Sherkbox Project: Putting Sherlock Holmes in an AI Box
An automated reasoning system for criminal investigation & intelligence analysis
This article introduces an AI-driven criminal investigation and intelligence analysis system, built upon an automated reasoning architecture. The goal of Sherkbox is to propose and develop plausible investigation strategies by continuously evaluating the available evidence at any given time (T).
Multidimensional automated reasoning sequences for criminal analysis
The software model is designed to establish and automate investigation and analytical processes using the Sequences of Search and Reasoning (SSR) approach. This enables the system to methodically generate investigative hypotheses, validate them with evidence, and refine the reasoning process dynamically.
Operational view of Sherkbox
Sherkbox operates in three distinct modes:
Real-time automatic mode – autonomously producing investigative results.
Investigation-assist mode – supporting investigators and analysts by continuously generating investigative leads and testing their feasibility against the available evidence.
Simulation and training mode – allowing users to test case studies and simulate results for training purposes.
Modular architecture for intelligent criminal analysis
At its core, Sherkbox employs intelligence inference meta-rules, which include deductive, inductive, abductive, and retroductive reasoning to formulate and refine plausible investigative hypotheses. These rules are structured around a set of predefined, domain-independent analytical questions that guide the reasoning process. Currently, the system encompasses:
40+ investigative questions
150+ intelligence inference rules
By linking analytical sequences to investigative queries, Sherkbox provides structured insights into complex criminal scenarios.
Search & Reasoning Secuences
Beyond human limits: Enhancing analytical capabilities
While human investigators excel at motivational, resultative, predictive, and intentional inferences, integrating automated reasoning enables the system to handle complex inference combinations efficiently. This capability significantly enhances investigative outcomes, offering high-value intelligence insights.
Sherkbox represents a new frontier in AI-driven security and intelligence, bridging automated reasoning with real-world criminal investigations. By systematically analyzing evidence, generating hypotheses, and testing investigative strategies, it enhances both the speed and accuracy of criminal intelligence analysis.
Intelligence inference meta rules
Specific inferences: What conceptual components are likely missing in an incomplete conceptual group?
Causal inferences: What were the primary causes of an action or state?
Resultative inferences: What are the likely outcomes (effects on the world) of an action or state?
Motivational inferences: Why did or would an actor perform a specific action? What were their intentions?
Inferences about capabilities: What states in the world must be (or should have been) true for an action to occur?
Functional inferences: Why do people desire certain objects?
Inferences on predictions and qualifications: If a person wants the world to be in a specific "state," is it because such a state will enable them to take a predictable action?
Inferences on limitations: If a person cannot perform a desired action, can it be explained by a "state" of the world that is a prerequisite and has not yet occurred?
Inferences about mediation: When an action is causing (or will cause) undesirable outcomes in the world.
Inferences about predicting actions: Knowing the needs or desires of a person, what actions are they likely to take to fulfill those needs or achieve those desires?
Inferences about spread of knowledge: If a person knows certain information, what additional knowledge can be inferred that they also possess?
Inferences about standards: Relating to what is considered normal in the world, how believable is a story in the absence of specific knowledge?
Inferences about permanence: For a lengthy state or action, how long is it likely to last?
Inferences about traits: Knowing some traits of an entity and the contexts in which it operates, what additional characteristics or behaviors can be predicted?
Inferences on situations: What other information can be imagined (or inferred) from a familiar situation?
Inferences of intention: What can be inferred from the way something is expressed? Why was it expressed in that manner?
Relational inferences: Diachronic (who or what can be related to an actor over time) or synchronic (who or what has been related to an actor during the course of an action or event).
Inferences about spread of relationships: Knowing that an actor is related to certain entities, what other entities can be inferred to have connections to them?
High-value insights from these inferences
The objectives of an opponent.
The emotional state of an opponent.
The preferences of an opponent.
The plans of an opponent.
Anticipating the actions of an opponent.
Manipulations and traps of an opponent.
The strategic center of gravity of an organization.
The social and psychological profile of a person or group.
Hidden relationships between people or organizations.
The relationships of a person or organization
Past, present, and intended social network of a person or organization.
Profiles and perceptions of individuals and organizations.
Alternative ways of solving problems and making selections.
Alternatives in light of the situation and objectives.
The possibility of future events based on past occurrences.
The credibility of information, people, and organizations.
Identifying the correct targets for specific goals and actions.
The processes that can lead to new events or situations.
The opinions of experts on a given topic.
The necessary trade-offs to perform an action.
Trends in the escalation of a situation or conflict.
The strength of actors in a potential operational scenario.
The strategy employed by actors in a possible operation.
The disposition of a target's forces in terms of political, geographical, and readiness factors.
The significance of numerical patterns and trends in people, objects, and organizations.
The degree and nature of risk associated with an actor, target, or action.
The level of opportunity available to an actor, factor, or situation.
The ability to influence a person or organization.
The vulnerabilities of a person or organization.
Predicting future outcomes based on present evidence.
Systematic problems within an organization.
The various conflict scenarios emerging from present or future conditions.
Key influencing trends affecting people and organizations within a specific time frame.
Desirable / Undesirable events and their potential consequences.
The risk levels of actions and their associated costs.
The dissolution of structural problems.
Temporary or total inhibition about situations in the environment.
Critical incidents for an organization or individual.
Detection of critical incidents for an organization or individual.
Situations that would arise if we make statements contrary to our objectives.
Possible and achievable scenarios depending on the morphology of the proposed actions.
New scenarios based on the strategies of the actors.
Possibility of occurrence of unthinkable situations.
Processes that could lead to unthinkable situations.
Generating intuitions.
Generating clichés.
Generating dreams and fantasy questions.
Validation or refutation of hypnotic images.
Contradictions and paradoxes of the organization.
Implicit communication flows in organizations and individuals.
Situations resulting from erroneous actions.
The influence of external organizations and individuals.
Extreme situations.
Factors that allow for establishing early warnings of threats, risks, and opportunities..
Refining automated reasoning for criminal and terrorist intelligence analysis
An essential part of my experimental design process involves identifying the meta-rules and inference rules that govern automated reasoning. These rules must be scientifically validated through cognitive psychology, probabilistic reasoning, and critical thinking to ensure their robustness and applicability. Once validated, they are coded for integration into an AI-powered investigative system.
Structured intelligence training for analysts
Building upon this foundation, the meta-rules and reasoning inferences can be leveraged for training intelligence analysts in criminal and counterterrorism investigations. By using structured reasoning processes, analysts can internalize systematic thinking patterns, enhancing their ability to:
Understand complex situations
Propose plausible investigative and analytical pathways
Formulate hypotheses quickly and coherently
This approach fosters a methodical, logic-driven mindset, crucial for intelligence and security operations.
Applying Witmore’s Model to criminal and terror analysis
A structured approach to verifying or refuting binary hypotheses about criminal or terrorist situations aligns with Witmore’s model of evidence marshaling and hypothesis testing. This model provides a systematic framework for evaluating evidence, ensuring that intelligence assessments are grounded in logic and empirical data.
Witmore's Model of evidence marshalling and hypothesis testing
An evolving cognitive AI system for real-world analysis
For effective deployment, the AI system must operate as an evolutionary cognitive system, capable of:
Processing and managing real-time intelligence
Relating new information to an existing knowledge base
Generating comprehensive, evidence-backed investigative inferences
By correlating physical evidence, documents, and detected patterns, the system enhances real-world intelligence analysis. This adaptive AI framework enables a dynamic understanding of evolving criminal and terrorist threats, allowing security professionals to anticipate and counteract risks effectively.