, hexadecimal, and rot13 encodings ([Cai, 0028] discloses classifying malicious hexadecimal files and extracting single byte patterns using a SVM classifier) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi, Inan and Schmidtler to use a method of detecting obfuscation patterns using the classifier of Cai to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the security of the model by enabling the classifier to capture malicious patterns in various types of input data. Claim 32 is rejected under 35 U.S.C. 103 as being unpatentable over Goswami in view of Shi and further in view of Elthakeb. Regarding claim 32, Goswami teaches: A computing system comprising: one or more hardware processors and memory storing instructions which, when executed by the one or more hardware processors, cause the computing system to provide: ([Goswami, 0077-0078] Goswami is implemented using a processor and storage medium having instructions) a model computing environment executing a [Goswami, 0108, Fig. 9] The new network activation data 920 which is the means for the intermediate layers of the DNN 820 is generated in response to the new input data and input to the SVL classifier 870. The computing environment in which the DNN and the SVL classifier run is the model computing environment) [Goswami, 0108, Fig. 9] The new network activation data 920 which is the means for the intermediate layers of the DNN 820 is generated in response to the new input data and input to the SVL classifier 870. And then, the SVL classifier 870 which is trained to detect distortion in the data is used to classify the data) an analysis engine configured to assemble the activations into an array of numerical data having a consistent shape and to apply a machine learning-based classifier trained using a policy mapping intermediate results to prohibited actions comprising at least a response in a non-approved spoken or written language and computer code; and ([Goswami, 0108, Fig. 9] The new network activation data 920 which is the means for the intermediate layers of the DNN 820 is generated in response to the new input data and input to the SVL classifier 870. And then, the SVL classifier 870 which is trained to detect distortion in the data is used to classify the data. [0122] discloses in response to the output of the SVM classifier, a responsive operation may be performed such as alerting contingent systems or operating or shutting down the downstream computing system) a remediation engine configured, in response to a classifier output indicating a prohibited action, to block, at a network gateway, transmission of an output of the GenAI model to a requestor. ([Goswami, Fig. 9 and 0108] discloses generating an output 930 indicating whether or not the SVM classifier 870 has determined that the new input data 910 is distorted or undistorted. [0122] discloses in response to the output of the SVM classifier, a responsive operation may be performed such as alerting contingent systems or operating or shutting down the downstream computing system) Goswami does not specifically disclose: a model computing environment executing a generative artificial intelligence (GenAI) model having an input layer, an output layer, and a plurality of intermediate transformer layers; a monitoring computing environment distinct from the model computing environment; wherein the monitoring computing environment comprises: a proxy of the GenAI model executing in the monitoring computing environment, the proxy comprising a quantized version of the GenAI model having a same topology and producing intermediate residual streams compatible with the GenAI model; runtime hooks applied to one or more transformer layers of the proxy and configured to capture, during a forward pass responsive to a prompt, activations in residual streams; Shi teaches: a model computing environment executing a generative artificial intelligence (GenAI) model having an input layer, an output layer, and a plurality of intermediate transformer layers; ([Shi, page 1974, Fig. 1] and [Shi, page 1975, left col, 3.2 Embedding layer, lines 1-4] show that the embedding layer obtains the input sequence of sentence representation, which are stream input. Fig.1 further shows that the model having an embedding layer (i.e., input layer), an output layer (i.e., softmax layer), and a plurality of intermediate transformer layers (i.e., PL transformer layer with a plurality of C-Encoder layers). The transformer layer is the GenAI model) runtime hooks applied to one or more transformer layers of the proxy and configured to capture, during a forward pass responsive to a prompt, activations in residual streams; ([Shi, page 1975, left col, 3.2 Embedding layer, lines 1-4] shows that the embedding layer obtains the input sequence of sentence representation, which are stream input. [Shi, page 1974, Fig. 1] and [page 1976, left col, 3.3.2 Layer ensemble mechanisms, line 1 – right col, line 13] collectively discloses capturing PL-Transformer intermediate layers (C-Encoder Layers, which are interpreted as transformer layers) using hooks, and then ensembles the captured layer outputs h1(1)h2(1)…hn(1) and h1(2)h2(2)…hn(2) … h1(l)…hn(l). The transformer layer outputs are combined and sent to Linear Layer and Softmax to generate the final text classification output) Goswami in view of Shi does not specifically disclose: a monitoring computing environment distinct from the model computing environment; wherein the monitoring computing environment comprises: a proxy of the GenAI model executing in the monitoring computing environment, the proxy comprising a quantized version of the GenAI model having a same topology and producing intermediate residual streams compatible with the GenAI model; Elthakeb teaches: a monitoring computing environment distinct from the model computing environment; wherein the monitoring computing environment comprises: ([Elthakeb, page 3, 2.1. Matching Activations for Intermediate Layers, lines 1-23], [Figure 3] collectively discloses that the quantized models have the same topology and the produce intermediate streams compatible with the full precision network (the GenAI model). [page 3, 2.2. Splitting, Training and Merging, lines 1-4] discloses that the models are trained in isolation and in parallel) a proxy of the GenAI model executing in the monitoring computing environment, the proxy comprising a quantized version of the GenAI model having a same topology and producing intermediate residual streams compatible with the GenAI model; ([Elthakeb, page 3, 2.1. Matching Activations for Intermediate Layers, lines 1-23], [Figure 3] collectively discloses that the quantized models have the same topology and the produce intermediate streams compatible with the full precision network (the GenAI model)) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami, Shi and Elthakeb to use the method of quantizing the generative model of Elthakeb to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to increase the performance of the system, as quantizing a model makes running inference faster and cheaper by reducing the file size of model weight files. Double Patenting The nonstatutory double patenting rejection is based on a judicially created doctrine grounded in public policy (a policy reflected in the statute) so as to prevent the unjustified or improper timewise extension of the “right to exclude” granted by a patent and to prevent possible harassment by multiple assignees. A nonstatutory double patenting rejection is appropriate where the conflicting claims are not identical, but at least one examined application claim is not patentably distinct from the reference claim(s) because the examined application claim is either anticipated by, or would have been obvious over, the reference claim(s). See, e.g., In re Berg, 140 F.3d 1428, 46 USPQ2d 1226 (Fed. Cir. 1998); In re Goodman, 11 F.3d 1046, 29 USPQ2d 2010 (Fed. Cir. 1993); In re Longi, 759 F.2d 887, 225 USPQ 645 (Fed. Cir. 1985); In re Van Ornum, 686 F.2d 937, 214 USPQ 761 (CCPA 1982); In re Vogel, 422 F.2d 438, 164 USPQ 619 (CCPA 1970); In re Thorington, 418 F.2d 528, 163 USPQ 644 (CCPA 1969). A timely filed terminal disclaimer in compliance with 37 CFR 1.321(c) or 1.321(d) may be used to overcome an actual or provisional rejection based on nonstatutory double patenting provided the reference application or patent either is shown to be commonly owned with the examined application, or claims an invention made as a result of activities undertaken within the scope of a joint research agreement. See MPEP § 717.02 for applications subject to examination under the first inventor to file provisions of the AIA as explained in MPEP § 2159. See MPEP § 2146 et seq. for applications not subject to examination under the first inventor to file provisions of the AIA . A terminal disclaimer must be signed in compliance with 37 CFR 1.321(b). The filing of a terminal disclaimer by itself is not a complete reply to a nonstatutory double patenting (NSDP) rejection. A complete reply requires that the terminal disclaimer be accompanied by a reply requesting reconsideration of the prior Office action. Even where the NSDP rejection is provisional the reply must be complete. See MPEP § 804, subsection I.B.1. For a reply to a non-final Office action, see 37 CFR 1.111(a). For a reply to final Office action, see 37 CFR 1.113(c). A request for reconsideration while not provided for in 37 CFR 1.113(c) may be filed after final for consideration. See MPEP §§ 706.07(e) and 714.13. The USPTO Internet website contains terminal disclaimer forms which may be used. Please visit www.uspto.gov/patent/patents-forms. The actual filing date of the application in which the form is filed determines what form (e.g., PTO/SB/25, PTO/SB/26, PTO/AIA /25, or PTO/AIA /26) should be used. A web-based eTerminal Disclaimer may be filled out completely online using web-screens. An eTerminal Disclaimer that meets all requirements is auto-processed and approved immediately upon submission. For more information about eTerminal Disclaimers, refer to www.uspto.gov/patents/apply/applying-online/eterminal-disclaimer. Claim 20 is rejected on the ground of nonstatutory double patenting as being obvious under claim 16 of US 12572777 in view of Shi and further in view of Elthakeb. Present application 18/669,366 Reference patent US 12572777 20 A computer-implemented method comprising: receiving, by a computing system, data characterizing a prompt for ingestion by a generative artificial intelligence (GenAI); 16 A computer-implemented method for policy-based gating of multimodal prompts in a generative artificial intelligence (GenAI) system, the method comprising: receiving, by a proxy computing system, data characterizing a multimodal prompt for ingestion by a production GenAI model, the GenAI model comprising a transformer architecture with a plurality of layers including at least one intermediate layer and residual streams; redirecting, by the computing system, the prompt from the GenAI model to a quantized version of the GenAI model having an input layer, an output layer, and a plurality of intermediate transformer layers positioned immediately after the input layer and immediately before the output layer, the quantized version of the GenAI model executing on separate computing resources and being different than the GenAI model; redirecting the multimodal prompt to a quantized proxy model that shares a layer topology with the production GenAI model, the quantized proxy model configured to process the prompt without generating a final output; capturing, during inference of the quantized version of the GenAI model and using runtime hooks, an intermediate result comprising activations in residual streams of one or more of the intermediate transformer layers of the quantized version of the GenAI model, the capturing comprising applying the runtime hooks to each of the intermediate transformer layers and, after completion of a forward pass responsive to the prompt, querying each runtime hook to retrieve the activations and storing the retrieved activations in an array of floating-point values having a consistent shape, regardless of a length or content of the prompt, for input to the classifier; capturing, by the proxy computing system and via forward-pass hooks applied to at least one transformer layer of the quantized proxy model, per-token activation vectors from residual streams at a predetermined intermediate layer, wherein the capturing occurs after encoding of the prompt and before any output is generated by the production GenAI model; determining, using a machine learning-based classifier trained using a policy that maps intermediate residual stream activations to acceptable and undesired actions and based on the intermediate result, whether the prompt elicits undesired actions by the GenAI model, the classifier being trained using a data set generated according to a policy which maps acceptable and undesired actions to intermediate results; applying, by the proxy computing system, a classifier trained using a policy-labeled dataset that maps internal activation patterns to a plurality of enumerated prohibited actions, the prohibited actions comprising at least: (i) generation of computer code, (ii) response in a non-approved language, (iii) response to an encrypted prompt, or (iv) cross-modal obfuscation; returning, automatically by the computing system in response to a determination that the prompt does not elicit undesired actions by the GenAI model, an output of the GenAI model to a requestor; and determining, by the proxy computing system and based on the per-token activation vectors, whether the prompt is likely to elicit a prohibited action by the production GenAI model; and initiating, automatically by the computing system in response to a determination that the prompt elicits undesired actions by the GenAI model, at least one remediation action preventing the output as generated by the GenAI model from being returned to the requestor. in response to the determination, gating the prompt by preventing the prompt from being ingested by the production GenAI model when the determination indicates a prohibited action, and allowing ingestion otherwise, such that the production GenAI model does not process prompts likely to elicit prohibited actions. However, the reference patent US 12572777 does not suggest or disclose: the quantized version of the GenAI model executing on separate computing resources and being different than the GenAI model; the capturing comprising applying the runtime hooks to each of the intermediate transformer layers and, after completion of a forward pass responsive to the prompt, querying each runtime hook to retrieve the activations and storing the retrieved activations in an array of floating-point values having a consistent shape, regardless of a length or content of the prompt, for input to the classifier; Shi teaches: the capturing comprising applying the runtime hooks to each of the intermediate transformer layers and, after completion of a forward pass responsive to the prompt, querying each runtime hook to retrieve the activations and storing the retrieved activations in an array of floating-point values having a consistent shape, regardless of a length or content of the prompt, for input to the classifier; ([Shi, page 1975, left col, 3.2 Embedding layer, lines 1-4] shows that the embedding layer obtains the input sequence of sentence representation, which are stream input. The stored values are floating values because the C-Encoder Layer inherently involves Attention softmax and layer normalization as shown in the equation (7) which requires floating-point operations. [Shi, page 1974, Fig. 1] and [page 1976, left col, 3.3.2 Layer ensemble mechanisms, line 1 – right col, line 13] collectively discloses capturing PL-Transformer intermediate layers (C-Encoder Layers, which are interpreted as transformer layers) using hooks, and then ensembles the captured layer outputs h1(1)h2(1)…hn(1) and h1(2)h2(2)…hn(2) … h1(l)…hn(l). The transformer layer outputs are combined and sent to Linear Layer and Softmax to generate the final text classification output) Before the effective filing date of the invention to a person of ordinary skill in the art, it would have been obvious, having the teachings of Goswami and Shi to use a method of capturing an intermediate result of the transformer model of Shi to implement the machine learning method of Goswami. The suggestion and/or motivation for doing so is to improve the performance of the model by providing more minority class data from the trained latent space. However, the reference patent in view of Shi do not specifically disclose: the quantized version of the GenAI model executing on separate computing resources and being different than the GenAI model; Elthakeb teaches: the quantized version of the GenAI model executing on separate computing resources and being different than the GenAI model; ([Elthakeb, page 3, 2.1. Matching Activations for Intermediate Layers, lines 1-23], [Figure 3] collectively discloses that the quantized models have the same topology and the produce intermediate streams compatible with the full precision network (the GenAI model). [page 3, 2.2. Splitting, Training and Merging, lines 1-4] discloses that the models are trained in isolation and in parallel) It would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to apply the method of running the proxy of the GenAI model on distinct computing hardware by Elthakeb to improve the performance of the machine learning method of the present invention. The suggestion and/or motivation for doing so is to improve the efficiency of the method by offering additional speedup through enabling parallelization [Elthakeb, ABSTRACT]. Response to Arguments Response to Arguments under 35 U.S.C. 101 35 U.S.C. 101 rejections based on MPEP 2106.04(a) Abstract Ideas have been withdrawn. Response to Arguments under 35 U.S.C. 103 Arguments: Applicant asserts that (a) Claim 1, none of the cited references teach the amended claim 1 [Remarks, pages 12-13], (b) Claim 20, none of the cited references teaches the specific capture-and-assembly mechanism of “capturing, during inference of the quantized version of the GenAI model and using runtime hooks, an intermediate result comprising … storing the retrieved activations in an array of floating-point values having a consistent shape, regardless of a length or content of the prompt, for input to the classifier” [Remarks, pages 13-14], (c) Claim 21, none of the cited references teaches the specific capture-and-assembly mechanism because Goswami’s SVM produces only a binary distorted/undistorted output, and its “responsive operation” is generic and is not selected from a multi-class label identifying a type of prohibited action [page 15], (d) Claim 32, Goswami runs its neural network and classifier within a single data processing system and has no separate monitoring environment executing a proxy of the model whose activation drive classification and gateway-level remediation [Remarks, page 15]. Examiner’s response: Applicant’s argument (a) have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument. Regarding (b), examiner respectfully disagrees that Shi do not teach applying hooks to each layer, querying each hook after the forward pass, and storing the activations in a consistent-shape floating point array for a separate classifier. [Shi, page 1975, right col, 3.3.1 C-encoder layer] The layer-wise output of C-Encoder are floating-point array because C-encoder layer operation involves Cor-Att operation in equation (7) and normalization operation that inherently requires floating point operations. Even though Goswami does not specifically disclose “applying hooks”, Goswami teaches capturing intermediate result and inputting the intermediate result to a separate classifier. O’Connor teaches that the input data to the quantized model (student) is redirected from another model (teacher), and Elthakeb teaches running the quantized version of the GenAI model on separate computing resources. Therefore, examiner believes that the combination teaches the specific capture-and-assembly mechanism of the present invention. Applicant’s argument (c) have been considered but moot because the new ground of rejection relies on Inan which teaches multi-class classifier that outputs a category and a policy. Regarding (d), claim 32 does not specifically disclose that the ‘classifier’ and the ‘neural network’ are run in a multiple data processing systems. The claim recites “a remediation engine configured, in response to a classifier output indicating a prohibited action, to block, at a network gateway.” This limitation can be interpreted in two different interpretations, (a) the remediation engine receives classifier (located anywhere) output and use it to prohibit actions, or the classifier is in the remediation engine and used to prohibit actions. Therefore, argument (d) is not persuasive. Conclusion THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a). A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action. Any inquiry concerning this communication or earlier communications from the examiner should be directed to JUN KWON whose telephone number is (571)272-2072. The examiner can normally be reached Monday – Friday 8:00AM – 4:00PM ET. Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice. If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Abdullah Kawsar can be reached at (571)270-3169. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300. Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000. /JUN KWON/Examiner, Art Unit 2127 /ABDULLAH AL KAWSAR/Supervisory Patent Examiner, Art Unit 2127
Read full office action

Prosecution Timeline

Show 10 earlier events
Aug 08, 2025
Final Rejection mailed — §101, §103, §112
Sep 26, 2025
Response after Non-Final Action
Oct 08, 2025
Request for Continued Examination
Oct 15, 2025
Response after Non-Final Action
Jul 22, 2026
Non-Final Rejection mailed — §101, §103, §112
Jul 23, 2026
Response Filed
Oct 09, 2026
Final Rejection mailed — §101, §103, §112
Oct 12, 2026
Response after Non-Final Action

Precedent Cases

Applications granted by this same examiner with similar technology

Patent 12737633
TRAINING A FEDERATED GENERATIVE ADVERSARIAL NETWORK
3y 9m to grant Granted Sep 15, 2026
Patent 12731035
LABEL INFERENCE IN SPLIT LEARNING DEFENSES
3y 8m to grant Granted Sep 08, 2026
Patent 12718063
DEEP LEARNING ARCHITECTURE FOR ADVERSE MEDIA SCREENING
3y 11m to grant Granted Aug 25, 2026
Patent 12711383
ACCURATE ENSEMBLE BY MUTATING NEURAL NETWORK PARAMETERS
7y 3m to grant Granted Aug 18, 2026
Patent 12705504
KNOWLEDGE BASE CONSTRUCTION
8y 5m to grant Granted Aug 11, 2026
Study what changed to get past this examiner. Based on 5 most recent grants.

Strategy Recommendation AI-generated — please review before filing

Get a prosecution strategy drawn from examiner precedents, rejection analysis, and claim mapping.
Typically takes 5-10 seconds — AI-generated, attorney review required before filing

Prosecution Projections

6-7
Expected OA Rounds
42%
Grant Probability
83%
With Interview (+40.5%)
4y 8m (~2y 2m remaining)
Median Time to Grant
High
PTA Risk
Based on 83 resolved cases by this examiner. Grant probability derived from career allowance rate.

Sign in with your work email

Enter your email to receive a magic link. No password needed.

Personal email addresses (Gmail, Yahoo, etc.) are not accepted.

Free tier: 3 strategy analyses per month