Hierarchical State Intelligence (HSI)
From the next word to the next action.
ABOUT HSI
HSI is an AI model that learns the relationship between the current state and actions, and predicts future changes.
Drawing on past experience and goals, it supports decisions about the next action to take.
Most AI today has developed its ability to write natural text by learning to predict the next word that follows a passage. What HSI learns is how the state of a target changes, and which action to choose.
HSI is an AI model that places the relationship between "State," "Action," and "Next State" at the center of learning. It supports goal-aligned decision-making while predicting changes in the situation and the results of actions.
HSI is not intended simply to replace large language models (LLMs). Generating text is where LLMs excel. HSI is an AI model with a different role: understanding the state, predicting change, evaluating alignment with goals, and deciding on actions. We also envision it working alongside existing AI technologies.
HSI is built on three core technologies—"State Learning," which learns the relationships among pieces of state information; "State Transition Prediction," which predicts actions and the changes in state that follow; and "Hierarchical Intelligence," which evaluates the alignment between higher-level goals and individual actions.
State Learning
HSI reads a situation through the connections between items—something no single item can reveal.
For example, when a software development issue is "Investigating," the state is represented as progress stage = Investigating, priority = High, stakeholder reaction = Negative, time to deadline = 2 days, assignee = Unassigned, number of reopenings = 2. Which items are related, and how strongly, is something HSI learns—without people specifying the correct answer—by adjusting in proportion to how far each prediction missed.
HSI also puts past experience to work. Rather than just looking for similar situations, it considers the action taken at the time, its result, how recent it was, and more, and selects and uses the experiences most useful for the current state.
State Transition Prediction
From the current state, HSI predicts the next action to be taken and the state that follows from it.
01
For each candidate action, it simulates changes in state several steps ahead and compares them in terms of reward, risk, authority, safety, budget, and more. It does not choose an operation it lacks the authority to perform, even if that operation has high value.
02
For example, it compares assigning an owner, continuing the investigation, and postponing action—looking not only at the immediate outcome but at the outcomes further ahead—and selects the best action within the rules.
Hierarchical Intelligence
HSI treats everything from organizational goals down to actual actions as a hierarchy, and evaluates how well the levels align (the degree of alignment).
01
It shows between which levels the misalignment lies—for example, the upper levels are well aligned, but there is a large gap between the tasks and the actual actions.
02
By tracking changes in the degree of alignment, it catches signs of deviation, stalling, and failure at an early stage, leading to warnings, halted processing, requests for approval, or changes in action. Its predictions of changes in alignment also come with a range of possible variation.
HSI Judge
HSI Judge is a decision engine that goes beyond judging whether an answer is correct, evaluating whether an action should be carried out at all.
The function that "proposes" an action and the function that "executes it automatically" are separate. In actual operation, it is meant to be used together with approval and control mechanisms.
Example: a development project
Applied to issue management in software development, HSI works as follows.
01
Issue "Investigating" / high priority / 2 days to deadline / no assignee
02
Evaluates assigning an owner, continuing the investigation, and postponing, and predicts the next state for each action
03
Checks authority, then requests approval or sends a notification. Records the actual result and uses it for improvement
HSI is not limited to a specific use. It is a general-purpose AI model technology intended for application across many fields.
01
Each time the state of an issue changes, it shows the next move and signs of stalling or failure
02
It detects actions that stray far from the goal, then holds back execution or requests approval
03
It tracks the progress, risks, and goal alignment of long-running work, and supports business decisions
04
It predicts changes in the state of equipment and processes, and supports production management and operational decisions
05
It uses changes in state and historical information to support the decisions of healthcare professionals
06
It supports risk evaluation and response decisions based on changes in the situation and signs of anomalies
07
It can be applied to motion planning and action selection based on predicted environmental states and action outcomes
TIGEREYE has filed a patent application for the core technology of HSI.
Toward workplaces where AI is simply part of everyday work.
