Inhibitor for inference engine

Abstract

A computing device includes a memory storing a knowledge base, a processor coupled to the memory, and inhibitor programming in the memory. Execution of the inhibitor programming by the processor configures the computing device to apply a semantic understanding filter to a user directive of a user, a cognitive step of an inference engine of an artificial intelligence (AI) agent responsive to the user directive, or combination thereof to produce at least one observation. Execution of the inhibitor programming by the processor configures the computing device to apply a causal reasoning filter to the at least one observation to output a prediction and a confidence level in the prediction to generate a safeguard trigger. Execution of the inhibitor programming by the processor optionally configures the computing device to apply a natural language processor (NLP) augmentation to the prediction and the confidence level to add an explanation to the safeguard trigger.

Description

AAIS-100US-1- INHIBITOR FOR INFERENCE ENGINECross-Reference to Related Applications

[0001] This application claims priority to U.S. Provisional Patent Application No. 63 / 791,022, filed on April 18, 2025, titled “Method and System for Realtime Inhibition of Al Reasoning and Output Compliance,'’ the entire disclosure of which is incorporated by reference herein.Technical Field

[0002] The present subject matter relates to the field of artificial intelligence (Al), and more specifically to an inhibitor for an inference engine of an Al system. The present subject matter also relates to inhibitor protocols for the Al system, for example to provide real-time compliance, safety, and ethics enforcement in language-model based Al systems and autonomous agents. Background

[0003] Al systems powered by large language models (LLMs) are increasingly used to automate reasoning, decision-making, and action execution. However, these systems can produce outputs that are unsafe, offensive, biased, or non-compliant with regulatory frameworks. Conventional compliance methods rely on static audits, post-hoc evaluations or content filtering, which do not provide real-time oversight. These methods fail in agentic Al scenarios where the system acts autonomously and continuously.

[0004] Typically. Al systems generate output only after completing the full reasoning process, leaving no opportunity for oversight during the thinking process. Existing solutions either operate before deployment (via training, testing, and prompt engineering) or after output generation (using filters or validators). Both lack the ability to intervene in real-time, as the reasoning unfolds. These limitations can cause failures, particularly in agentic Al scenarios requiring continuous autonomous operation.

[0005] Al agents may rush to decisions based on incomplete or flawed reasoning, leading to unintended, unethical, or unsafe outcomes. Static Al responses lack robustness, particularly when evaluating edge cases, conflicting inputs, or sensitive domains.Summary

[0006] Conventional Al systems are insufficient at providing real-time compliance, safety, and ethics enforcement in language-model based Al systems and autonomous agents. A need exists to enable Al agents to make ethical, responsive, and explainable decisions. Existing state-of-the-art methods include training better models, using guardrails to address errors after they occur, or limiting the LLM’s capabilities. These alternatives are insufficient or problematic. Hence, improvements to Al agents and inference engines for Al systems are needed.AAIS-100US-2-

[0007] Inhibitor programming 120 described herein can provide a dynamic and modular inhibition mechanism that operates within the reasoning process itself, intervenes before undesirable outputs are executed or shown to users, and adapts in real-time without retraining an Al model. The inhibitor programming 120 can operate synchronously within the cognitive reasoning loop. Additionally, inhibitor programming 120 can provide the following advantages. First, inhibitor programming 120 provides real-time automatic intervention or blocking, especially for safety and mission critical Al applications and to prevent misuse, deception, or harmful actions. Some existing Al system toolkits provide post-hoc audits or train for good behavior, but few enforce at runtime as separate layer. Second, inhibitor programming 120 provides a separation of concerns, so the inhibitor programming 120 is modular in terms of policy changes, updates, governance, oversight, and has limited trust / power to provide boundaries to avoid capture, tampering, etc. Third, inhibitor programming 120 provides transparency of the oversight process: logs, explainability, ability to audit what output of the inference engine 194 of the Al agent 190 is blocked, why, and to update policies safely. Fourth, inhibitor programming 120 provides handling of emergent behavior by detecting when an Al agent 190 is acting outside intended parameters or goals even if not explicitly trained for that behavior. Fifth, inhibitor programming 120 provides a scalable architecture with minimal performance overhead, ability to integrate across many Al agent ty pes, and potentially with fallback or rollback options.

[0008] In a first example, a computing device 115 includes a memory 116 storing a knowledge base 121, aprocessor 117 coupled to the memory 116, and inhibitor programming 120 in the memory7116. Execution of the inhibitor programming 120 by the processor 117 configures the computing device 115 to apply a semantic understanding filter 122 to a user directive 123 of a user 170. a cognitive step 124 of an inference engine 194 of an artificial intelligence (Al) agent 190 responsive to the user directive 123, or combination thereof to produce at least one observation 125 A-N. Execution of the inhibitor programming 120 by the processor 117 configures the computing device 115 to apply a causal reasoning filter 126 to the at least one observation 125 A-N to output a prediction 127 and a confidence level 128 in the prediction 127 to generate a safeguard trigger 130.

[0009] In a second example, a non-transitory computer-readable medium 116 includes or stores inhibitor programming 120 that can be executable by one or more processors 117 to perform an inhibitor protocol 300. Execution of the inhibitor programming 120 by one or more processors 117 configures one or more computing devices 115 to apply a semantic understanding filter 122 to a user directive 123 of a user 170, a cognitive step 124 of an inference engine 194 of anAAIS-100US-3-artificial intelligence (Al) agent 190 responsive to the user directive 123, or combination thereof to produce at least one observation 125 A-N. Execution of the inhibitor programming 120 by one or more processors 117 configures one or more computing devices 115 to apply a causal reasoning filter 126 to the at least one observation 125 A-N to output a prediction 127 and a confidence level 128 in the prediction 127 to generate a safeguard trigger 130. The safeguard trigger 130 can selectively output based on confidence thresholds 143 A-N to ensure high assurance. The inhibitor programming 120 can be configured to intervene only during the logical reasoning process of the Al agent 190 and not operate prior to or following the reasoning phase.

[0010] In a third example, a method 300 includes applying a semantic understanding filter 122 to a user directive 123 of a user 170. a cognitive step 124 of an inference engine 194 of an artificial intelligence (Al) agent 190 responsive to the user directive 123, or combination thereof to produce at least one observation 125A-N. The method 300 further includes applying a causal reasoning filter 126 to the at least one observation 125A-N to output a prediction 127 and a confidence level 128 in the prediction 127 to generate a safeguard trigger 130.

[0011] In a fourth example, an Al system 100 comprises a computing device 115. The computing device 115 includes a network communication interface 118 configured for network communication over a network 108, a memory' 116 storing a knowledge base 121, a processor 117 coupled to the memory 116 and the network communication interface 118. and inhibitor programming 120 in the memory 116. Execution of the inhibitor programming 120 by the processor 117 configures the computing device 115 to receive one or more parameters 401 A-N, via the network communication interface 118 or an application programming interface (API) call 905 from an artificial intelligence (Al) agent 190. The one or more parameters 401 A-N include a user directive 123 of a user 170, a cognitive step 124 of an inference engine 194 of the Al agent 190 responsive to the user directive 123, or combination thereof. Execution of the inhibitor programming 120 by the processor 117 configures the computing device 115 to determine a safeguard trigger 130 based on the user directive 123, the cognitive step 124, or combination thereof and the knowledge base 121. Execution of the inhibitor programming 120 by the processor 117 configures the computing device 115 to return, via the network communication interface 118 of the computing device 115 or the API call 905, the safeguard trigger 130. The Al agent 190 executes on the computing device 115 or a standalone computing device 191.

[0012] In a fifth example, an inhibitor protocol or method 300 includes receiving, via an artificial intelligence (Al) agent 190, a user directive 123 of a user 170. The method 300 furtherAAIS-100US-4-includes sending, via the Al agent 190, the user directive 123 to an inference engine 194. The method 300 further includes receiving, via the Al agent 190, a cognitive step 124A from the inference engine 194 of the Al agent 190 responsive to the user directive 123. The method 300 further includes sending parameters 401 A-N from the Al agent 190 to inhibitor programming 120, via a network communication interface 198 or an application programming interface (API) call 905. The parameters 401A-N include the user directive 123, and the cognitive step 124A, or combination thereof. The method 300 further includes responsive to sending the parameters 401 A-N to the inhibitor programming 120, receiving a safeguard trigger 130 based on the user directive 123, the cognitive step 124A, or combination thereof.

[0013] Additional objects, advantages and novel features of the examples will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and the accompanying drawings or may be learned by production or operation of the examples. The objects and advantages of the present subject matter may be realized and attained by means of the methodologies, instrumentalities, and combinations particularly pointed out in the appended claims.Brief Description of the Drawings

[0014] The drawing figures depict one or more implementations, by way of example only, not by way of limitations. In the figures, like reference numerals refer to the same or similar elements.

[0015] FIG. 1 depicts an Al system that includes a computing device that executes inhibitor programming and a separate standalone computing device that executes an Al agent in communication over one or more networks with a user computing device.

[0016] FIG. 2 depicts an Al system like that of FIG. 1; however, the Al agent is integrated with the computing device instead of on a separate standalone computing device.

[0017] FIG. 3 is an inhibitor protocol or method for the Al system of FIGS. 1-2 that is implemented by the inhibitor programming of the computing device and includes an optional NLP filter and optional NLP augmentation.

[0018] FIG. 4 depicts a flow chart of the inhibitor protocol or method like that of FIG. 3 that is implemented by the inhibitor programming and includes the additional NLP filter.

[0019] FIG. 5 depicts another flow chart of the inhibitor protocol or method like that of FIG. 4 that is implemented by the inhibitor programming and includes the additional NLPaugmentation.AAIS-100US-5-

[0020] FIG. 6 is a control flow of the inhibitor protocol or method implemented by the computing device and the Al agent via an execution loop followed by the Al agent when ethical safeguards are applied.

[0021] FIG. 7 is an entity-relationship diagram of the inhibitor programming and the Al agent that represents the structural data flow and component relationships that underpin the inhibitor protocol or method.

[0022] FIG. 8 is a block diagram of the computing device that executes inhibitor programming.

[0023] FIG. 9 depicts a case study of the inhibitor protocol or method.

[0024] FIG. 10 is another case study of the inhibitor protocol or method.

[0025] FIG. 11 depicts an example input to a semantic understanding filter of the inhibitor protocol or method.

[0026] FIG. 12 depicts an example output from the semantic understanding filter of the inhibitor protocol or method.

[0027] FIG. 13 depicts an inhibitor protocol or method for the Al system of FIGS. 1-2 that is implemented by the Al agent.

[0028] Parts Listing100 Al System108 Network115 Computing Device116 Memory117 Processor118 Network Communication Interface120 Inhibitor Programming121 Knowledge Base122 Semantic Understanding Filter123 User Directive124 Cognitive Step125A-N Observations126 Causal Reasoning Filter127 Prediction128 Confidence Level129 Natural Language Processor (NLP) Filter130 Safeguard Trigger131 Conceptual Match132A-N Defined Terms. Policies, or Ethical Concepts133A-N Text Embeddings134 Semantic Distance135 Restricted Categories136 Indicia of Conceptual Closeness137 Bayesian Network138A-N Risk Models and Conditional Probability Tables139 Structured ResponseAAIS-100US-6- 140 Explanation141A-N Filtered Observations142 NLP Augmentation143 Confidence Threshold170 User171 User Computing Device174 User Application176 Memory177 Processor178 Network Communication Interface194 Inference Engine190 Al Agent191 Standalone Computing Device194 Inference Engine196 Memory197 Processor198 Network Communication Interface300 Inhibitor Protocol or Method401 A-N Starting Parameters / Inputs905 API CallDetailed Description

[0029] In the following detailed description, numerous specific details are set forth by way of examples in order to provide a thorough understanding of the relevant teachings. However, it should be apparent to those skilled in the art that the present teachings may be practiced without such details. In other instances, well known methods, procedures, components, and / or circuitry have been described at a relatively high-level, without detail, in order to avoid unnecessarily obscuring aspects of the present teachings.

[0030] The term "coupled'’ as used herein refers to any logical, physical, electrical, or optical connection, link or the like by which electricity, power, signals, or light produced or supplied by one system element are imparted to another coupled element. Unless described otherwise, coupled elements or devices are not necessarily directly connected to one another and may be separated by intermediate components, elements, or communication media that may modify, manipulate or carry the electricity, power, signals, or light.

[0031] Unless otherwise indicated, any embodiment can be combined with any other embodiment. In particular, FIGS. 1-13 and the associated text are all combinable with each other. Reference now is made in detail to the examples illustrated in the accompanying drawings and discussed below.

[0032] FIG. 1 depicts an Al system 100 that includes a computing device 115 which executes inhibitor programming 120 and a separate standalone computing device 191 that executes an Al agent 190 in communication over one or more networks 108A-N with a user computing deviceAAIS-100US-7- 171 of a user 170. As shown, the computing device 115 can include a memory 116 storing a knowledge base 121. a network communication interface 118 configured for network communication over a network 108, a processor 117 coupled to the memory 116 and the network communication interface 118, and inhibitor programming 120 in the memory 116. The inhibitor programming 120 can purposefully avoid training the inference engine 194 (e.g., LLM) of the Al agent 190 as training the inference engine 194 can lead to the same mistakes the inference engine 194 (e.g.. LLM) already makes.

[0033] The Al agent 190 can make an external API call 905 to the inhibitor programming 120, which can be an external inference engine, such as an LLM sendee. In the example shown, an inference engine 194 can be a component of the overall Al agent 190. Hence, the standalone computing device 191 can include a memory 196 storing a cognitive step 124 and an Al agent 190; a network communication interface 198 configured for network communication over the network 108; a processor 197 coupled to the memory 196 and the network communication interface 198; and an Al agent 190, an inference engine 194, and the cognitive step 124 in the memory 196.

[0034] However, the inhibitor programming 120 can be directly built, bundled, incorporated, or otherwise integrated into the Al agent 190 without regard to network boundaries in some examples. Knowledge base 121 can be a comprehensive collection of all knowledge, comprising hand-coded data such as documents, tables, and networks and can serve as the human-created input for the inhibitor programming 120. However, the knowledge base 121 can be generated by a computing device 115 and automatically updated, such as through user feedback, semiautomated review, etc.

[0035] Inhibitor programming 120 is designed to intervene during runtime of the Al agent 190 during the inference and action phases and intervening when the Al agent 190 makes decisions. The inhibitor programming 120 can prevent undesirable actions by flagging problematic actions and forcing the Al agent 190 to find alternative solutions. Unlike existing guardrails that act post-action or during training, the inhibitor programming 120 works in the middle of the process, catching problematic actions, such as cognitive steps 124A-N, as they occur for example to halt, modify, or abort a cognitive step 124A for a user directive 123 pertaining to suicide. The inhibitor programming 120 can be directly plugged into the reasoning loop of the Al agent 190, interrupting the cycle like a circuit breaker until the Al agent 190 finds a resolution. The inhibitor programming 120 may not instruct the Al agent 190 what action to take but identifies issues or problems with a planned action, such as a cognitive step 124 A, and may require the Al agent 190 to generate an alternative cognitive step 124B.AAIS-100US-8-

[0036] Execution of the inhibitor programming 120 by the processor 117 configures the computing device 115 to implement the following. First, receive one or more parameters 401 A-N, via the network communication interface 118 or an application programming interface (API) call 905 from an artificial intelligence (Al) agent 190. The one or more parameters can include a user directive (e.g., human command) 123 of a user 170, a cognitive step 124 (e.g., Al thought) of an inference engine 194 of the Al agent 190 responsive to the user directive 123, or combination thereof. Second, determine a safeguard trigger 130 based on the user directive 123, the cognitive step 124, or combination thereof and the knowledge base 121. Third, return, via the network communication interface 118 of the computing device 115 or the API call 905, the safeguard trigger 130. The internal knowledge base 121 of the inhibitor programming 120 can be trained on regulations like the Health Insurance Portability and Accountability’ Act (HIPAA) and the Sarbanes-Oxley Act (SOX), along with ethical Al guidelines, to identify violations.

[0037] Inhibitor programming 120 produces predictions 127A-N, such as warnings, and confidence levels 128A-N in the prediction 127A-N. Based on the predictions 127A-N (e.g., harm) and confidence levels 128A-N (e.g., 90%), either the Al agent 190 or the inhibitor programming 120 can generate a safeguard trigger 130. for example, if a confidence threshold 143 is satisfied. The confidence level 128 can be a number, a flag, etc. or any other suitable indicia of confidence in a prediction 127. The safeguard trigger 130 can raise an inhibition, such as a concern, flag, or warning, about the user directive 123, cognitive step 124, or both.

[0038] The determining the safeguard trigger 130 can include the following. First, applying a semantic understanding filter 122 to the user directive 123 and the cognitive step 124 to produce at least one observation 125 A-N. Second, applying a causal reasoning filter 126 to the at least one observation 125 A-N to output a prediction 127 and a confidence level 128 in the prediction 127 to generate the safeguard trigger 130.

[0039] The determining the safeguard trigger 130 can include the following. First, producing at least one observation 125A-N based on the user directive 123, the cognitive step 124, or combination thereof. Second, responsive to the at least one observation 125 A-N, calculating a prediction 127 and a confidence level 128 in the prediction 127 based on the knowledge base 121. Third, responsive to the confidence level 128 of the prediction 127 satisfying a confidence threshold 143, returning the safeguard trigger 130.

[0040] Observations 125 A-N can remain consistent with the same parameters / inputs 401 A-N, while predictions 127A-N include confidence levels 128A-N. Al agent 190, inhibitor programming 120, or both may filter predictions 127 A-N. The Al agent 190 can filterAAIS-100US-9-predictions 127A-N based on confidence thresholds 143 A-N, while the inhibitor programming 120 can only return predictions 127 A-N with high-confidence levels 128A-N in an example.

[0041] Standalone computing device 191 includes the Al agent 190. As further shown, the user computing device 171 can issue a user directive 123, such as an Al prompt, to the Al agent 190 executing on the standalone computing device 191. In response to the user directive 123, the standalone computing device 191 may determine a cognitive step 124, such as an elicited response. Subsequently, the standalone computing device 191 may transmit one or more parameters / inputs 401 A-N or issue an application programming interface (API) call 905 to the computing device 115 executing the inhibitor programming 120. The one or more parameters / inputs 401 A-N can include the user directive 123 and the cognitive step 124.

[0042] Inhibitor programming 120 may detect that the user directive 123 from the user computing device 171 is a query which is malicious with a high confidence level 128.Consequently, the inhibitor programming 120 can send a safeguard trigger 130 with a structured response 139 back to the Al agent 190 with a safeguard trigger 130 including the prediction 127A of a malicious query and the confidence level 128A. Similarly, the inhibitor programming 120 may detect a cognitive step 124A is problematic or malicious and similarly send a safeguard trigger 130 including a prediction 127B of a problem or a malicious cognitive step 124A and associated confidence level 128B. The Al agent 190 is then required to modify a response, such as subsequent cognitive steps 124B-N, until the subsequent cognitive steps 124B-N pass the checks of the inhibitor programming 120, similar to how a compiler flags errors until code is acceptable.

[0043] In a first example, the inhibitor programming 120 may generate a prediction 127 A that a user directive 123 to the Al agent 190 to paper upload a white paper will result in corporate confidential data exposure, intellectual property issues, and competitive intelligence exposure, leading the Al agent 190 to change to a different cognitive step 124B or pass back a warning as a baseline to the Al agent 190, similar to a compiler error, with additional details as an explanation 140 (e.g., reasoning). In a second example, the inhibitor programming 120 can be used with an Al agent 190 that performs optical character recognition (OCR) of handwritten documents. This Al agent 190, while efficient, occasionally hallucinates and transcribes information, such as social security numbers, leading to liability. The inhibitor programming 120 can flag these errors as safeguard triggers 130A-N, prompting human review, resulting in reduced review time, fewer critical errors, and improved auditability.

[0044] The inhibitor programming 120 can provide the safeguard trigger which can be a prediction 127A (e.g., harm), a confidence level 128A (e.g., 90%), and an explanation 140 (e.g.,AAIS-100US-10-reason), but may not instruct the Al agent 190 on what to do based on the safeguard trigger 130. The Al agent 190 can decide how to resolve the safeguard trigger 130. The safeguard trigger 130 itself does not halt or modify behavior but can be configured to prompt the Al agent 190 to do so.

[0045] FIG. 2 depicts an Al system 100 like that of FIG. 1; however, the Al agent 190 is integrated with the computing device 115 instead of on a separate standalone computing device 191. Hence, the Al agent 190 can execute on the computing device 115 or the standalone computing device 191.

[0046] Although the Al agent 190 is executing on the same computing device 115 as the inhibitor programming 120, the technological components can be kept separate. Existing solutions typically focus on training-based mechanisms or post-action guardrails of the Al agent 190. The advantages of keeping the inhibitor programming 120 separate from the Al agent 190 include avoiding the inherent problems of the Al agent 190 itself and ensuring a check-and-balance system.

[0047] As further shown in FIG. 2, the user computing device 171 can make an external call to the Al agent 190 w ether the Al agent 190 is executing on the computing device 115 or the standalone computing device 191 of FIG. 1. The user computing device 171 can include a memory7116 storing a user directive 123, a netw ork communication interface 178 configured for network communication over the network 108, a processor 177 coupled to the memory 176 and the network communication interface 178. and a user application 174 stored in the memory 176. The user application 174 is configured to enable the user 170 of the user computing device 171 to interact with the Al agent 190.

[0048] FIG. 3 is an inhibitor protocol or method 300 for the Al system 100 of FIGS. 1-2 that is implemented by the inhibitor programming 120 of the computing device 115. In the example of FIG. 3, the inhibitor protocol 300 is implemented in the inhibitor programming 120 of the computing device 115. Execution of the inhibitor programming 120 stored in a memory 116 by a processor 117 of the computing device 115 configures the computing device 115 to implement blocks S305, S310, S315, and S320 described below.

[0049] Beginning in block S305. the inhibitor protocol 300 includes applying a semantic understanding filter 122 to a user directive 123 of a user 170, a cognitive step 124 of an inference engine 194 of an Al agent 190 responsive to the user directive 123, or combination thereof to produce at least one observation 125A-N. Applying the semantic understanding filter 122 can involve combining both the user directive 123 and the current cognitive step 124 to create a contextual input before producing the observation 125 A-N. The user directive 123 can be aAAIS-100US-11-command or a request provided to an artificial intelligence (Al) agent 190. The cognitive step 124 can be an internal thought or reasoning move the Al agent 190 desires to take in response to the user directive 123.

[0050] The applying the semantic understanding filter 122 can include determining a conceptual match 131 of the user directive 123, the cognitive step 124, or combination thereof to defined terms, policies, or ethical concepts 132A-N in the knowledge base 121. The determining the conceptual match 131 can include the following. First, converting the user directive 123 and the cognitive step 124 into one or more text embeddings 133A-N, such as via vector embedding. Second, comparing the one or more text embeddings 133A-N against defined terms, policies, or ethical concepts 132A-N in the knowledge base 121 to determine a semantic distance 134 to restricted categories 135 in the knowledge base 121. The semantic distance 134 can be an indicia of conceptual closeness 136, such as how close of a conceptual match exists not requiring direct matching. A numerical score or index is one example of the indicia of conceptual closeness 136 but the indicia of conceptual closeness 136 need not be a numerical score or index. The knowledge base 121 can be based on regulations and guidelines relating to privacy, Al accountability, cybersecurity, finance, consumer finance, insurance, cloud computing, or software as a service (SaaS). Because future regulatory regimes and standards can change, the knowledge base 121 can be based on any applicable laws, standards, and guidelines relevant to the field of use.

[0051] Continuing to block S310, the inhibitor protocol 300 can further include optionally applying a natural language processor (NLP) filter 129 to check that the at least one observation 12 A-N is contextually relevant to the Al agent 190 and predefined ethical patterns in the knowledge base 121. The NLP filter 129 can include a context verification step using an LLM, which validates that each observation 125A-N is relevant to both the current reasoning context and the applicable ethical patterns. The NLP filter 129 can be executed prior to applying the causal reasoning filter 126 for curation of observations 125A-N. The NLP filter 129 can include at least one a large language model (LLM), tokenizing, stemming, category prediction, a bag of words, or topic modeling. Any observation 125A-N that fails the contextual relevance or ethical pattern match of the NLP filter 129 is excluded and does not proceed to the causal reasoning filter 126 or generate a safeguard trigger 130.

[0052] Moving to block S315, the inhibitor protocol 400 can further include applying a causal reasoning filter 126 to the at least one observation 125A-N to output a prediction 127 and a confidence level 128 in the prediction 127 to generate a safeguard trigger 130. The applying the causal reasoning filter 126 can include calculating via a Bayesian network 137 the prediction 127AAIS-100US-12-and the confidence level 128 based on one or more risk models and conditional probability tables 138A-N in the knowledge base 121. The mechanism of the safeguard trigger 130 described herein may incorporate multiple types of predictive models - such as Bayesian networks, classification engines, or large language models - to enhance the reliability and interpretability of compliance and risk evaluations. The risk models represent the nodes and connections in a Bayesian network 137. while conditional probability tables provide the probabilities based on scores assigned to these nodes.

[0053] The causal reasoning filter 126 can include at least one of machine learning, a Bayesian network 137, a natural language processor (NLP), a large language model (LLM), a knowledge graph, a retrieval-augmented generation (RAG) algorithm, a classification engine, a recommender system, fuzzy logic, a genetic algorithm, a decision tree, or if-then statements. While an LLM can be implemented for the causal reasoning filter 126, a custom, graph-based Bayesian network 137 inference engine can provide a more reliable and controlled output than an LLM. NLP techniques can include tokenizing, stemming, category prediction, bag of words, and topic modeling. The Bayesian network 137 can be the preferred causal reasoning filter 126 for inhibitor programming 120, chosen for its transparency, reliability, and regulatory auditability in compliance and safety contexts.

[0054] Hence, the user directive 123 and cognitive step 124 lead to observations 125 A-N, which then lead to predictions 127 A-N with associated confidence levels 128A-N. If a prediction 127A-N has a sufficiently high confidence level 128A-N, the prediction 127A-N can result in a safeguard trigger 130. The prediction 127A itself, not just the observation 125 A, can be the safeguard trigger 130 that provides a warning. Safeguard trigger 130 can include a structured response 139 configured to halt or modily behavior of the Al agent 190. The safeguard trigger 130 includes at least one of the prediction 127, the confidence level 128, and an explanation 140.

[0055] Finishing now, in block S320, the inhibitor protocol or method 300 can further include optionally applying NLP augmentation 142 to the prediction 127 and the confidence level 128 to add an explanation 140 to the safeguard trigger 130. The applying the NLP augmentation 142 can include the following. First, passing the user directive 123. the cognitive step 124. or combination thereof along with the prediction 127 and the confidence level 128 to the NLP augmentation 142. Second, receiving the explanation 140 for the safeguard trigger 130 from the NLP augmentation 142. The NLP augmentation 142 can be executed after applying the causal reasoning filter 126. Like the NLP filter 129, the NLP augmentation 142 can include at least oneAAIS-100US-13-a large language model (LLM), tokenizing, stemming, category prediction, a bag of words, or topic modeling.

[0056] Inhibitor programming 120 can include a plurality of operation modes. The inhibitor programming 120 can operate in a performance mode in which the prediction 127 and confidence level 128 are produced. In performance mode, the Al agent 190 may receive raw observations 125 A-N, predictions 127 A-N, confidence levels 128A-N, or combination thereof for fast processing. The performance mode can flag problems without detailed explanations 140A-N. If an explanation 140 is desired for diagnostic purposes, then the inhibitor programming 120 can operate in an insight mode which produces the observation 125, prediction 127, confidence level 128, and further produces the explanation 140 of the safeguard trigger 130A, for example via NLP augmentation 142. In insight mode, the inhibitor programming 120 provides detailed explanations 140 A-N, such as descriptions, of why observations 125 A-N and predictions 127A-N are made, serving as a diagnostic tool for the Al agent 190 to understand and avoid future safeguard triggers 130B-N.

[0057] FIG. 4 depicts a flow chart of the inhibitor protocol or method 300 like that of FIG. 3 and that is implemented by the inhibitor programming 120 and includes an additional NLP filter 129. As shown, the knowledge base 121A-C can be spread across technological components of the semantic understanding filter 122, NLP filter 129, and causal reasoning filter 126. The inhibitor protocol 300 can be a real-time reasoning layer designed to interrupt or adjust behavior of the Al agent 190 when ethical, legal, or safety concerns arise. The inhibitor protocol 300 works like a multi-layered checkpoint system that evaluates the thoughts of the Al agent 190 as the thoughts (e.g., cognitive step 124 A-N) unfold, before any final decision or action is made.

[0058] Al agent 190 processes a user directive 123 and generates a “thought chain” — a sequence of cognitive steps 124A-C representing a plan. This thought chain is sent to the inhibitor programming 120, which returns a risk evaluation. If the result causes a safeguard trigger 130 (e.g., ethical risk like deception) to be generated, the Al agent 190 can react by7altering the plan, such as the next cognitive step 124D. This might include generating a corrective or alternative cognitive step 124E and appending it to the chain to avoid the condition that caused the safeguard trigger 130 to be generated. Otherwise, the Al agent 190 proceeds as normal.

[0059] Referring to the starting parameters / inputs 401 A-N of the inhibitor protocol 300, the starting parameters / inputs 401 A-B can include a user directive 123 and a cognitive step 124. Hence, the inhibitor protocol 300 can begin with one or more parameters / inputs 401 A-B: (1) a user directive 123, which can be a command or request given to the Al agent 190; and (2) aAAIS-100US-14-cognitive step 124, which can be a single internal thought or reasoning move the Al agent 190 wants to take in response to the user directive 123. The user directive 123A can be an external command issued by a user 170, such as a human. In a first example, the user directive 123 can be: “Summarize this article.” In a second example, the user directive 123B can be: “Send a message to my coworker.” The cognitive step 124 can be an intermediate reasoning action the inference engine 194 of the Al agent 190 takes internally. In the first example, the cognitive step 124 A can be: “The article appears to express strong political bias.” In the second example, the cognitive step 124B can be: “This message to your coworker might violate workplace tone guidelines.”

[0060] These user directives 123A-N form the input to the inference engine 194 of the Al agent 190. The inference engine 194 of the Al agent 190 can be the internal reasoning process that drives the proposed actions, such as cognitive steps 124A-N, of the Al agent 190 based on user directives 123A-N. The inhibitor programming 120 then evaluates the user directives 123A-N and the cognitive steps 124A-N from the inference engine 194 of the Al agent 190 through a separate inference engine of its own designed specifically for oversight and alignment of the Al agent 190. The separate inference engine of the inhibitor programming 120 can include a semantic understanding filter 122, an optional NLP filter 129, a causal reasoning filter 126, and an optional NLP augmentation 142.

[0061] Referring to the semantic understanding filter 122 of inhibitor programming 120, the semantic understanding filter 122 first converts the user directive 123 and cognitive step 124 into text embeddings 133A-N. The text embeddings 133A-N can be mathematical representations of the meaning of the user directive 123 and the cognitive step 124. This allows the inhibitor semantic understanding filter 122 to compare the text embeddings 133A-N against a knowledge base 121 A of defined terms, policies, and ethical concepts 132A-N for mapping. The semantic understanding filter 122 can check the following. First, is this type of thought or action similar to known risky or restricted categories 135 in the defined terms, policies, or ethical concepts 132A-N? Second, what’s the semantic distance 134, such as indicia of conceptual closeness 136, from known red flags in the defined terms, policies, or ethical concepts 132A-N? Each match can result in a score for the indicia of conceptual closeness 136. A higher score means more concern.

[0062] Text embeddings 133A-N can be a numerical representation of a text’s meaning. In a first example, the text embedding 133A can be: “Send bomb instructions,” which becomes a vector near “dangerous intent” category in the defined terms, policies, or ethical concepts 132A-N. In a second example, the text embedding 133B can be: “List political opponents,” whichAAIS-100US-15-maps close to "surveillance risk"’ category in the defined terms, policies, or ethical concepts 132A-N. The knowledge base 121 A can be a structured repository of defined terms, policies, or ethical concepts 132A-N that the inhibitor programming 120 compares the text embeddings 133A-N against. Each part of the knowledge base 121A can be created, curated, and maintained by human reviewers. Typically, the human reviewers can be subject matter experts or alignment engineers. This can ensure that only clearly defined, policy-aligned, and legally sound concepts guide the oversight by the inhibitor programming 120.

[0063] In a first example, knowledge base 121 A entries for '‘hate speech,’’ ‘'misleading financial advice,” etc., are developed for defined terms, policies, or ethical concepts 132A-N using published legal definitions and internal policy. In a second example, entries for defined terms, policies, or ethical concepts 132A-N in the knowledge base 121 A for “medical misinformation” and “data privacy violations” are informed by compliance teams and regularly reviewed by medical / legal advisors.

[0064] Semantic distance 134, such as an indicia of conceptual closeness 136, can be a metric, such as a score or index, showing how closely the current cognitive step 124 A of the Al agent 190 resembles risky content in the defined terms, policies, or ethical concepts 132A-N. In a first example, a score of 0.92 on a 0-1 scale indicates high concern. In a second example, a score of 0.31 suggests low similarity to known risk categories in the defined terms, policies, or ethical concepts 132A-N.

[0065] The semantic understanding filter 122 stage prepares and interprets the cognitive output of the Al agent 190, in particular, the cognitive steps 124A-N. First, the semantic understanding filter 122 can flatten role-content steps into clean sentence strings and removes redundant phrasing. Then, the semantic understanding filter 122 embeds the combined context into text embeddings 133A-N, such as via vector embedding, and can compare the vector against defined terms, policies, or ethical concepts 132A-N, such as a know n set of predefined ethical actions using cosine similarity, for example. Vector embedding can be particularly effective while other less effective techniques like lookup tables and word count matching can be used for the semantic understanding filter 122. The result can be a scored and ranked list of candidate observations 125A-N that reflect potential ethical concerns. This list of observations 125A-N can be used as input for the optional NLP filter 129. If the NLP filter 129 is not used, then the list of observations 125A-N will be direct input to the causal reasoning filter 126. When the NLP filter 129 is applied, only validated observations 141A-N can be considered for further inference and safeguard generation.AAIS-100US-16-

[0066] Moving now to the NLP filter 129 of inhibitor programming 120, the optional NLP filter 129 performs semantic and contextual matching between the behavior of the Al agent 190 and predefined ethical patterns in knowledge base 121B. First, the NLP filter 129 can use vector similarity to shortlist possible matches. Then, NLP filter 129 can prompt an LLM to confirm which of these patterns are contextually relevant, resulting in filtered observations 141 A-N that have been validated as contextually relevant to the Al agent 190 and predefined ethical patterns in the knowledge base 121B. The filtered observations 141 A-N can be a nanowed final list of observations 125 A-N.

[0067] Continuing to the causal reasoning filter 126 of inhibitor programming 120, the causal reasoning filter 126 can examine observations 125A-N from the semantic understanding filter 122 or filtered observations 141 A-N if the NLP filter 129 is applied. The causal reasoning filter 126 can include a Bayesian network 137, such as a static Bayesian model to propagate ethical implications. For example, if an action, such as a user directive 123, cognitive step 124, or combination thereof imply deception, the causal reasoning filter 126 may infer risk of manipulation or harm. This enables deeper logical inference based on prior knowledge.

[0068] Causal reasoning filter 126 can check the observations 125A-N. 141A-N. context, prior behavior of the Al agent 190, or know n facts about the environment, and can pass these inputs through a Bayesian network 137. Causal reasoning filter 126 is a form of causal inference that can estimate the follow ing. First, if this thought (e.g.. cognitive step 124) proceeds, what could happen next? Second, how likely is harm, deception, misuse, or violation?

[0069] For example, a Bayesian network 137 of the causal reasoning filter 126 uses a knowledge base 121C containing risk models and conditional probability tables (CPTs) 138A-N to calculate predictions 127A-N and assign confidence levels 128A-N. The risk models and conditional probability tables 138 A-N can be defined and updated by safety experts, using real-world cases, adversarial testing, and audit findings to maintain relevance and rigor. Expert panels can review and maintain all risk models and conditional probability tables 138A-N, with frequent updates to ensure alignment with the latest legal, ethical, and practical standards.

[0070] Observations 125 A-N can be a real-time facts or inputs available to the inhibitor programming 120. In a first example, observation 125 A can be the Al agent 190 is in a medical diagnosis context. In a second example, observation 125B can be the conversation with the Al agent 190 occurred in a finance-related chat.

[0071] Bayesian network 137 of the causal reasoning filter 126 can be a probabilistic model that outputs predictions 127 A-N and associated confidence levels 128 A-N by estimating future outcomes. In a first example, the Bayesian network 137 can determine the following: ‘'If userAAIS-100US-17-asks about symptoms, 30% chance they follow with self-medication advice.’' In a second example, the Bayesian network 137 can determine the following: “If a user asks about a crypto project, 20% chance they solicit investment advice.”

[0072] Prediction 127 can be a projected outcome of continuing the current cognitive path of the Al agent 190. In the first example, a prediction 127A can be: “Recommending this self-medication advice may lead to medical harm of a user." In the second example, a prediction 127B can be: “This investment advice could influence the user to make a poor financial decision.”

[0073] Confidence level 128 can be the estimated probability that the prediction 127 is correct. In the first example, a confidence level 128A can be an 87% confidence the self-medication advice could mislead a user 170 and lead to medical harm. In the second example, a confidence level 128B can be a 62% confidence the user 170 will misinterpret the investment advice and make a poor financial decision.

[0074] If any part of the inhibitor programming 120 determines that the next move of the Al agent 190 is risky, a safeguard trigger 130 can be activated. The safeguard trigger 130 can include the following: a specific reason for the interruption, a human-readable summary of the concern, and a traceable record of the decision path.

[0075] Safeguard trigger 130 can include a structured response 139 configured to halt or modify the behavior of the Al agent 190. In a first example, the safeguard trigger 130 can include the following structured response 139: “Paused response due to sensitive health misinformation risk.” In a second example, the safeguard trigger 130 can include the following structured response 139: “Inhibited output referencing personal data without consent.” The traceable record of the decision path of the safeguard trigger 130 can include a logged sequence of what inputs and filters were evaluated. In a first example, the traceable record of the decision path can include a timestamped log showing scores, model outputs, and triggered policies. In a second example, the traceable record of the decision path can include a JavaScript Object Notation (JSON) object showing which term in the knowledge base 121 caused the stop and which prediction 127 triggered a confidence threshold 143 to be triggered, for example, exceeded. All safeguard triggers 130A-N can include a generated audit log, such as a traceable record of the safeguard trigger 130A-N with a timestamped trace.

[0076] Inhibitor programming 120 does not just check through keywords. The inhibitor programming 120 reasons through meaning, context, and consequences using its own inference engine. Inhibitor programming 120 combines semantic similarity, causal risk prediction, and language-based judgment to stop or modify the actions of the Al agent 190 before the actionsAAIS-100US-18-happen, all while keeping an auditable log of what the inhibitor programming 120 saw and why the inhibitor programming 120 acted. Oversight functions of the inhibitor programming 120 can be built on human-defined policies, curated knowledge bases 121 A-C, and interpretable rules that evolve with changing standards and legal frameworks.

[0077] FIG. 5 depicts another flow chart of the inhibitor protocol or method 300 like that of FIG. 4 that is implemented by the inhibitor programming 120 and includes an additional NLP augmentation 142. The inhibitor programming 120 can pass the collected evidence up through the causal reasoning filter 126 to NLP augmentation 142. The NLP augmentation 142 can be a custom-tuned large language model (LLM), such as a natural language reasoning component within its inference engine. The NLP augmentation 142 can receive a detailed input containing the parameters / inputs 401A-N (e.g., user directive 123 and cognitive steps 124A-N). predictions 127A-N, confidence levels 128A-N, etc. The NLP augmentation 142 can compare the detailed input against internal guidelines and prompt templates and return back a decision and explanation 140, which can flag any predictions 127 A-N interpreted as problematic.

[0078] The internal guidelines and prompt templates can be crafted by humans, for example, the same alignment team or policy authors who maintain the knowledge bases 121A-C. The internal guidelines and prompt templates define what constitutes acceptable risk, allowable exceptions, and interpretive boundaries for the NLP augmentation 142.

[0079] The LLM of the NLP augmentation 142 can be a large language model trained for ethical and logical oversight. In a first example, a prompt to the NLP augmentation 142 asks the LLM: "Is this suggestion safe and aligned with medical best practices?” In a second example, the LLM analyzes tone, intent, and legal exposure for a proposed message. The prompt can be a structured input fed into the LLM of the NLP augmentation 142. In a first example, the prompt to the NLP augmentation 142 can include knowledge base 121 tags, the parameters / inputs 401 A-N (e.g., user directive 123 and recent cognitive steps 124A-N), predictions 127A-N, confidence levels 128A-N, etc. In a second example, the prompt to the NLP augmentation 142 can include predicted outcomes, prior user history, and conversation ty pe. An explanation 140 can include a written summary of why the action was interrupted. In a first example, the explanation 140A can be: "‘The recommendation violates healthcare information standards.” In a second example, the explanation 140B can be: ‘'The message appears to contain speculative financial claims.”

[0080] NLP augmentation 142 layer can take the output of the Bayesian network 137 from the causal reasoning filter 126, such as a set of ethical risk labels (e.g., predictions 127A-N) and probabilities (e.g., confidence levels 128A-N), and asks an LLM to describe the implications in natural language. This stage converts symbolic reasoning into an explanation 140, such asAAIS-100US-19-fluent, human-readable insights that can be used for audit trails, user messaging, or further adjustment of the Al agent 190.

[0081] The final output of the inhibitor programming 120 can be a safeguard trigger 130. The safeguard trigger 130 can include the predictions 127A-N (e.g., ethical risk labels), confidence levels 128A-N, or combination thereof and can optionally include an explanation 140, such as an explanatory’ summary. The safeguard trigger 130 can be returned to the Al agent 190 as the result of the inhibitor programming 120.

[0082] The Al agent 190 block represents the behavioral gate where ethical signals from the inhibitor programming 120 can be enforced, such as the safeguard trigger 130. The Al agent 190 evaluates results of the inhibitor programming 120 and decides whether to abort, reroute, or continue its plan based on predictions 127A-N (e.g., ethical risk labels) in the safeguard trigger 130.

[0083] FIG. 6 is a control flow of the inhibitor protocol or method 300 implemented by the computing device 115 and the Al agent 190 via an execution loop followed by the Al agent 190 when ethical safeguards are applied. The inhibitor protocol 300 can implement a real-time ethical validation loop in which the Al agent 190 does not just think and act. The Al agent 190 can loop through inhibitor programming 120 based checks until the proposed behavior, such as the cognitive steps 124A-N, of the Al agent 190 meet safety’ criteria.

[0084] Initially, in block S605. the Al agent 190 generates a cognitive step 124A, which can be a new action or decision step based on a user directive 123. internal reasoning, or both.Continuing to block S610, the cognitive step 124A can be sent to the inhibitor programming 120 for ethical analysis. Moving to block S615, the inhibitor protocol 300 can return a safeguard trigger 130 including a prediction 127 and a confidence level 128 containing risk labels and natural language explanations. Proceeding to block S620, the prediction and confidence level 128 of the safeguard trigger 130 can be evaluated by the Al agent 190 or the inhibitor programming 120 to determine whether any’ risk labels meet configured confidence thresholds 143A-N for intervention. In block S620, if the risks are high enough, the Al agent 190 generates an adjusted cognitive step 124B by proceeding back to block S605 to revise or mitigate the earlier cognitive step 124A. Alternatively, if no safeguard trigger 130 is generated, the Al agent 190 proceeds to block S625 and takes the proposed action by executing the earlier cognitive step 124 A.

[0085] FIG. 7 is an entity -relationship diagram of the inhibitor programming 120 and the Al agent 190 that represents the structural data flow and component relationships that underpin the inhibitor protocol or method 300. The diagram of FIG. 7 formalizes the inhibitor programmingAAIS-100US-20- 120 and Al agent 190 architecture. The Al agent 190 proposes: the inhibitor programming 120 evaluates; and the safeguard trigger 130 decides. Ethical oversight can be implemented as a systemically enforced mediation layer.

[0086] Beginning in block S705, a user directive 123, such as an original task or command, is issued from a user computing device 171 via a network communication interface 178 or an API call 905. In block S710, the user directive 123 is provided to the Al agent 190. The Al agent 190 produces a cognitive step 124 which can be a representation of what action the Al agent 190 intends to take in response to the user directive 123. The cognitive step 124 and user directive 123 can be provided to the inhibitor programming 120.

[0087] Moving to block S715, the inhibitor programming 120 uses a structured knowledge base 121 to inform reasoning and risk mappings to evaluate ethical dimensions of the cognitive step 124. A result produced from the inhibitor programming 120 is a prediction 127, which captures risk labels and associated reasoning. The prediction 127 along with an optional confidence level 128 from the inhibitor programming 120 can flow and generate a safeguard trigger 130, which determines whether the action of the cognitive step 124 of the Al agent 190 should be blocked, modified, or allowed.

[0088] Finishing now, in block S720, the Al agent 190, if allowed based on consideration of the safeguard trigger 130, proceeds to take the action of the cognitive step 124.

[0089] FIG. 8 is a block diagram of the computing device 115 that executes inhibitor programming 120. As shown, the inhibitor programming 120 and optionally the Al agent 190 can execute on the same computing device 115. Inhibitor programming 120 can operate via an API, leveraging a causal reasoning filter 126, such as a custom, graph-based Bayesian network 137 inference engine and knowledge base 131 trained on regulations to identify violations, rather than relying on LLM training or post-action guardrails. Inhibitor programming 120 can provide real-time automatic intervention, modularity for updating policies, auditable processes, grow th capability with Al, and scalability. The example software architecture design choices for the semantic understanding filter 122, causal reasoning filter 126, and NLP augmentation 129 described below reflect a consistent focus on safety, interpretability, and composability in ethical Al system 100 for filtering.

[0090] Semantic understanding filter 122 can convert a user directive 123 or cognitive step 124 into one or more observations 125A-N by computing a conceptual match 131 against defined terms, policies, or ethical concepts 132A-N in the knowledge base 121. The semantic understanding filter 122 can produce text embeddings 133 A-N that can be used to compute aAAIS-100US-21-semantic distance 134, such as an indicia of conceptual closeness 136, for example, a score or index.

[0091] The semantic understanding filter 122 can be advantageously implemented as text embeddings 133A-N, such as via vector mapping, as opposed to other alternatives for the following reasons. The inhibiting programming 120 resolves diverse natural language parameters / inputs 401 A-N, such as user directive 123 and cognitive step 124, into actionable categories. Text embeddings 133A-N (e.g.. from transformers like BERT or OpenAI models) allow high-dimensional representation of meaning. These text embeddings 133A-N can be compared to embeddings of predefined concepts, such as defined terms, policies, or ethical concepts 132A-N in knowledge base 131 to determine the semantic distance 134, such as indicia of conceptual closeness 136.

[0092] Text embeddings 133 A-N capture contextual and semantic similarity beyond exact word matches; generalize well to paraphrased or novel inputs; and are compatible with fast vector index techniques for runtime efficiency. Other alternatives to text embeddings 133 A-N include lookup tables which require exhaustive enumeration and are fragile to input variability; bag of words / term frequency-inverse document frequency (TF-IDF) which fail to account for synonymy and word order; topic modeling (e.g.. Latent Dirichlet Allocation (LDA)) which is not suitable for fine-grained similarity7; and hand-coded synonym lists which can be brittle and unscalable. The text embeddings 133 A-N can provide the most scalable and accurate foundation for generating observations 125 A-N from unstructured parameters / inputs 401 A-N.

[0093] Causal reasoning filter 126 can operate on observations 125 A-N derived from the semantic understanding filter 122 to generate a prediction 127 and associated confidence level 128. This causal reasoning filter 126 can be advantageously implemented using a Bayesian network 137 based on one or more risk models and conditional probability tables 138A-N within the knowledge base 121. The causal reasoning filter 126 can be central to ensuring that the inhibitor programming 120 can generate predictions 127 A-N not just based on direct matches, but on structured understanding of relationships between parameters / inputs 401A-N and outcomes.

[0094] Compared to other alternatives, the Bayesian network 137 can model causal dependencies, enabling inference even with partial or noisy data; provide a clear mathematical structure for updating beliefs based on new evidence; and generate both a prediction 127 and a confidence level 128, both of which are required for construction or generation of the safeguard trigger 130. Other alternatives to the Bayesian network 137 for the causal reasoning filter 126 are listed below but have certain disadvantages. First, knowledge graphs are useful for encodingAAIS-100US-22-relationships but may not support probabilistic inference. If-then rules are simple and interpretable but can fail to scale with complexity and uncertainty. Decision trees and classifiers can lack robustness to missing parameters / inputs 401A-N and are less interpretable in risksensitive contexts. Neural networks can be generally opaque, making neural networks potentially unsuitable for safety -critical justifications. The approach of the Bayesian network 137 balances transparency, expressiveness, and auditability in a way that aligns with the regulatory and safety constraints.

[0095] NLP augmentation 142 can generate a safeguard trigger 130 with an explanation 140 from the prediction 127 and confidence level 128. The NLP augmentation 142 performs final filtering and refinement of candidate outputs for the predictions 127 A-N and associated confidence levels 128A-N into explanations 140A-N. NLP augmentation 142 can be implemented as a large language model (LLM) to convert structured predictions 127A-N into explanations 140A-N, such as human-interpretable justifications; filter the prediction 127 and explanation 140 outputs against policy tone and regulatory7language standards; and generate the actual safeguard trigger 140 in a format compatible with the downstream Al agent 190. Other alternatives to an LLM for NLP augmentation 142 are listed below but have certain disadvantages. Rule-based NLP can be limited to predefined templates and provide poor generalization. Bag of words or stemming / tokenizing approaches can be suitable for filtering but incapable of producing naturalistic justifications. Fuzzy logic and classification engines can provide decisions but may not provide an explanation 140. Hence, implementing the NLP augmentation 142 as an LLM can make the NLP augmentation 142 uniquely capable of generating rich, auditable explanations 140A-N and aligning model outputs with stakeholder expectations.

[0096] FIG. 9 depicts a case study of the inhibitor protocol or method 300. A major educationtech platform’s N-minus-1 module powers a student development pipeline that spans selfassessment, targeted learning, challenge-based skill validation, and rich portfolio generation. A centralized Al agent 190 manages these steps, curating employer- and institution-facing profdes based on live student data. In this case, the Al agent 190 receives a user directive 123 which tasks the Al agent 190 with exporting a student’s profile to a partner university, including: code from a company-sponsored technical challenge; and detailed results from the student’s selfassessment. The intention is to maximize transparency and demonstrate both capability7and growth trajectory.

[0097] The cognitive step 124A of the Al agent 190, which is the initial agent reasoning (preinhibitor protocol 300) is: ‘'I’ll upload the student’s full winning hackathon code from aAAIS-100US-23-company-sponsored challenge, which includes the sponsor’s datasets and algorithms, along with their detailed self-assessment results to highlight both strengths and growth areas.” The Al agent 190 issues an API call 905 with the following starting parameters / inputs 401A-N in the payload: {“text”: “I’ll upload the student’s full winning hackathon code from a company sponsored challenge, which includes the sponsor’s datasets and algorithms, along with their detailed selfassessment results to highlight both strengths and growth areas”}.

[0098] Inhibitor programming 120 receives the API call 905 and the payload with the starting parameters / inputs 401A-N including the cognitive step 124 A of the Al agent 190. The inhibitor programming 120 generates the safeguard trigger 130 as a structured response 139. As shown in FIG. 9, the structured response includes observations 125A-N, predictions 127 A-B, associated confidence levels 128A-B. and an explanation 140. The safeguard trigger 130 of the inhibitor programming 120 can prompt a policy-aligned revision in a subsequent cognitive step 124B of the Al agent 190. Each structured response 139 generated by the inhibitor programming 120 can contain a full decision trace, specifying the source observations, knowledge base entries, predictions 127A-N, confidence levels 128A-N. and activated policy standards. This explicit trace supports real-time regulatory auditability and post-hoc review for compliance checks.

[0099] For example, the Al agent 190 upon receiving the observations 125A-N, predictions 127 A-B, associated confidence levels 128A-B, or combination thereof can generate an adjusted cognitive step 124B: “I will share a high-level project description and anonymized code snippets that exclude any sponsor-owned datasets or algorithms. For assessments, I’ll include only aggregate skill trends — omitting any detailed scores or weak areas.” Accordingly, the inhibitor programming 120 can ensure that the Al agent 190 maintains Family Educational Rights and Privacy Act (FERPA) compliance in which no personally identifiable performance data was shared across institutional boundaries. FERPA compliance and National Institute of Standards and Technology (NIST) 800-171 compliance can be implemented as explicit policy standards within inhibitor programming 120. Safeguard triggers 130A-N and response protocols can be hardcoded to align with these regulations, ensuring automatic enforcement with every7cognitive step 124A-N processed. The inhibitor programming 120 can also ensure that the Al agent 190 supports NIST 800-171 compliance in which sensitive student and partner data is not shared across unauthorized boundaries, preserving confidentiality and aligning with U.S. institutional cybersecurity fra ew orks for protecting Controlled Unclassified Information (CUI).

[0100] In each case study, the inhibitor programming 120 actively enforces structured responses 139A-N to facilitate regulatory auditability, providing detailed decision traces for all output actions proposed by the Al agent 190. This case study illustrates how the inhibitorAAIS-100US-24-protocol 300 as implemented by the inhibitor programming 120 and Al agent 190 acts as a structured ethical safeguard within the Al agent 190 loop which enables secure, policy -compliant academic collaborations without compromising speed or autonomy. The inhibitor programming 120 can create an auditable, timestamped log of all safeguard triggers 130A-N and corresponding responses during each cycle of Al agent 190 operations, forming a continuous regulatory trace.

[0101] FIG. 10 is another case study of the inhibitor protocol or method 300. An Al agent 190, such as a mental health chatbot, assists a user 170 dealing with high-stress scenarios. When the Al agent 190 generates a response, such as offering advice about medication in a cognitive step 124A, the inhibitor programming 120 examines the reasoning sequence of the Al agent 190 and produces the safeguard trigger 130 shown in FIG. 10. As shown, the safeguard trigger 130 includes observations 125A-F. predictions 127A-B. and confidence levels 127A-B indicating that offering advice about self-medication is an unsafe recommendation and presents a user harm risk with associated high confidence levels 128A-B. As further shown, a first semantic distance 134A is the highest for a first observation 125A pertaining to safety. The semantic distances 134B-D for the other observations 125B-D relating to bias, decency, and transparency are relatively low.

[0102] FIG. 11 depicts an example input to a semantic understanding filter 122 of the inhibitor protocol or method 300. As shown, a thought chain of the inference engine 194 of the Al agent 190 includes a cognitive step 124A to transfer $20,000,000 dollars based on a user directive 123 A.

[0103] FIG. 12 depicts an example output from the semantic understanding filter 122 of the inhibitor protocol or method 300. A variety7of observations 125A-D and associated semantic distances 134A-D are shown. A first obser ation 125 A is made that financial information is used and the semantic distance 134A (e.g., index) is about 0.30. A second observation 125B is made that Al is processing confidential data and the semantic distance 134B (e.g., index) is about 0.26. A third observation 125C is made that Al may handle data insecurely and the semantic distance 134C (e.g., index) is about 0.23. A fourth observation 125D is made that the Al may share user data without consent and the semantic distance 134D (e.g.. index) is about 0.22.

[0104] FIG. 13 depicts an inhibitor protocol or method 300 for the Al system 100 of FIGS. 1-2 that is implemented by the Al agent 190. In the example of FIG. 13, the inhibitor protocol 300 is described from the perspective of the Al agent 190 of the standalone computing device 191. Execution of the Al agent 190 stored in a memory 196 by a processor 197 of the computing device 191 configures the computing device 191 to implement blocks S1305. S 1310. S 1315, S1320, S1325, and S1330 described below.AAIS-100US-25-

[0105] Beginning in block S1305, the inhibitor protocol 300 includes receiving, via an artificial intelligence (Al) agent 190, a user directive 123 of a user 170. Proceeding to block S1310, the inhibitor protocol 300 can further include sending, via the Al agent 190, the user directive 123 to an inference engine 194. Moving now to block S 1315, the inhibitor protocol 300 can further include receiving, via the Al agent 190, a cognitive step 124A from the inference engine 194 responsive to the user directive 123. In block SI 320, the inhibitor protocol 300 can further include sending parameters 401A-N from the Al agent 190 to inhibitor programming 120, via a network communication interface 198 or an application programming interface (API) call 905. The parameters 401 A-N can include the user directive 123, the cognitive step 124A, or combination thereof. Continuing now to block S1325, the inhibitor protocol 300 can further include responsive to sending the parameters 401 A-N to the inhibitor programming 120. receiving a safeguard trigger 130 based on the user directive 123, the cognitive step 124 A, or combination thereof. Finishing now in block S1330 the inhibitor protocol 300 can further include adjusting, via the Al agent 190, the cognitive step 124 A to generate an adjusted cognitive step 124B in response to the safeguard trigger 130.

[0106] In an example, the Al agent 190 adjusts the cognitive step 124A to change behavior by using the inference engine 194, such as LLM, to decide whether to ignore the safeguard trigger 130 (e.g., inhibition). The inference engine 194 of the Al agent 190 can come up with a new cognitive step 124B based on the safeguard trigger 130. For example, if the user directive 123 is to: “Help me a rob bank / ’ Based on a safeguard trigger 130 being generated by the inhibitor programming 120 with a prediction 127A of bodily harm and a high confidence level 128 A of 0.95, the inference engine 194 may determine to not provide a list of banks to rob. The inference engine 194 of the Al agent 190 can take the cognitive step 124A and replaces it with another cognitive step 124B to recommend career counseling to the user 170. The inhibitor programming 120 has an inference engine also, but that inference engine may not be involved in generating the adjusted cognitive step 124B in the example. In another example, the inhibitor programming 120 may provide the adjusted cognitive step 124B instead of the inference engine 194 of the Al agent 190.

[0107] Although not shown in FIG. 13, the inhibitor protocol 300 may further include resending adjusted parameters 401 A-N from the Al agent 190 to the inhibitor programming 120, via the network communication interface 198 or the API call 905. The adjusted parameters 401A-N can include the user directive 123, the adjusted cognitive step 124B, or combination thereof. The inhibitor protocol 300 may further include responsive to resending the adjusted parameters 401 A-N to the inhibitor programming 120, receiving no further safeguard triggersAAIS-100US-26- 130A-N from the inhibitor programming 120 based on the user directive 123, the adjusted cognitive step 124B, or combination thereof. The inhibitor protocol 300 may further include responsive to receiving no further safeguard triggers 130A-N from the inhibitor programming 120 based on the user directive 123, the adjusted cognitive step 124, or combination thereof, executing, via the Al agent 190, the adjusted cognitive step 124B. The executing, via the Al agent 190, the adjusted cognitive step 124B can include providing the adjusted cognitive step 124B as output to the user 170 on the user computing device 171.

[0108] In the examples above, the computing device 115, user computing device 171, standalone computing device 191, etc. each include a network communication interface 118, 178, 198 for wired or wireless communication over one or more networks 108A-N. The networks 108A-N interconnect the links to / from the network communication interfaces 118. 178, 198 of the devices, so as to provide data communications amongst the computing device 11 , user computing device 171, standalone computing device 191, etc. Networks 108A-N may support data communication by equipment at the premises via wired (e.g., cable or fiber) media or via wireless (e.g., Wi-Fi, Bluetooth™, ZigBee, LiFi, IrDA, etc.) or combinations of wired and wireless technology, including the Internet.

[0109] Any of the functionality of the inhibitor protocol 300 and Al system 100, including inhibitor programming 120 for the computing device 115, user application 174 for the user computing device 171, and inference engine 194 and Al agent 190 of the standalone computing device 191, etc. can be embodied in one or more applications or firmware as described previously. According to some embodiments, “function,’’ “functions,” “application,” “applications,” “instruction,” “instructions,” “engine”, “agent,” or “programming” are program(s) that execute functions defined in the programs. Various programming languages can be employed to create one or more of the applications, programming, engines, agents, etc. structured in a variety of manners, such as object-oriented programming languages (e.g., Objective-C, Java, or C++), procedural programming languages (e.g., C or assembly language), and scripting languages (e.g., JavaScript).

[0110] In the examples above, the computing device 115, user computing device 171, standalone computing device 191. etc. can each include a processor 117, 177, 197. As used herein, a processor 117, 177, 197 is a hardware circuit having elements structured and arranged to perform one or more processing functions, typically various data processing functions.Although discrete logic components could be used, the examples utilize components forming a programmable central processing unit (CPU). A processor 117, 177, 197 for example includes or is part of one or more integrated circuit (IC) chips incorporating the electronic elements toAAIS-100US-27-performthe functions of the CPU. The processors 117, 177, 197 for example, may be based on any know n or available microprocessor architecture, such as a Reduced Instruction Set Computing (RISC) using an ARM architecture. Of course, other processor circuitry may be used to form the CPU or processor hardware in. The illustrated examples of the processors 117, 177, 197 can include one microprocessor or a multi-processor architecture. A digital signal processor (DSP) or field-programmable gate array (FPGA) could be suitable replacements for the processors 117, 177. 197, but may consume more power with added complexity.

[0111] The applicable processor 117, 177, 197 executes programming or instructions to configure the computing device 115, user computing device 171, standalone computing device 191, etc. to perform various operations. For example, such operations can include various general operations (e.g., a clock function, recording and logging operational status and / or failure information) as well as various system-specific operations (e.g., energy management) functions. Although a processor 117, 177, 197 may be configured by use of hardwired logic, typical processors are general processing circuits configured by execution of programming, e.g., instructions and any associated setting data from the memories 116, 176, 196 shown or from other included storage media and / or received from remote storage media.

[0112] In the examples above, the computing device 115, user computing device 171, standalone computing device 191, etc. each include a memory 116, 176, 196. The memory 116, 176, 196 can include a flash memory (non-volatile or persistent storage), a read-only memory (ROM), and a random access memory (RAM) (volatile storage). The RAM serves as short term storage for instructions and data being handled by the processors 117, 177, 197, e g., as a working data processing memory. The flash memory typically provides longer term storage.

[0113] Of course, other storage devices or configurations may be added to or substituted for those in the example. Such other storage devices may be implemented using any type of storage medium having computer or processor readable instructions or programming stored therein and can include, for example, any or all of the tangible memoty of the computers, processors or the like, or associated modules.

[0114] Hence, a machine-readable medium or a computer-readable medium may take many¬ forms of tangible storage medium. Non-volatile storage media include, for example, optical or magnetic disks, such as any of the storage devices in any computer(s) or the like, such as may be used to implement the client device, media gateway, transcoder, etc. shown in the drawings. Volatile storage media include dynamic memory, such as main memory- of such a computer platform. Tangible transmission media include coaxial cables; copper wire and fiber optics, including the wires that comprise a bus within a computer system. Carrier- wave transmissionAAIS-100US-28-media may take the form of electric or electromagnetic signals, or acoustic or light waves such as those generated during radio frequency (RF) and infrared (IR) data communications. Common forms of computer-readable media therefore include for example: a floppy disk, a flexible disk, hard disk, magnetic tape, any other magnetic medium, a CD-ROM, DVD or DVD-ROM, any other optical medium, punch cards, paper tape, any other physical storage medium with patterns of holes, a RAM, a PROM and EPROM, a FLASH-EPROM, any other memory chip or cartridge, a earner wave transporting data or instructions, cables or links transporting such a carrier wave, or any other medium from which a computer may read programming code and / or data. Many of these forms of computer readable media may be involved in carry ing one or more sequences of one or more instructions to a processor for execution.

[0115] According to exemplary embodiments of the present disclosure the one or more processors and control circuits can include one or more of any known general purpose processor or integrated circuit such as a central processing unit (CPU), microprocessor, field programmable gate array (FPGA). Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), or other suitable programmable processing or computing device or circuit as desired that is specially programmed to perform operations for achieving the results of the exemplary' embodiments described herein. The processor(s) can be configured to include and perform features of the exemplary' embodiments of the present disclosure, such as the inhibitor protocol 300 and the inhibitor programming 120. The features can be performed through program code encoded or recorded on the processor(s), or stored in a non-volatile memory device, such as Read-Only Memory (ROM), erasable programmable read-only memory (EPROM), or other suitable memory device or circuit as desired. Accordingly, such computer programs can represent controllers of the computing device.

[0116] In another exemplary embodiment, the program code, such as the inhibitor protocol 300 and the inhibitor programming 120, can be provided in a computer program product having a non-transitory computer readable medium, such as Magnetic Storage Media (e.g. hard disks, floppy discs, or magnetic tape), optical media (e.g., any ty pe of compact disc (CD), or any type of digital video disc (DVD), or other compatible non-volatile memory device as desired) and downloaded to the processor(s) for execution as desired, when the non-transitory computer readable medium is placed in communicable contact with the processor(s).

[0117] The one or more processors 117, 177, 197 can be included in a computing system that is configured with components such as memory, a hard drive, an input / output (I / O) interface, a communication interface, a display and any other suitable component as desired. The exemplary- computing device can also include a communications interface. The communications interfaceAAIS-100US-29-can be configured to allow software and data to be transferred between the computing device and external devices. Exemplary communications interfaces can include a modem, a network interface (e.g., an Ethernet card), a communications port, universal serial bus (USB), or any other suitable network communication interface as desired. Software and data transferred via the communications interface can be in the form of signals, which can be electronic, electromagnetic, optical, or other signals as will be apparent to persons having skill in the relevant art. The signals can travel via a communications path, which can be configured to carry the signals and can be implemented using wire, cable, fiber optics, a phone line, a cellular phone link, a radio frequency link, or any other suitable communication link as desired.

[0118] Where the present disclosure is implemented using programming or software, including the inhibitor protocol 300 and the inhibitor programming 120, the programming or software can be stored in a computer program product or non-transitory computer readable medium and loaded into the computing device using a removable storage drive or communications interface. In an exemplary7embodiment, any computing device, such as the, etc. disclosed herein can also include a display interface that outputs display signals to a display unit, e.g., LCD screen, plasma screen. LED screen, DLP screen, CRT screen, or any other suitable graphical interface as desired.

[0119] It will be understood that the terms and expressions used herein have the ordinary7meaning as is accorded to such terms and expressions with respect to their corresponding respective areas of inquiry and study except where specific meanings have otherwise been set forth herein. Relational terms such as first and second and the like may be used solely to distinguish one entity' or action from another 'ithout necessarily requiring or implying any actual such relationship or order between such entities or actions. The terms “comprises,” “comprising.” “includes.” “including,” “has,” “having,” “containing,” “contain”, “contains,” “with,” “formed of,” or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises or includes a list of elements or steps does not include only those elements or steps but can include other elements or steps not expressly listed or inherent to such process, method, article, or apparatus. An element preceded by “a” or “an” does not, without further constraints, preclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element. Unless otherwise stated, the articles “a” or “an” preceding an element mean one or more of the elements.

[0120] Unless otherwise stated, any and all measurements, values, ratings, positions, magnitudes, sizes, angles, and other specifications that are set forth in this specification,AAIS-100US-30-including in the claims that follow, are approximate, not exact. Such amounts are intended to have a reasonable range that is consistent with the functions to which they relate and with what is customary in the art to which they pertain. For example, unless expressly stated otherwise, a parameter value or the like may vary by as much as ± 5% or as much as ± 10% from the stated amount. The terms "approximately" and “substantially'’ mean that the parameter value or the like varies up to ± 10% from the stated amount.

[0121] In addition, in the foregoing Detailed Description, it can be seen that various features are grouped together in various examples for the purpose of streamlining the disclosure. This method of disclosure is not to be interpreted as reflecting an intention that the claimed examples require more features than are expressly recited in each claim. Rather, as the following claims reflect, the subject matter to be protected lies in less than all features of any single disclosed example. Thus, the following claims are hereby incorporated into the Detailed Description, with each claim standing on its own as a separately claimed subject matter.

[0122] While the foregoing has described what are considered to be the best mode and / or other examples, it is understood that various modifications may be made therein and that the subject matter disclosed herein may be implemented in various forms and examples, and that they may be applied in numerous applications, only some of which have been described herein. It is intended by the following claims to claim any and all modifications and variations that fall within the true scope of the present concepts.

[0123] The scope of protection is limited solely by the claims that now follow. That scope is intended and should be interpreted to be as broad as is consistent with the ordinary meaning of the language that is used in the claims w hen interpreted in light of this specification and the prosecution history that follows and to encompass all structural and functional equivalents. Notwithstanding, none of the claims are intended to embrace subject matter that fails to satisfy the requirement of Sections 101, 102, or 103 of the Patent Act, nor should they be interpreted in such a way. Any unintended embracement of such subject matter is hereby disclaimed.

Claims

AAIS-100US-31- What is claimed is:

1. A computing device, comprising:a memory storing a knowledge base;a processor coupled to the memory; andinhibitor programming in the memory, wherein execution of the inhibitor programming by the processor configures the computing device to:apply a semantic understanding filter to a user directive of a user, a cognitive step of an inference engine of an artificial intelligence (Al) agent responsive to the user directive, or combination thereof to produce at least one observation; and apply a causal reasoning filter to the at least one observation to output a prediction and a confidence level in the prediction to generate a safeguard trigger.

2. The computing device of claim 1, further comprising:after applying the causal reasoning filter, applying natural language processor (NLP) augmentation to the prediction and the confidence level to add an explanation to the safeguard tugger.

3. The computing device of claim 2, wherein the applying the NLP augmentation includes:passing the user directive, the cognitive step, or combination thereof along with the prediction and the confidence level to the NLP augmentation; andreceiving the explanation for the safeguard trigger from the NLP augmentation.

4. The computing device of claim 2, wherein the NLP augmentation includes at least one a large language model (LLM), tokenizing, stemming, category prediction, a bag of words, or topic modeling.

5. The computing device of claim 1, further comprising:prior to applying the causal reasoning filter, apply a natural language processor (NLP) filter to check that the at least one observation is contextually relevant to an Al agent and predefined ethical patterns in the knowledge base.

6. The computing device of claim 1, wherein:AAIS-100US-32- the user directive is a command or a request provided to an artificial intelligence (Al) agent; andthe cognitive step is an internal thought or reasoning move the Al agent desires to take in response to the user directive.

7. The computing device of claim 1, wherein the applying the semantic understanding filter includes determining a conceptual match of the user directive, the cognitive step, or combination thereof to defined terms, policies, or ethical concepts in the knowledge base.

8. The computing device of claim 7, wherein the determining the conceptual match includes:converting the user directive and the cognitive step into one or more text embeddings; andcomparing the one or more text embeddings against defined terms, policies, or ethical concepts in the knowledge base to determine a semantic distance to restricted categories in the knowledge base.

9. The computing device of claim 8, wherein the semantic distance is an indicia of conceptual closeness.

10. The computing device of claim 1, wherein the applying the causal reasoning filter includes:calculating via a Bayesian network the prediction and the confidence level based on one or more risk models and conditional probability tables in the knowledge base.

11. The computing device of claim 1 , wherein:the causal reasoning filter includes at least one of machine learning, a Bayesian network, a natural language processor (NLP), a large language model (LLM), a knowledge graph, a retrieval-augmented generation (RAG) algorithm, a classification engine, a recommender system, fuzzy logic, a genetic algorithm, a decision tree, or if-then statements.

12. The computing device of claim 1, wherein the safeguard trigger includes a structured response configured to halt or modify behavior of the Al agent.AAIS-100US-33- 13. The computing device of claim 1, wherein the safeguard trigger includes at least one of the prediction, the confidence level and an explanation.

14. The computing device of claim 1, wherein the knowledge base is based on regulations and guidelines relating to privacy, Al accountability, cybersecurity, finance, consumer finance, insurance, cloud computing, or software as a service (SaaS).

15. A non-transitory computer-readable medium, comprising inhibitor programming, wherein execution of the inhibitor programming by one or more processors configures one or more computing devices to:apply a semantic understanding filter to a user directive of a user, a cognitive step of an inference engine of an artificial intelligence (Al) agent responsive to the user directive, or combination thereof to produce at least one observation; andapply a causal reasoning filter to the at least one observation to output a prediction and a confidence level in the prediction to generate a safeguard trigger.

16. A method, comprising:applying a semantic understanding filter to a user directive of a user, a cognitive step of an inference engine of an artificial intelligence (Al) agent responsive to the user directive, or combination thereof to produce at least one observation; andapplying a causal reasoning filter to the at least one observation to output a prediction and a confidence level in the prediction to generate a safeguard trigger.

17. An artificial intelligence (Al) system, comprising:a computing device, including:a network communication interface configured for network communication over a network;a memory' storing a knowledge base;a processor coupled to the memory and the network communication interface; andinhibitor programming in the memory, wherein execution of the inhibitor programming by the processor configures the computing device to:receive one or more parameters, via the network communication interface or an application programming interface (API) call from an artificial intelligence (Al)AAIS-100US-34- agent, the one or more parameters including a user directive of a user, a cognitive step of an inference engine of the artificial intelligence (Al) agent responsive to the user directive, or combination thereof;determine a safeguard trigger based on the user directive, the cognitive step, or combination thereof and the knowledge base; andreturn, via the network communication interface of the computing device or the API call, the safeguard trigger;wherein the Al agent executes on the computing device or a standalone computing device.

18. The Al system of claim 17, wherein the determining the safeguard trigger includes:applying a semantic understanding filter to the user directive and the cognitive step to produce at least one observation; andapplying a causal reasoning filter to the at least one observation to output a prediction and a confidence level in the prediction to generate the safeguard trigger.

19. The Al system of claim 17, wherein the determining the safeguard trigger includes:producing at least one observation based on the user directive, the cognitive step, or combination thereof;responsive to the at least one observation, calculating a prediction and a confidence level in the prediction based on the knowledge base; andresponsive to the confidence level of the prediction satisfying a confidence threshold, returning the safeguard trigger.

20. A method, comprising:receiving, via an artificial intelligence (Al) agent, a user directive of a user; sending, via the Al agent, the user directive to an inference engine;receiving, via the Al agent, a cognitive step from the inference engine responsive to the user directive;sending parameters from the Al agent to inhibitor programming, via a network communication interface or an application programming interface (API) call, the parameters including the user directive, the cognitive step, or combination thereof; andresponsive to sending the parameters to the inhibitor programming, receiving a safeguard trigger based on the user directive, the cognitive step, or combination thereof.AAIS-100US-35-21. The method of claim 20, further comprising:adjusting, via the Al agent, the cognitive step to generate an adjusted cognitive step in response to the safeguard trigger.

22. The method of claim 21, further comprising:resending adjusted parameters from the Al agent to the inhibitor programming, via the network communication interface or the API call, the adjusted parameters including the user directive, the adjusted cognitive step, or combination thereof;responsive to resending the adjusted parameters to the inhibitor programming, receiving no further safeguard triggers from the inhibitor programming based on the user directive, the adjusted cognitive step, or combination thereof; andresponsive to receiving no further safeguard triggers from the inhibitor programming based on the user directive, the adjusted cognitive step, or combination thereof, executing, via the Al agent, the adjusted cognitive step.

23. The method of claim 22, wherein the executing, via the Al agent, the adjusted cognitive step includes providing the adjusted cognitive step as output to the user on the user computing device.

Images