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 .
Priority
Applicant’s claim for the benefit of a prior-filed application under 35 U.S.C. 119(e) or under 35 U.S.C. 120, 121, 365(c), or 386(c) is acknowledged. Applicant has not complied with one or more conditions for receiving the benefit of an earlier filing date under 35 U.S.C. 119(e) as follows:
The later-filed application must be an application for a patent for an invention which is also disclosed in the prior application (the parent or original nonprovisional application or provisional application). The disclosure of the invention in the parent application and in the later-filed application must be sufficient to comply with the requirements of 35 U.S.C. 112(a) or the first paragraph of pre-AIA 35 U.S.C. 112, except for the best mode requirement. See Transco Products, Inc. v. Performance Contracting, Inc., 38 F.3d 551, 32 USPQ2d 1077 (Fed. Cir. 1994).
The disclosure of the prior-filed application, Provisional Application No. 63/752,806, fails to provide adequate support or enablement in the manner provided by 35 U.S.C. 112(a) or pre-AIA 35 U.S.C. 112, first paragraph for one or more claims of this application.
The specification provided in 63/752,806 consists of two sections.
The second section, which begins at page 9, describes transformers, encoder-decoders, embeddings, and various attention mechanisms generally, and does not apply these general descriptions of known AI components to the particular invention claimed in the present applications. The drawings provided in 63/752,806 correspond to this second section, and thus only aid explanation of the general concepts described in the second section.
The first section, pages 1–7, lists features of a specific system, but does not describe how any feature is achieved, and thus reads as a proposal for a system and not a description of its implementation. Further, Examiner finds that some elements of the claims are not even proposed by any of the features listed in the provisional specification, including:
“decision logic […] in a current state represented by a machine-readable state vector stored by the decision-logic runtime” (Claims 1–19)
“a state-transition directive identifying at least one transition rule of the decision logic” (Claims 1–19)
“a scoring rubric specifying at least one measurable criterion for evaluating alignment of the response with a learning goal” (Claims 1–19)
“an agent configured to generate a stimulant and further configured to apply the stimulant to the decision logic to transition the decision logic from the current state to a next state” (Claims 1–20)
“wherein the prompt comprises one or more of a user query, a system-generated query, a probing input, a context-adding input, or an adversarial input” (Claim 2)
“a state of the decision logic comprises one or more of a rule, a condition, or a causal path identifying a reason the response was generated by the AI model” (Claims 3–4 and 17)
“the state vector encodes at least one of: a selected strategy identifier, a constraint identifier, or a reasoning-path identifier” (Claims 3–4 and 17)
“wherein the decision trace further comprises: (i) a hash of the current state vector, (ii) a hash of the next state vector, and (iii) an identifier of the transition rule applied, wherein the decision trace is stored as a tamper-evident record linked to at least one prior decision trace via a hash chain” (Claims 5–12 and 16)
“increase storage duration for the decision trace when the decision trace comprises a positive associated emotional state and when the learning-goal gap satisfies a reinforcement threshold” (Claim 7)
“apply an exponential decay function to the decision trace when the decision trace is not associated with a positive outcome under the scoring rubric” (Claim 8)
This list should not be taken as exhaustive.
Examiner finds that a person having ordinary skill in the art would not recognize, from the disclosure of provisional application 63/752,806, that the inventors had possession of the invention of any claim in the present non-provisional application. Examiner further finds that the 63/752,806 disclosure provides no implementation information specific to the claims of the present invention, and therefore, a person having ordinary skill in the art would be unable to implement the claimed inventions of the present non-provisional application without undue experimentation. As the 63/752,806 disclosure does not support the claims under either the enablement or the written description requirements of 35 U.S.C. 112(a), the application for the benefit of the 63/752,806 is denied for all present claims, which will be examined using the present application’s filing date of 02/02/2026.
Claim Interpretation
Examiner here notes that the term “stimulant,” used in claims 1–19, is being interpreted according to its use in independent claims 1 or 18, which recite that a stimulant “comprises (i) a state-transition directive identifying at least one transition rule of the decision logic, and (ii) a scoring rubric specifying at least one measurable criterion for evaluating alignment of the response with a learning goal.”
Likewise, the term “temperature,” used in claim 16, is interpreted as any data which is used such that “a higher … parameter increases exploration in selection of transition rules and a lower … parameter enforces constraint-satisfaction in selection of transition rules.”
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1–19 rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Regarding claims 1–19:
Claim 1 recites “wherein the transition updates the state vector without updating parameters of the AI model.” Examiner finds this limitation indefinite in the context of the claim. The language does not make clear what is included in, or excluded from, “the parameters of the AI model.” Examiner notes, for example, that “the state vector” which is updated by the limitation is a representation of “the decision logic” which is executed to generate a response from the AI model. Hence, “the state vector” partly determines the operation of the “AI model” and therefore is itself reasonably considered a parameter of the AI model. A person having ordinary skill in the art would be unable to determine which values within of the AI model or controlling the AI model could be updated without violating the restriction “without updating parameters of the AI model” and thereby determine the metes and bounds of the claim.
Claims 2–17 are dependent upon claim 1 and are rejected by the same arguments.
Claim 18 recites the same language described as indefinite in claim 1, above, and is rejected by the same arguments.
Claim 19 is dependent on claim 18 and is rejected by the same arguments.
In further examination, the phrase “wherein the transition updates the state vector without updating parameters of the AI model” will be interpreted to mean only that the AI model is not explicitly re-trained as part of a transition update.
Regarding claims 6–7 and 16:
In addition to the above, claim 6 recites “the decision trace further comprises an associated emotional state inferred from a learning-goal gap determined using the scoring rubric, the associated emotional state comprising at least one of a positive state, an altered state, a neutral state, or a cautious state.”
Examiner finds the term “emotional state” indefinite in the context of this claim. Neither the claim nor the specification define to what entity the state refers; whether, for example, the “emotional state” refers to the emotional state of a user, of a “decision trace,” or of the AI model. If a “decision trace” or AI model is intended, the claim does not make clear how such an inference is measured. Further, the claim does not define what qualifies a state as “a positive state, an altered state, a neutral state, or a cautious state,” and the specification provides no explicit definitions.
Examiner notes that one disclosed embodiment (paragraph 0142 of the present specification) suggests that the emotional state is merely a label for a range of values from a scoring rubric; however, the term is not explicitly defined such as to limit the interpretation of the claim to this embodiment.
Therefore, a person having ordinary skill in the art would be unable to determine a definite meaning for “emotional state” or any of the particular emotional states of “a positive state, an altered state, a neutral state, or a cautious state” and thereby determine the metes and bounds of the claim.
Claims 7 and 16 are dependent upon claim 6 and are rejected by the same arguments.
In further examination, “emotional state” will be interpreted as any measure of a “learning-goal gap.” Examiner finds that no particular meaning can be ascribed to each of “a positive state, an altered state, a neutral state, or a cautious state,” and therefore the limitation “the associated emotional state comprising at least one of a positive state, an altered state, a neutral state, or a cautious state” will be interpreted as that the “emotional state” takes on one or more values or ranges of values.
In claim 7, the limitation “including logic to increase storage duration for the decision trace when the decision trace comprises a positive associated emotional state” will be interpreted as that a decision trace storage duration can increase based on the value of an emotional state, hence, based on a particular value from a scoring rubric.
Regarding claim 7:
In addition to the above, claim 7 recites “further including logic to increase storage duration for the decision trace when the decision trace comprises a positive associated emotional state and when the learning-goal gap satisfies a reinforcement threshold.” This limitation appears to indicate that the duration individual decision traces spend in memory or storage is variable.
Through dependency with claim 5, claim 7 also recites “wherein the decision trace is stored as a tamper-evident record linked to at least one prior decision trace via a hash chain.” Examiner finds that a “hash chain” is a term of art, describing a chain of hashed data blocks, in which each block contains a hash of the previous block. Thus, if any block is modified, subsequent blocks will be altered as well, which allows verification that individual blocks have not been altered, providing the “tamper-evident” attribute of hash chains, as explained in Rosenberg, “Hash Chains Explained: How Cryptographic Integrity Works,” February 2025, https://www.hyrelog.com/blog/hash-chains-explained (hereafter Rosenberg): (“Verification Process” section), “To verify the integrity of a hash chain, you: 1. Start with the first entry and compute its hash 2. Compare that hash with the ‘Previous Hash’ stored in the second entry 3. If they match, compute the hash of the second entry 4. Compare that with the ‘Previous Hash’ in the third entry 5. Continue this process through the entire chain. If any entry has been modified, the hash comparison will fail, immediately revealing the tampering.” Thus, detecting tampering requires verification using all blocks in the chain. Examiner finds the resolution of these two limitations indefinite, as they appear to state that individual decision traces may be removed, but also that all traces remain to constitute a hash chain. A person having ordinary skill in the art would be unable to determine the intended meaning of removing a trace in the context of this claim, for example, whether this removal is intended to break the hash chain, or whether the trace in the hash chain is somehow separate from the trace in storage.
Regarding claim 8:
In addition to the above, claim 8 recites “further including logic to apply an exponential decay function to the decision trace when the decision trace is not associated with a positive outcome under the scoring rubric, wherein decision traces not reinforced by positive outcomes are removed from memory over time.” This limitation appears to indicate that the duration individual decision traces spend in memory or storage is variable, and Examiner finds the combination of this limitation with “wherein the decision trace is stored as a tamper-evident record linked to at least one prior decision trace via a hash chain” of claim 5 indefinite for reasons analogous to those of claim 7, above.
Regarding claim 11:
In addition to the above, claim 11 recites “wherein the decision trace comprises a causal chain linking the response to one or more prior decision operations that contributed to generation of the response, enabling reinforcement of effective early decisions when a positive outcome is delayed.” Examiner finds the phrase “enabling reinforcement of effective early decisions when a positive outcome is delayed” indefinite in the context of the claim. Neither the claim nor the specification explicitly indicate what criteria must be met to enable this reinforcement. A person having ordinary skill in the art would therefore be unable to determine the metes and bounds of the claim. Examiner notes that paragraph 0037 of the specification states:
The decision trace additionally comprises a causal chain linking the response to one or more prior decision operations that contributed to generation of the response. This causal chain enables reinforcement of effective early decisions when a positive outcome is delayed.
This suggests that the causal chain inherently enables the reinforcement. In further examination, the phrase “enabling reinforcement of effective early decisions when a positive outcome is delayed” will accordingly be interpreted as stating an inherent property of the causal chain and not a narrowing limitation.
Regarding claims 13–15:
In addition to the above, claim 13 recites “wherein the agent is selected from a plurality of agents including one or more role-based agents configured to generate role-based stimulants that probe the decision logic from distinct reasoning perspectives.” Examiner finds the term “distinct reasoning perspectives” indefinite in the context of the claim. Neither the claim nor the specification provide a definition for what constitutes a “reasoning perspective,” or how two agent would need to differ in order to have different reasoning perspectives. Further, the term “distinct” in this context is a term of degree, and neither the claim nor the specification provides instruction to the degree of difference necessary for two reasoning perspectives to be distinct.
In further examination, the phrase “role-based stimulants that probe the decision logic from distinct reasoning perspectives” is interpreted to mean only that the stimulants are in some aspect different or distinguishable.
Further, the claim recites “wherein each stimulant is generated in a typed format comprising the state-transition directive and the scoring rubric.” Examiner finds the term “typed format” indefinite in the context of this claim. Neither the claim nor the specification provide criteria for indicating whether a format is typed, and the term is a not a term of art outside of text file formats, which is not applicable here.
In further examination, the term “typed format” is interpreted to mean only that the each stimulant comprises “the state-transition directive and the scoring rubric.”
Further, Examiner notes that in claim 1, upon which claim 13 depends, a single agent generates a single “stimulant.” Claim 13 recites that this agent is chosen from a plurality of agents. Examiner finds the further steps of claim 13, “a plurality of agents including one or more role-based agents configured to generate role-based stimulants […] and one or more context-based agents configured to generate context-based stimulant […], and wherein each stimulant is generated in a typed format comprising the state-transition directive and the scoring rubric” indefinite in the context of the claim. Examiner finds the phrasing indefinite as to whether the agents in the two groups are merely capable of generating stimulants or whether the system necessarily includes having the agents in the plurality generate stimulants.
Claim 14, which depends upon claim 13, further recites “combine stimulants generated by the one or more role-based agents and the one or more context-based agents into an ordered stimulus plan for application to the decision-logic runtime,” which explicitly states that the plurality of agents generate stimulants. However, Examiner finds the step indefinite in the context of the claim. As stated above, in claim 1, a single agent generates a single stimulant. A person having ordinary skill in the art would be unable to determine how the process of claim 14 relates to the single-agent process of claim 1. For example, in claim 1, after the one stimulant produced by the one agent is applied to the decision logic of an AI model, “the decision logic in its next state is used to generate the response to the prompt.” Whether the multiple-agent combination of stimulants of claim 14 is applied to the decision logic to affect the response generation is indefinite, and if the combination of stimulants is to be applied, at what point in the system process this application takes place would also be indefinite.
In claim 15, “the plurality of agents is configured to iteratively generate stimulants and apply the stimulants to the decision-logic runtime until a convergence condition is satisfied” is found indefinite by the same arguments as claim 14; a person having ordinary skill in the art cannot determine how this iterative application of stimulants relates to the single-agent process of claim 1.
In further examination, the stimulant generation of claims 13–15 is interpreted merely as a capability of each agent, and not that every agent will necessarily be employed to generate a stimulant. The multiple-agent combination of stimulants of claim 14 and the iterative application of stimulants of claim 15 are interpreted to not interact with the single-agent stimulant process of claim 1.
Regarding claim 17:
In addition to the above, claim 17 recites “wherein each decision logic in the plurality of decision logics is associated with a learning goal identifier, and wherein one or more decision logics, in the plurality of decision logics, associated with the learning goal identifier are retained in memory and one or more decision logics, in the plurality of decision logics, not associated with the learning goal identifier are subject to a decay function.”
Examiner finds this claim indefinite. A plain reading of the language is self-contradictory, first reciting that “each decision logic in the plurality of decision logics is associated with a learning goal identifier” but then also “one or more decision logics, in the plurality of decision logics, not associated with the learning goal identifier are subject to a decay function,” which implies that some of the decision logics are both associated and not associated with a learning goal identifier. Examiner finds that a person having ordinary skill in the art would be unable to determine how the decision logics are separated into groups that are retained or discarded, and thereby determine the metes and bounds of the claim. In further examination, the language of claim 17 is interpreted to mean only that some decision logics are retained and some are subject to a decay function.
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 1–4, 18, and 20 rejected under 35 U.S.C. 103 over Lehr, US Pre-Grant Publication No. 2021/0110293 (hereafter Lehr) in view of Lucas et al., US Pre-Grant Publication No. 2026/0119384 (hereafter Lucas).
Regarding claim 1 and analogous claim 18:
Lehr teaches:
(bold only) “A system, comprising: an artificial intelligence (AI) model that receives a prompt, and is configured with a decision logic executed by a decision-logic runtime to generate a response to the prompt”: Lehr, paragraph 0001, “The present disclosure relates to computer-implemented methods, software, and systems [system] for automated data processing in a process management environment”; Lehr, paragraph 0004, “One example method may include operations such as extracting a plurality of rules from a machine learning model [an artificial intelligence (AI) model] for prediction [configured with a decision logic executed by a decision-logic runtime] of execution results of instances of a process workflow.”
“wherein the decision logic is in a current state represented by a machine-readable state vector stored by the decision-logic runtime”: Lehr, paragraph 0079, "The process rules 225 may be provided in a selectable form, where the user 205 may select a rule from the process rules 225 and incorporate that rule into implemented process logic 260 for the rules service logic of the workflow service [the decision logic is in a current state represented by a machine-readable state vector stored by the decision-logic runtime]. it is favored to be displayed to the end user."
“an agent configured to generate a stimulant and further configured to apply the stimulant to the decision-logic runtime to transition the decision logic from the current state to a next state”: Lehr, paragraph 0052, "Based on computations of metrics for the decision trees and decision rules part of a model from the machine learning models 130, a selection of decision rules may be performed [an agent configured to generate a stimulant, stimulant interpreted as anything that comprises the two items listed in the later limitation] and these selected decision rules may be input and implemented as part of process rules 145 to be used during process execution [further configured to apply the stimulant to the decision-logic runtime to transition the decision logic from the current state to a next state]."
“wherein the stimulant comprises (i) a state-transition directive identifying at least one transition rule of the decision logic”: Lehr, paragraph 0052, "Based on computations of metrics for the decision trees and decision rules part of a model from the machine learning models 130, a selection of decision rules may be performed and these selected decision rules may be input and implemented as part of process rules 145 to be used during process execution [a state-transition directive identifying at least one transition rule of the decision logic]."
“and (ii) a scoring rubric specifying at least one measurable criterion for evaluating alignment of the response with a learning goal”: Lehr, paragraph 0051, "In some instances, the models at the machine learning models 130 may be evaluated based on different qualitative and quantitative parameters to evaluate performance of prediction of the generated models. As discussed above, rules extracted from a model from the machine learning models 130 may be evaluated in relation to precision, data coverage, and complexity [a scoring rubric specifying at least one measurable criterion for evaluating alignment of the response with a learning goal]."
“wherein the transition updates the state vector without updating parameters of the AI model”: Lehr, paragraph 0052, "Based on computations of metrics for the decision trees and decision rules part of a model from the machine learning models 130, a selection of decision rules may be performed and these selected decision rules may be input and implemented as part of process rules 145 to be used during process execution [wherein the transition updates the state vector without updating parameters of the AI model, interpreted as affected the execution of the process without explicitly retraining a model]."
(bold only) “wherein the decision logic in its next state is used to generate the response to the prompt”: Lehr, paragraph 002, “Software application may execute processes in relation to providing user-requested services. Processes may be defined as workflows that include multiple steps taking input and providing output [the decision logic in its next state is used to generate the response].”
Lehr does not explicitly teach:
(bold only) “A system, comprising: an artificial intelligence (AI) model that receives a prompt, and is configured with a decision logic executed by a decision-logic runtime to generate a response to the prompt”
(bold only) “wherein the decision logic in its next state is used to generate the response to the prompt”
Lucas teaches (bold only) “A system, comprising: an artificial intelligence (AI) model that receives a prompt, and is configured with a decision logic executed by a decision-logic runtime to generate a response to the prompt” and (bold only) “wherein the decision logic in its next state is used to generate the response to the prompt”: Lucas, paragraph 0037, "An AI model 162 may be a generative AI model, such as a generative language model ( e.g., small language model, large language model, etc., that responds to natural language prompts in natural language)[receives a prompt][generate a response to the prompt], generative image model (e.g., that responds to natural-language prompts with an image), generative video model (e.g., that responds to natural-language prompts with a video), generative coding model (e.g., that responds to natural-language prompts with software code), or the like.”
Lucas and Lehr are analogous arts as they are both related to analyzing process execution. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date of the claimed invention to have combined the prompt response of Lucas with the teachings of Lehr to arrive at the present invention, in order to produce user-readable output, as stated in Lucas, paragraph 0037, “An AI model 162 may be a generative AI model, such as a generative language model ( e.g., small language model, large language model, etc., that responds to natural language prompts in natural language), generative image model (e.g., that responds to natural-language prompts with an image), generative video model (e.g., that responds to natural-language prompts with a video), generative coding model (e.g., that responds to natural-language prompts with software code), or the like. As used herein, the term ‘natural language’ or ‘natural-language’ refers to language, including grammar, that would be expected in a normal conversation between two humans.”
Regarding claim 2:
Lehr as modified by Lucas teaches “[t]he system of claim 1.”
Lucas further teaches “wherein the prompt comprises one or more of a user query, a system-generated query, a probing input, a context-adding input, or an adversarial input”: Lucas , paragraphs 0036–0037, “AI agent 160 may collect data, analyze data, communicate with human users [a user query] and/or other software entities, collaborate with other AI agents 160 to complete a complex task, execute actions, learn and improve over time, and/or the like. Each AI agent 160 comprises or is communicatively coupled to at least one AI model 162. AI model 162 may be internal to AI agent 160, external but local (i.e., within computing environment 150) to AI agent 160, or external and remote (i.e., outside computing environment 150, e.g., hosted on third-party system 140, etc.) from AI agent 160. An AI model 162 may be a generative AI model, such as a generative language model ( e.g., small language
model, large language model, etc., that responds to natural language prompts in natural language), generative image model (e.g., that responds to natural-language prompts with an image), generative video model (e.g., that responds to natural-language prompts with a video), generative coding model (e.g., that responds to natural-language prompts with software code), or the like.”
Lucas and Lehr are combinable for the rationale given under claim 1.
Regarding claim 3:
Lehr as modified by Lucas teaches “[t]he system of claim 1.”
Lehr further teaches:
“wherein a state of the decision logic comprises one or more of a rule, a condition, or a causal path identifying a reason the response was generated by the AI model”: Lehr, paragraph 0079, "The process rules 225 may be provided in a selectable form, where the user 205 may select a rule [a rule] from the process rules 225 and incorporate that rule into implemented process logic 260 for the rules service logic of the workflow service. it is favored to be displayed to the end user."
“and wherein the state vector encodes at least one of: a selected strategy identifier, a constraint identifier, or a reasoning-path identifier”: Lehr, paragraph 0079, "The process rules 225 may be provided in a selectable form [a selected strategy identifier], where the user 205 may select a rule from the process rules 225 and incorporate that rule into implemented process logic 260 for the rules service logic of the workflow service. it is favored to be displayed to the end user."
Regarding claim 4:
Lehr as modified by Lucas teaches “[t]he system of claim 3.”
Lehr further teaches “further comprising a memory storing a plurality of decision logics, wherein each decision logic in the plurality of decision logics has a respective associated state vector and a respective set of transition rules executable by the decision-logic runtime”: Lehr, paragraph 0014, "Similar operations and processes may be performed in a system comprising at least one process and a memory [comprising a memory storing a plurality of decision logics] communicatively coupled to the at least one processor where the memory stores instructions that when executed cause the at least one processor to perform the operations”; Lehr, paragraph 0079, "The process rules 225 may be provided in a selectable form, where the user 205 may select a rule from the process rules 225 and incorporate that rule into implemented process logic 260 for the rules service logic [each decision logic in the plurality of decision logics has a respective associated state vector and a respective set of transition rules executable by the decision-logic runtime] of the workflow service. it is favored to be displayed to the end user."
Regarding claim 20:
Lehr teaches:
“[a] system, comprising”: Lehr, paragraph 0001, “The present disclosure relates to computer-implemented methods, software, and systems [system] for automated data processing in a process management environment.”
(bold only) “an artificial intelligence (AI) model that receives a prompt, and is configured with a decision logic to generate a response to the prompt”: Lehr, paragraph 0004, “One example method may include operations such as extracting a plurality of rules from a machine learning model [an artificial intelligence (AI) model] for prediction [configured with a decision logic to generate a response] of execution results of instances of a process workflow.”
“wherein the decision logic is in a current state; an agent configured to generate a stimulant and further configured to apply the stimulant to the decision logic to transition the decision logic from the current state to a next state”: Lehr, paragraph 0079, "The process rules 225 may be provided in a selectable form, where the user 205 may select a rule from the process rules 225 and incorporate that rule into implemented process logic 260 for the rules service logic of the workflow service [the decision logic is in a current state]. it is favored to be displayed to the end user"; Lehr, paragraph 0052, "Based on computations of metrics for the decision trees and decision rules part of a model from the machine learning models 130, a selection of decision rules may be performed [] and these selected decision rules may be input and implemented as part of process rules 145 to be used during process execution [further configured to apply the stimulant to the decision logic to transition the decision logic from the current state to a next state]."
(bold only) “wherein the decision logic in its next state is used to generate the response to the prompt”: Lehr, paragraph 002, “Software application may execute processes in relation to providing user-requested services. Processes may be defined as workflows that include multiple steps taking input and providing output [the decision logic in its next state is used to generate the response].”
Lehr does not explicitly teach:
(bold only) “an artificial intelligence (AI) model that receives a prompt, and is configured with a decision logic to generate a response to the prompt”
(bold only) “wherein the decision logic in its next state is used to generate the response to the prompt”
Lucas teaches (bold only) “an artificial intelligence (AI) model that receives a prompt, and is configured with a decision logic to generate a response to the prompt” and (bold only) “wherein the decision logic in its next state is used to generate the response to the prompt”: Lucas, paragraph 0037, "An AI model 162 may be a generative AI model, such as a generative language model ( e.g., small language model, large language model, etc., that responds to natural language prompts in natural language)[receives a prompt][generate a response to the prompt], generative image model (e.g., that responds to natural-language prompts with an image), generative video model (e.g., that responds to natural-language prompts with a video), generative coding model (e.g., that responds to natural-language prompts with software code), or the like.”
Lucas and Lehr are analogous arts as they are both related to analyzing process execution. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date of the claimed invention to have combined the prompt response of Lucas with the teachings of Lehr to arrive at the present invention, in order to produce user-readable output, as stated in Lucas, paragraph 0037, “An AI model 162 may be a generative AI model, such as a generative language model ( e.g., small language model, large language model, etc., that responds to natural language prompts in natural language), generative image model (e.g., that responds to natural-language prompts with an image), generative video model (e.g., that responds to natural-language prompts with a video), generative coding model (e.g., that responds to natural-language prompts with software code), or the like. As used herein, the term ‘natural language’ or ‘natural-language’ refers to language, including grammar, that would be expected in a normal conversation between two humans.”
Claims 5, 10–11, and 19 rejected under 35 U.S.C. 103 over Lehr as modified by Lucas in view of Rosenberg.
Regarding claim 5 and analogous claim 19:
Lehr as modified by Lucas teaches “[t]he system of claim 1.”
Lucas further teaches:
(bold only) “(i) a hash of the current state vector, (ii) a hash of the next state vector, and (iii) an identifier of the transition rule applied, wherein the decision trace is stored as a tamper-evident record linked to at least one prior decision trace via a hash chain”: Lucas, paragraph 0014, "The one or more visual elements may comprise a state evolution timeline, wherein the state evolution timeline comprises a timeline and a plurality of points, wherein each of the plurality of points represents a state transition [current state vector][next state vector] and is positioned on the timeline at a location that is representative of a timing of that state transition [an identifier of the transition rule applied] relative to the state transitions represented by other ones of the plurality of points, and wherein each of one or more of the plurality of points are expandable to reveal a state snapshot of the AI agent at the timing of that point.”
“further including logic to generate a decision trace identifying a strategy adopted by the AI model to generate the response, wherein the decision trace further comprises”: Lucas, paragraph 0077, "Subprocess 320, which may be implemented by trace engine 116, may generate a decision trace structure based on the trace in the telemetry data, obtained in subprocess 310. The decision trace structure may represent a decision subset of the plurality of spans, in the trace, that represent decision-making operations performed by AI agent 160 during execution. In other words, the decision trace structure may comprise the spans that have been classified as decision-level [generate a decision trace identifying a strategy adopted by the AI model to generate the response].”
Lucas and Lehr are analogous arts as they are both related to analyzing process execution. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date of the claimed invention to have combined the decision traces of Lucas with the teachings of Lehr to arrive at the present invention, in order to enable better analysis of model output, as stated in Lucas, paragraph 0027, “In particular, platform 110 may execute a server application 112, execute a trace engine 116 that organizes raw trace data into a queryable and hierarchical trace structure for analysis and visualization, and/or host a database 114 that may store data used by server application 112 and/or trace engine 116.”
Lehr as modified by Lucas does not explicitly teach “(i) a hash of the current state vector, (ii) a hash of the next state vector, and (iii) an identifier of the transition rule applied, wherein the decision trace is stored as a tamper-evident record linked to at least one prior decision trace via a hash chain.”
Rosenberg teaches (bold only) “(i) a hash of the current state vector, (ii) a hash of the next state vector, and (iii) an identifier of the transition rule applied, wherein the decision trace is stored as a tamper-evident record linked to at least one prior decision trace via a hash chain”: Rosenberg, “Verification Process” section, “To verify the integrity of a hash chain, you: 1. Start with the first entry and compute its hash [a hash of the next] 2. Compare that hash with the ‘Previous Hash’ stored in the second entry [a hash of the current] 3. If they match, compute the hash of the second entry 4. Compare that with the ‘Previous Hash’ in the third entry 5. Continue this process through the entire chain. If any entry has been modified, the hash comparison will fail, immediately revealing the tampering [wherein the decision trace is stored as a tamper-evident record linked to at least one prior decision trace via a hash chain].”
Rosenberg and Lehr as modified by Lucas are analogous arts as they are both related to record trace chains. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date of the claimed invention to have combined the hash chain of Rosenberg with the teachings of Lehr as modified by Lucas to arrive at the present invention, in order to detect tampering of records, as stated in Rosenberg, “Verification Process” section, “If any entry has been modified, the hash comparison will fail, immediately revealing the tampering.”
Regarding claim 10:
Lehr as modified by Lucas and Rosenberg teaches “[t]he system of claim 5.”
Lucas further teaches “wherein the decision trace comprises a cognitive strategy vector identifying the strategy applied by the decision-logic runtime to generate the response, the cognitive strategy vector including at least one of: a planning weight, an exploration weight, or a constraint-satisfaction weight”: Lucas, paragraph 0012, “The contextual data may comprise a relationship map that represents relationships between operations of the AI agent. The relationships, represented in the relationship map, may comprise temporal relationships, causal relationships, dependency relationships, and semantic relationships. Enriching the hierarchical trace structure with contextual data may comprise generating the relationship map by: deriving a plurality of features from the hierarchical trace structure, wherein the plurality of features comprise one or more temporal features, one or more contextual features, and one or more technical features; applying a plurality of analyses to the plurality of features to identify the relationships between operations of the AI agent [a cognitive strategy vector identifying the strategy applied by the decision-logic runtime to generate the response]; and classifying each of the identified relationships based on type, strength, and impact [the cognitive strategy vector including at least one of: a planning weight, planning weight interpreted as a measure of potential impact].”
Lucas and Lehr are combinable for the rationale given under claim 5.
Regarding claim 11:
Lehr as modified by Lucas and Rosenberg teaches “[t]he system of claim 5.”
Lucas further teaches “wherein the decision trace comprises a causal chain linking the response to one or more prior decision operations that contributed to generation of the response, enabling reinforcement of effective early decisions when a positive outcome is delayed”: Lucas, paragraph 0077, "Subprocess 320, which may be implemented by trace engine 116, may generate a decision trace structure based on the trace in the telemetry data, obtained in subprocess 310. The decision trace structure may represent a decision subset of the plurality of spans, in the trace, that represent decision-making operations performed by AI agent 160 during execution. In other words, the decision trace structure may comprise the spans that have been classified as decision-level [a causal chain linking the response to one or more prior decision operations that contributed to generation of the response, enabling reinforcement of effective early decisions when a positive outcome is delayed].”
Lucas and Lehr are combinable for the rationale given under claim 5.
Rosenberg further teaches “wherein the causal chain includes at least one intermediate state hash for each prior decision operation”: Rosenberg, “Basic Structure” section, “In a hash chain, each entry contains: 1. The event data itself 2. A hash of the previous entry 3. Optionally, a hash of the current entry (for verification) [at least one intermediate state hash for each prior decision operation].”
Rosenberg and Lehr as modified by Lucas are combinable for the rationale given under claim 5.
Claims 6–7 and 17 rejected under 35 U.S.C. 103 over Lehr as modified by Lucas and Rosenberg in view of Chung et al., US Pre-Grant Publication No. 2020/0302273 (hereafter Chung).
Regarding claim 6:
Lehr as modified by Lucas and Rosenberg teaches “[t]he system of claim 5.”
Lehr as modified by Lucas and Rosenberg does not explicitly teach “wherein the decision trace further comprises an associated emotional state inferred from a learning-goal gap determined using the scoring rubric, the associated emotional state comprising at least one of a positive state, an altered state, a neutral state, or a cautious state.”
Chung teaches “wherein the decision trace further comprises an associated emotional state inferred from a learning-goal gap determined using the scoring rubric, the associated emotional state comprising at least one of a positive state, an altered state, a neutral state, or a cautious state”: Chung, paragraph 0083, “Example 1: A computer-implemented method, comprising: following a forward training pass for an artificial neural network (ANN) model, computing a perplexity score for a current training data sample in a training data set, the perplexity score comprising a divergence between data defining a class associated with the current training data sample and a probability vector generated by the ANN model specifying probabilities that the current training data sample belongs to each of a plurality of classes [an associated emotional state inferred from a learning-goal gap determined using the scoring rubric, the associated emotional state comprising at least one of a positive state, an altered state, a neutral state, or a cautious state], determining if the perplexity score is lower than a first threshold, removing the current training data sample from the training data set if the perplexity score is lower than the first threshold, and continuing training of the ANN model following the removal of the current training data sample from the training data set.”
Chung and Lehr as modified by Lucas and Rosenberg are analogous arts as they are both related to the analysis and improvement of models and system. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date of the claimed invention to have combined the measurement and use of the scoring rubric for data used in the decision process of a model, from Chung, with the teachings of Lehr as modified by Lucas and Rosenberg to arrive at the present invention, in order to improve model results, as stated in Chung, paragraph 0007, “Technologies are disclosed herein for subsampling training data during ANN training. Through implementations of the disclosed technologies, ANN models can be trained to a desired level of accuracy using a reduced number of training data samples, thereby reducing training time.”
Regarding claim 7:
Lehr as modified by Lucas, Rosenberg, and Chung teaches “[t]he system of claim 6.”
Chung further teaches “further including logic to increase storage duration for the decision trace when the decision trace comprises a positive associated emotional state and when the learning-goal gap satisfies a reinforcement threshold”: Chung, paragraph 0083, “Example 1: A computer-implemented method, comprising: following a forward training pass for an artificial neural network (ANN) model, computing a perplexity score for a current training data sample in a training data set, the perplexity score comprising a divergence between data defining a class associated with the current training data sample and a probability vector generated by the ANN model specifying probabilities that the current training data sample belongs to each of a plurality of classes, determining if the perplexity score is lower than a first threshold, removing the current training data sample from the training data set if the perplexity score is lower than the first threshold [increase storage duration for the decision trace when the decision trace comprises a positive associated emotional state and when the learning-goal gap satisfies a reinforcement threshold], and continuing training of the ANN model following the removal of the current training data sample from the training data set.”
Chung and Lehr as modified by Lucas and Rosenberg are combinable for the rationale given under claim 6.
Regarding claim 17:
(See the 112(b) rejection, above, for interpretation of terms in this claim).
Lehr as modified by Lucas teaches “[t]he system of claim 4.”
Lehr as modified by Lucas does not explicitly teach “wherein each decision logic in the plurality of decision logics is associated with a learning goal identifier, and wherein one or more decision logics, in the plurality of decision logics, associated with the learning goal identifier are retained in memory and one or more decision logics, in the plurality of decision logics, not associated with the learning goal identifier are subject to a decay function.”
Chung teaches “wherein each decision logic in the plurality of decision logics is associated with a learning goal identifier, and wherein one or more decision logics, in the plurality of decision logics, associated with the learning goal identifier are retained in memory and one or more decision logics, in the plurality of decision logics, not associated with the learning goal identifier are subject to a decay function”: Chung, paragraph 0083, “Example 1: A computer-implemented method, comprising: following a forward training pass for an artificial neural network (ANN) model, computing a perplexity score for a current training data sample in a training data set, the perplexity score comprising a divergence between data defining a class associated with the current training data sample and a probability vector generated by the ANN model specifying probabilities that the current training data sample belongs to each of a plurality of classes, determining if the perplexity score is lower than a first threshold, removing the current training data sample from the training data set if the perplexity score is lower than the first threshold [wherein one or more decision logics, in the plurality of decision logics, associated with the learning goal identifier are retained in memory and one or more decision logics, in the plurality of decision logics, not associated with the learning goal identifier are subject to a decay function], and continuing training of the ANN model following the removal of the current training data sample from the training data set.”
Chung and Lehr as modified by Lucas and Rosenberg are analogous arts as they are both related to the analysis and improvement of models and system. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date of the claimed invention to have combined the measurement and use of the scoring rubric for data used in the decision process of a model, from Chung, with the teachings of Lehr as modified by Lucas and Rosenberg to arrive at the present invention, in order to improve model results, as stated in Chung, paragraph 0007, “Technologies are disclosed herein for subsampling training data during ANN training. Through implementations of the disclosed technologies, ANN models can be trained to a desired level of accuracy using a reduced number of training data samples, thereby reducing training time.”
Claims 8–9 rejected under 35 U.S.C. 103 over Lehr as modified by Lucas and Rosenberg in view of Chung and Lim et al., US Pre-Grant Publication No. 2026/0105118 (hereafter Lim).
Regarding claim 8:
Lehr as modified by Lucas and Rosenberg teaches “[t]he system of claim 5.”
Lehr as modified by Lucas and Rosenberg does not explicitly teach “further including logic to apply an exponential decay function to the decision trace when the decision trace is not associated with a positive outcome under the scoring rubric, wherein decision traces not reinforced by positive outcomes are removed from memory over time.”
Chung teaches (bold only) “further including logic to apply an exponential decay function to the decision trace when the decision trace is not associated with a positive outcome under the scoring rubric, wherein decision traces not reinforced by positive outcomes are removed from memory over time”: Chung, paragraphh0083, “Example 1: A computer-implemented method, comprising: following a forward training pass for an artificial neural network (ANN) model, computing a perplexity score for a current training data sample in a training data set, the perplexity score comprising a divergence between data defining a class associated with the current training data sample and a probability vector generated by the ANN model specifying probabilities that the current training data sample belongs to each of a plurality of classes, determining if the perplexity score is lower than a first threshold, removing the current training data sample from the training data set if the perplexity score is lower than the first threshold [apply an […] decay function to the decision trace when the decision trace is not associated with a positive outcome under the scoring rubric, wherein decision traces not reinforced by positive outcomes are removed from memory over time], and continuing training of the ANN model following the removal of the current training data sample from the training data set.”
Chung and Lehr as modified by Lucas and Rosenberg are analogous arts as they are both related to the analysis and improvement of models and system. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date of the claimed invention to have combined the measurement and use of the scoring rubric for data used in the decision process of a model, from Chung, with the teachings of Lehr as modified by Lucas and Rosenberg to arrive at the present invention, in order to improve model results, as stated in Chung, paragraph 0007, “Technologies are disclosed herein for subsampling training data during ANN training. Through implementations of the disclosed technologies, ANN models can be trained to a desired level of accuracy using a reduced number of training data samples, thereby reducing training time.”
Lim teaches (bold only) “further including logic to apply an exponential decay function to the decision trace when the decision trace is not associated with a positive outcome under the scoring rubric, wherein decision traces not reinforced by positive outcomes are removed from memory over time”: Lim, paragraph 0092, “In some implementations, the ML pipeline may employ implemental training. The implemental training may update model parameters using only new or modified data, which beneficially reduce computational load and deployment latency. The trainer may adjust the learning rates and apply exponential decay to historical data so more recent data trends receive greater weight than less recent trends [an exponential decay function]. This approach avoids retraining the entire model from scratch while still adapting the model to various new data and patterns.”
Lim and Lehr as modified by Chung are analogous arts as they are both related to adapting model updates. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date of the claimed invention to have combined the exponential decay of training data of Lim with the teachings of Lehr as modified by Chung to arrive at the present invention, in order to modify training data slowly over time, as stated in Lim, paragraph 0092, “This approach avoids retraining the entire model from scratch while still adapting the model to various new data and patterns.”
Regarding claim 9:
Lehr as modified by Lucas and Rosenberg teaches “[t]he system of claim 5.”
Lucas further teaches “wherein the decision trace comprises one or more referenced memory identifiers identifying prior decision traces consulted during generation of the response, enabling reconstruction of contextual lineage”: Lucas, paragraph 0011, “The contextual data may comprise a state snapshot of the AI agent at each of one or more points in time during the execution, wherein each state snapshot represents an internal state of the AI agent [prior decision traces consulted during generation of the response, enabling reconstruction of contextual lineage]. Enriching the hierarchical trace structure with contextual data may comprise generating one or more semantics tags for each state snapshot, wherein each state snapshot comprises the one or more semantic tags generated for that state snapshot. The one or more semantic tags may be generated by a Bidirectional Encoder Representations from Transformers (BERT)-based AI model [one or more referenced memory identifiers].”
Lucas and Lehr are combinable for the rationale given under claim 5.
Rosenberg further teaches “and wherein the one or more referenced memory identifiers are validated against the hash chain”: Rosenberg, “Verification Process” section, “To verify the integrity of a hash chain, you: 1. Start with the first entry and compute its hash 2. Compare that hash with the ‘Previous Hash’ stored in the second entry 3. If they match, compute the hash of the second entry 4. Compare that with the ‘Previous Hash’ in the third entry 5. Continue this process through the entire chain. If any entry has been modified, the hash comparison will fail, immediately revealing the tampering [wherein the one or more referenced memory identifiers are validated against the hash chain].”
Rosenberg and Lehr as modified by Lucas are combinable for the rationale given under claim 5.
Claim 12 rejected under 35 U.S.C. 103 over Lehr as modified by Lucas and Rosenberg in view of Yue et al., US Pre-Grant Publication No. 2026/0212217 (hereafter Yue).
Lehr as modified by Lucas and Rosenberg teaches “[t]he system of claim 5.”
Lucas further teaches:
(bold only) “a second AI model configured to generate a second response to the prompt based on a second decision logic executed by a second decision-logic runtime”: Lucas, paragraph 0037, "An AI model 162 may be a generative AI model, such as a generative language model ( e.g., small language model, large language model, etc., that responds to natural language prompts in natural language)[AI model configured to generate a … response to the prompt based on a … decision logic executed by a … decision-logic runtime], generative image model (e.g., that responds to natural-language prompts with an image), generative video model (e.g., that responds to natural-language prompts with a video), generative coding model (e.g., that responds to natural-language prompts with software code), or the like.”
(bold only) “logic to generate a second decision trace identifying a second strategy adopted by the second AI model”: Lucas, paragraph 0077, "Subprocess 320, which may be implemented by trace engine 116, may generate a decision trace structure based on the trace in the telemetry data, obtained in subprocess 310. The decision trace structure may represent a decision subset of the plurality of spans, in the trace, that represent decision-making operations performed by AI agent 160 during execution. In other words, the decision trace structure may comprise the spans that have been classified as decision-level [logic to generate a … decision trace identifying a … strategy adopted by the … AI model].”
Lucas and Lehr are combinable for the rationale given under claim 5.
Rosenberg further teaches (bold only) “the second decision trace including pre-state and post-state hashes”: Rosenberg, “Basic Structure” section, “In a hash chain, each entry contains: 1. The event data itself 2. A hash of the previous entry 3. Optionally, a hash of the current entry (for verification) [the … decision trace including pre-state and post-state hashes].”
Rosenberg and Lehr as modified by Lucas are combinable for the rationale given under claim 5.
Lehr as modified by Lucas and Rosenberg does not explicitly teach:
(bold only) “a second AI model configured to generate a second response to the prompt based on a second decision logic executed by a second decision-logic runtime”
(bold only) “logic to generate a second decision trace identifying a second strategy adopted by the second AI model”
(bold only) “the second decision trace including pre-state and post-state hashes”
“logic to compare the decision trace with the second decision trace to determine which of the response and the second response is more trustworthy, wherein the comparison includes at least one of: (i) a consistency check of the scoring rubric outcomes, or (ii) a divergence check between state hashes under common transition rule identifiers”
Yue teaches:
(bold only) “a second AI model configured to generate a second response to the prompt based on a second decision logic executed by a second decision-logic runtime,” (bold only) “logic to generate a second decision trace identifying a second strategy adopted by the second AI model,” (bold only) “the second decision trace including pre-state and post-state hashes”: Yue, paragraph 0242, “The query response can be generated using one or more models [second AI model][second response][second decision logic]. In an example, multiple models generate candidate responses, wherein a final response is selected from the candidate responses (e.g., based on a confidence score, alignment with the query, etc.). However, a single response can be generated from the aggregated corpus for the query.”
“logic to compare the decision trace with the second decision trace to determine which of the response and the second response is more trustworthy, wherein the comparison includes at least one of: (i) a consistency check of the scoring rubric outcomes, or (ii) a divergence check between state hashes under common transition rule identifiers”: Yue, paragraph 0242, “The query response can be generated using one or more models. In an example, multiple models generate candidate responses, wherein a final response is selected from the candidate responses (e.g., based on a confidence score, alignment with the query, etc.) [wherein the comparison includes at least one of: (i) a consistency check of the scoring rubric outcomes]. However, a single response can be generated from the aggregated corpus for the query.”
Yue and Lehr as modified by Lucas are analogous arts as they are both related to analysis of AI models. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date of the claimed invention to have combined the inter-model comparison of Yue with the teachings of Lehr as modified by Lucas to arrive at the present invention, in order to produce a best result from an ensemble, as stated in Yue, paragraph 0242, “The query response can be generated using one or more models. In an example, multiple models generate candidate responses, wherein a final response is selected from the candidate responses (e.g., based on a confidence score, alignment with the query, etc.). However, a single response can be generated from the aggregated corpus for the query.”
Claims 13–14 rejected under 35 U.S.C. 103 over Lehr as modified by Lucas in view of Zhu et al., US Pre-Grant Publication No. 2026/0300782 (hereafter Zhu).
Regarding claim 13:
Lehr as modified by Lucas teaches “[t]he system of claim 1.”
Lehr further teaches “wherein each stimulant is generated in a typed format comprising the state-transition directive and the scoring rubric”: Lehr, paragraphs 0051–-0052, "In some instances, the models at the machine learning models 130 may be evaluated based on different qualitative and quantitative parameters to evaluate performance of prediction of the generated models. As discussed above, rules extracted from a model from the machine learning models 130 may be evaluated in relation to precision, data coverage, and complexity [scoring rubric]. Based on computations of metrics for the decision trees and decision rules part of a model from the machine learning models 130, a selection of decision rules may be performed and these selected decision rules may be input and implemented as part of process rules 145 to be used during process execution [state-transition directive]."
Lucas further teaches “one or more context-based agents configured to generate context-based stimulants that provide contextual information to the decision-logic runtime”: Lucas, paragraph 0012, “The contextual data may comprise a relationship map that represents relationships between operations of the AI agent. The relationships, represented in the relationship map, may comprise temporal relationships, causal relationships, dependency relationships, and semantic relationships [one or more context-based agents configured to generate context-based stimulants that provide contextual information to the decision-logic runtime]. Enriching the hierarchical trace structure with contextual data may comprise generating the relationship map by: deriving a plurality of features from the hierarchical trace structure, wherein the plurality of features comprise one or more temporal features, one or more contextual features, and one or more technical features; applying a plurality of analyses to the plurality of features to identify the relationships between operations of the AI agent; and classifying each of the identified relationships based on type, strength, and impact.”
Lucas and Lehr are combinable for the rationale given under claim 1.
Lehr as modified by Lucas does not explicitly teach:
“wherein the agent is selected from a plurality of agents”
“one or more role-based agents configured to generate role-based stimulants that probe the decision logic from distinct reasoning perspectives”
Zhu teaches:
“wherein the agent is selected from a plurality of agents”: Zhu, paragraph 0006, “In some aspects, the techniques described herein relate to a method of answering a query with a plurality of artificial intelligence agents including: receiving the query from a client; applying the query to an orchestrator to select a subset of agents to answer the query [wherein the agent is selected from a plurality of agents], the orchestrator applying a relevance ranking algorithm to the query and an agent description set including a description of each agent and a description of a separator class having a description of context outside of a domain of the plurality of artificial intelligence agents; providing the query to each selected agent in the subset of agents; and aggregating responses from each selected agent; and output the aggregated response to the client.”
““one or more role-based agents configured to generate role-based stimulants that probe the decision logic from distinct reasoning perspectives”: Zhu, paragraph 0025, “In some implementations, the orchestrator 142 utilizes a generative AI model 150 to apply a relevance ranking algorithm to the query and an agent description set 152 including a description of each of the agents 160. The description of each agent may include a textual description provided by a source of the agent (e.g., a programmer). In some implementations, the description also includes examples such as example queries and example answers [one or more role-based agents configured to generate role-based stimulants that probe the decision logic from distinct reasoning perspectives].”
Zhu and Lehr as modified by Lucas are analogous arts as they are both related to agentic AI. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date of the claimed invention to have combined the multiple agents of Zhu with the teachings of Lehr as modified by Lucas to arrive at the present invention, in order to use agents more flexibility and with better results, as stated in Zhu, paragraph 0020, “Accordingly, the AI service allows addition of new agents without a need to completely retrain a base model. Further, the base model can be shared among the agents to provide query processing. In an aspect, the AI services aggregates responses from each selected agent into a response to the client. The aggregation may be based on a trained meta-learning decision tree model.”
Regarding claim 14:
Lehr as modified by Lucas and Zhu teaches “[t]he system of claim 13.”
Lehr further teaches:
(bold only) “further comprising an orchestrator agent configured to coordinate the plurality of agents and to combine stimulants generated by the one or more role-based agents and the one or more context-based agents into an ordered stimulus plan for application to the decision-logic runtime”: Lehr, paragraph 0052, "Based on computations of metrics for the decision trees and decision rules part of a model from the machine learning models 130, a selection of decision rules may be performed and these selected decision rules may be input and implemented as part of process rules 145 to be used during process execution [an ordered stimulus plan for application to the decision-logic runtime]."
“wherein the ordered stimulus plan resolves conflicts between transition rules using the scoring rubric”: Lehr, paragraph 0052, "Based on computations of metrics for the decision trees and decision rules part of a model from the machine learning models 130, a selection of decision rules may be performed [the ordered stimulus plan resolves conflicts between transition rules using the scoring rubric, resolves conflicts between transition rules interpreted as including selection of rules] and these selected decision rules may be input and implemented as part of process rules 145 to be used during process execution.”
Zhu further teaches (bold only) “further comprising an orchestrator agent configured to coordinate the plurality of agents and to combine stimulants generated by the one or more role-based agents and the one or more context-based agents into an ordered stimulus plan for application to the decision-logic runtime”: Zhu, paragraph 0006, “In some aspects, the techniques described herein relate to a method of answering a query with a plurality of artificial intelligence agents including: receiving the query from a client; applying the query to an orchestrator to select a subset of agents to answer the query, the orchestrator applying a relevance ranking algorithm to the query and an agent description set including a description of each agent and a description of a separator class having a description of context outside of a domain of the plurality of artificial intelligence agents; providing the query to each selected agent in the subset of agents; and aggregating responses from each selected agent; and output the aggregated response to the client [further comprising an orchestrator agent configured to coordinate the plurality of agents and to combine stimulants generated by the one or more role-based agents and the one or more context-based agents].”
Zhu and Lehr as modified by Lucas are combinable for the rationale given under claim 13.
Claim 15 rejected under 35 U.S.C. 103 over Lehr as modified by Lucas and Zhu in view of Dey et al., US Pre-Grant Publication No. 2024/0161006 (hereafter Dey).
Lehr as modified by Lucas and Zhu teaches “[t]he system of claim 13.”
Lehr further teaches (bold only) “wherein the plurality of agents is configured to iteratively generate stimulants and apply the stimulants to the decision-logic runtime until a convergence condition is satisfied”: Lehr, paragraph 0052, "Based on computations of metrics for the decision trees and decision rules part of a model from the machine learning models 130, a selection of decision rules may be performed and these selected decision rules may be input and implemented as part of process rules 145 to be used during process execution [generate stimulants and apply the stimulants to the decision-logic runtime].”
Lehr as modified by Lucas and Zhu does not explicitly teach (bold only) “wherein the plurality of agents is configured to iteratively generate stimulants and apply the stimulants to the decision-logic runtime until a convergence condition is satisfied, the convergence condition based on one or more of a resource constraint, an iteration threshold, or a quality metric computed from the scoring rubric.”
Dey teaches (bold only) “wherein the plurality of agents is configured to iteratively generate stimulants and apply the stimulants to the decision-logic runtime until a convergence condition is satisfied, the convergence condition based on one or more of a resource constraint, an iteration threshold, or a quality metric computed from the scoring rubric”: Dey, paragraph 01012, “The method may further include iteratively updating a ranking of the agents based on variability of their underlying stochastic processes, sequentially updating local policies of the agents, and simultaneously executing selected actions based on the updated local policies until the ranking of agents does not change between successive iterations [the plurality of agents is configured to iteratively until a convergence condition is satisfied, the convergence condition based on […] a quality metric computed from the scoring rubric].”
Dey and Lehr as modified by Lucas and Zhu are analogous arts as they are both related to agent ensembles. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date of the claimed invention to have combined the iterative training of Dey with the teachings of Lehr as modified by Lucas and Zhu to arrive at the present invention, in order to train until convergence, as stated in Dey, paragraph 0079, “These steps are then repeated until convergence or until the rankings of the agents do not change between iterations.”
Claim 16 rejected under 35 U.S.C. 103 over Lehr as modified by Lucas, Rosenberg and Chung in view of Yoshida et al., “Reinforcement Learning-based Penalty Function Approach for Constrained Optimization,” 2025, 2025 IEEE International Conference on Systems, Man, and Cybernetics (hereafter Yoshida).
Lehr as modified by Lucas, Rosenberg, and Chung teaches “[t]he system of claim 6.”
Chung further teaches (bold only) “further comprising a temperature parameter that modulates behavior of the agent, wherein the temperature parameter is adjusted based on the associated emotional state and the learning-goal gap, and wherein a higher temperature parameter increases exploration in selection of transition rules and a lower temperature parameter enforces constraint-satisfaction in selection of transition rules”: Chung, paragraph 0083, “Example 1: A computer-implemented method, comprising: following a forward training pass for an artificial neural network (ANN) model, computing a perplexity score for a current training data sample in a training data set, the perplexity score comprising a divergence between data defining a class associated with the current training data sample and a probability vector generated by the ANN model specifying probabilities that the current training data sample belongs to each of a plurality of classes [an the associated emotional state and the learning-goal gap], determining if the perplexity score is lower than a first threshold, removing the current training data sample from the training data set if the perplexity score is lower than the first threshold, and continuing training of the ANN model following the removal of the current training data sample from the training data set.”
Chung and Lehr as modified by Lucas and Rosenberg are combinable for the rationale given under claim 6.
Lehr as modified by Lucas, Rosenberg, and Chung does not explicitly teach (bold only) “further comprising a temperature parameter that modulates behavior of the agent, wherein the temperature parameter is adjusted based on the associated emotional state and the learning-goal gap, and wherein a higher temperature parameter increases exploration in selection of transition rules and a lower temperature parameter enforces constraint-satisfaction in selection of transition rules.”
Yoshida teaches (bold only) “further comprising a temperature parameter that modulates behavior of the agent, wherein the temperature parameter is adjusted based on the associated emotional state and the learning-goal gap, and wherein a higher temperature parameter increases exploration in selection of transition rules and a lower temperature parameter enforces constraint-satisfaction in selection of transition rules”: Yoshida, section I. A. 1, “To overcome these difficulties, research has been conducted on dynamically adjusting penalty parameters [a temperature parameter that modulates behavior of the agent]. One such approach introduces a two-stage penalty adjustment strategy [13]. In this approach, the initial stage allows a certain degree of constraint violations to encourage broad exploration across the search space. This is based on the insight that ensuring search diversity in the early phase of a metaheuristic algorithm [wherein a higher temperature parameter increases exploration in selection of transition rules] helps prevent premature convergence to local optima and increases the likelihood of finding better global solutions. In contrast, during the latter stage, the penalty becomes stricter to guide the search toward the feasible region, thereby improving convergence efficiency [a lower temperature parameter enforces constraint-satisfaction in selection of transition rules]. This dynamic adjustment aims to balance exploration and exploitation, maintaining flexibility in the search while accelerating convergence toward the optimal solution.”
Yoshida and Lehr as modified by Lucas, Rosenberg, and Chung are analogous arts as they are both related to model optimization. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date of the claimed invention to have combined the constraint relaxation of Yoshida with the teachings of Lehr as modified by Lucas, Rosenberg, and Chung to arrive at the present invention, in order to balance exploration and exploitation, as stated in Yoshida, section I. A. 1, “This dynamic adjustment aims to balance exploration and exploitation, maintaining flexibility in the search while accelerating convergence toward the optimal solution.”
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Nandy, US Pre-Grant Publication No. 2008/0109392, discloses a method of evaluating a set of rules using a scoring rubric, in which rules are stored in a hash table.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to VINCENT SPRAUL whose telephone number is (703) 756-1511. The examiner can normally be reached M-F 9:00 am - 5:00 pm.
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, MICHAEL HUNTLEY can be reached at (303) 297-4307. 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.
/VAS/ Examiner, Art Unit 2129
/MICHAEL J HUNTLEY/Supervisory Patent Examiner, Art Unit 2129